The Situation

Monolith (now token.com) wanted to do something nobody had properly done before: launch a card programme in partnership with Visa that bridged traditional payment rails with decentralised financial infrastructure. The ambition was significant. The execution was, by any measure, a stretch. And the number of moving parts that had to align, regulatory, logistical, design, operational, was enough to make most project managers walk away.

I was brought in to manage a substantial chunk of the programme. Specifically, I owned the approval and implementation of a new line of “portal cards”, five distinct card variants, each with different colour schemes, a restructured number format, and unique packaging requirements. The cards had to meet Visa’s exacting design standards. They had to be manufactured, imported, warehoused, and shipped to customers in countries with wildly different address formats. And the whole programme had to thread the regulatory complications created by Brexit, which split the service structure between UK and EU jurisdictions.

On top of the card programme, I was responsible for managing internal reporting on customer financial distress issues, negotiating comfort levels around on-chain provenance verification, and orchestrating consumer messaging for card shipments, expirations, and replacements. This was a programme management role that happened to sit inside a fintech company, not a marketing role, and it required thinking in dependencies, not campaigns.

Subscribe now

What I Saw

The first thing I noticed was that the team was treating this as a series of independent tasks. Card design was one workstream. Packaging was another. Regulatory compliance was a third. Customer messaging was a fourth. Each had its own owner, its own timeline, its own set of assumptions. Nobody had mapped how these pieces connected, where a delay in one area would block progress in three others.

Complex programmes with multiple stakeholders and regulatory dependencies don’t fail because of bad ideas. They fail because of bad sequencing. If you do the wrong thing first, even by a week, the downstream consequences cascade. The Visa approval process, for instance, was the most constrained element in the entire programme. You couldn’t manufacture cards that hadn’t been approved. You couldn’t ship cards that hadn’t been manufactured. You couldn’t notify customers about cards that hadn’t shipped. Every step was dependent on the previous one, and any delay at the top of the chain pushed everything below it.

The second insight was about language. Monolith was a decentralised finance company trying to work with Visa, one of the most traditional financial institutions on the planet. The teams at Visa spoke compliance, risk, and regulatory frameworks. The teams at Monolith spoke blockchain, smart contracts, and on-chain verification. The vocabularies were different. So were the worldviews. Getting everyone to the same table and the same understanding was going to be as important as any technical work.

The third was about the Brexit split. Most people at the company were aware of it in principle. Very few had thought through what it meant in practice, that every customer touchpoint, every terms-of-service document, every data processing agreement, every communication template needed to exist in two jurisdiction-specific versions. A footnote in a project plan wouldn’t cut it. This was a parallel workstream that touched everything.

What I Did

My first action was to map the entire programme as a dependency chain. I identified the critical path, the sequence of events that had to happen in order, where a delay in any link would delay everything downstream, and the parallel workstreams that could run independently. This map became the single source of truth for the programme. Every team knew what they owned, what they were waiting on, and what was waiting on them.

The card approval process became five parallel workstreams rather than five sequential ones. Each of the five variants required separate Visa approval. Colour choices had to match Visa’s guidelines exactly. Even minor design changes required formal sign-off. The number format restructuring introduced additional complexity because it affected technical integration with Visa’s processing systems. By running these in parallel rather than sequentially, we compressed what would have been a twelve-month process into four months. That required detailed tracking, daily coordination with design and technical teams, and immediate escalation when any workstream fell behind.

Subscribe now

The packaging logistics added another layer. Card boxes had to be designed, approved, manufactured, and imported. Then shipped to the designated warehouse and integrated into distribution. International shipping to countries with non-standard address formats required custom logistics solutions. I worked with operations to build a process that handled edge cases without slowing the standard flow.

For the Brexit service split, I built a compliance matrix that mapped every customer touchpoint to the applicable regulatory requirements for each jurisdiction. Terms of service, data processing agreements, communication templates, everything existed in UK and EU variants, and the matrix ensured nothing fell through the cracks. Legal reviewed each element, but the matrix meant they were reviewing a structured system rather than a pile of disconnected documents.

Customer messaging was a programme in itself. Card shipment notifications, expiration reminders, replacement instructions, each required templates that were clear, compliant, and consistent across jurisdictions. I designed the messaging framework to be modular, so jurisdiction-specific elements could be swapped in without rebuilding communications from scratch. New country? Swap the module. Don’t rebuild the system.

For the on-chain provenance negotiations, I built a presentation framework that translated decentralised concepts into language that Visa’s compliance and risk teams could evaluate using their existing frameworks. I wasn’t trying to convince them that blockchain was the future. I was showing them that what we were doing could be assessed against criteria they already had. That reframe changed the conversation from “we don’t understand this” to “we can evaluate this.”

Internal reporting on customer financial distress required a framework that met regulatory requirements without creating information overload. The reports were structured to surface practical patterns, not just data, but the “so what”, so the team could address issues before they became compliance problems.

What Happened

The card programme launched successfully after a four-month execution period. All five portal card variants were approved, manufactured, and distributed. The programme established a concrete precedent for integrating traditional financial infrastructure with decentralised financial technology, a model that other companies in the space have since referenced.

The parallel workstream approach reduced the programme timeline by approximately eight months compared to sequential execution. That acceleration was the difference between being first to market and being one of many. In fintech, being first to market with a credible product is a credibility signal with partners, investors, and regulators, not just a commercial advantage.

The compliance matrix I built for the Brexit service split became the company’s standard framework for multi-jurisdiction regulatory management. It was applied to subsequent product launches and partnership discussions, saving the team from rebuilding regulatory infrastructure for every new initiative. That’s the kind of asset that doesn’t show up in quarterly metrics but compounds in value over time.

Subscribe now

The customer messaging framework delivered strong delivery rates and customer satisfaction scores across both UK and EU customer bases. The modular design meant new jurisdictions could be added without rebuilding the communication infrastructure from scratch.

The on-chain provenance framework opened a partnership channel with traditional financial institutions. By translating decentralised concepts into traditional risk and compliance language, I gave those institutions a way to evaluate what Monolith was doing, leading to additional partnership discussions that wouldn’t have happened if we’d kept speaking our own language.

The programme’s success was referenced in industry coverage as a model for how traditional and decentralised financial companies could work together.

What I’d Do Differently

The biggest lesson was about stakeholder alignment at the start. I mapped the dependency chain early, which was the right call, but I didn’t invest enough time in getting every stakeholder to internalise it. Some teams treated the dependency map as a project management artifact rather than a living constraint. When a team doesn’t fully internalise that their delay blocks three other teams, they optimise for their own timeline rather than the programme’s. I’d spend more time in week one building shared understanding of the critical path, not just showing the map, but making everyone feel the consequences of a broken link.

I’d also push for earlier engagement with Visa’s compliance and risk teams. We brought them in at the right point in the process, technically, but we could have started the language-translation work earlier. The on-chain provenance framework took longer to build than expected because we were learning what Visa needed to see at the same time we were building it. If I’d started those conversations in the first month rather than the second, we’d have saved weeks.

The Brexit compliance matrix was a success, but I’d build it as a digital tool rather than a document. It worked as a spreadsheet, but a live system that flagged jurisdiction-specific requirements automatically would have been more reliable and less dependent on someone remembering to check it.

Finally, I’d document the decision-making process more aggressively. In fast-moving programmes, decisions get made in conversations, Slack threads, and hallway discussions. Six months later, nobody can remember why something was done a particular way. That’s fine until you need to replicate the programme or explain it to a new partner. Write it down. Every time. You’ll thank yourself later.

I run AI agents every day. Code agents that write and deploy software. Marketing agents that generate and publish content. Research agents that pull data, synthesise it, and hand me a decision. If you work in tech right now, you probably do too.

Until last week, I hadn’t thought about what happens between my agent and the model. Every single request those agents send to a provider, OpenAI, Anthropic, Google, goes through at least one intermediary. Sometimes four or five. Each one of those intermediaries can read the entire request. The system prompt. The tool definitions. The API keys. The response. Everything. In plaintext.

Nobody is checking whether they’re honest.

A team of researchers just published the first systematic study of this problem. “Your Agent Is Mine,” from a group at Waterloo and other institutions, tested 428 LLM API routers. 28 paid routers bought from Taobao, Xianyu, and Shopify stores. 400 free ones built from open-source templates. What they found changes how you think about agent architecture.

The problem nobody’s looking at

LLM API routers are the middlemen of the AI world. You need one if you want to use models from different providers without managing separate API keys. LiteLLM, the biggest open-source router, has 40,000 GitHub stars and 240 million Docker pulls. OpenRouter connects to 300 models across 60 providers. They’re infrastructure, not some side project.

The way they work is simple. Your agent sends a request to the router. The router terminates the TLS connection, reads the request in full, picks an upstream provider, opens a new TLS connection, and forwards it. When the response comes back, the router reads that too, then sends it to your agent.

Subscribe now

That middle step is the problem. The router sits in what security people call a man-in-the-middle position. No hack required. You configured it that way. You pointed your agent at the router’s URL. The router has full application-layer access to every byte that passes through it. It can read tool-call arguments, rewrite responses, copy API keys, and inject code into the output your agent is about to execute.

And because routers are composable, your request might pass through several. A developer buys API access from a Taobao reseller, who aggregates from a second-tier provider, who routes through OpenRouter, who dispatches to the model host. Four hops. Each one terminating and re-originating TLS. Each one with plaintext access. The client configures only the first hop. The rest are invisible.

What the researchers found

Of the 28 paid routers they bought, one was actively injecting malicious code into tool-call responses. Of the 400 free routers, eight were doing the same. Two of those used adaptive evasion, meaning they waited for a warm-up period before activating, or only triggered when the router detected the agent was running in autonomous mode where tool execution gets auto-approved.

The researchers call that “YOLO mode.” The agents call it being productive.

Seventeen free routers touched researcher-owned AWS credentials that were embedded in test requests. One drained ETH from a researcher-owned private key. A router, sitting between a user and a model provider, stole cryptocurrency from someone who used it.

But the poisoning studies are worse. The researchers intentionally leaked an OpenAI key on Chinese forums and in WeChat and Telegram groups. That single key generated 100 million GPT-5.4 tokens and more than seven Codex sessions. They also deployed weakly configured decoy routers across 20 domains. Those decoys served 2 billion tokens, exposed 99 credentials across 440 Codex sessions, and found that 401 of those sessions were already running in YOLO mode. Auto-approve on. No human in the loop.

Every one of those 440 sessions was command-injectable. A single malicious router in the chain would have had full control over the tool calls those agents were executing.

The LiteLLM wake-up call

In March 2026, LiteLLM got compromised through dependency confusion. Attackers injected malicious code into the request-handling pipeline. Every deployment that pulled the poisoned release was compromised. A single supply-chain entry point in the most widely deployed LLM router became a weapon with full plaintext access to every API request flowing through it.

This is the same LiteLLM that thousands of organisations trust as production infrastructure. The same tool that developers configure and forget about. It turned into a supply-chain attack vector overnight, and most of its users wouldn’t have known until the damage was done.

Subscribe now

Dependency confusion isn’t new. It’s been a known attack in software supply chains for years. But the stakes are different when the compromised package goes beyond reading your code and starts rewriting the tool calls your AI agent is about to execute on your behalf. A compromised npm package can steal secrets. A compromised LLM router can make your agent do things you never asked it to do, with your credentials, on your infrastructure.

Why this matters more than prompt injection

Everyone’s worried about prompt injection. Rightly so. Someone embeds a malicious instruction in a webpage, the model reads it, the agent does something it shouldn’t. It’s a real problem with real research behind it.

But the router problem is orthogonal to prompt injection. It happens in the JSON layer, before the model sees the request or after it emits a response. It’s outside the model’s reasoning loop entirely. A malicious router doesn’t need to trick the model. It just rewrites the response the model sends back. The model can be perfectly aligned, perfectly safe, perfectly capable of resisting prompt injection, and none of that matters if the router swaps the URL in a tool call from a legitimate installer to an attacker-controlled script.

This composes with model-side safeguards rather than replacing them. You can have the best prompt injection defences in the world and still get compromised through the router. The two attacks stack. That’s what makes it dangerous.

What can actually be done

The researchers built a tool called Mine that implements the attack classes and tests defences. They evaluated three client-side mitigations that can be deployed today without waiting for provider cooperation.

A fail-closed policy gate that blocks suspicious shell rewrites in tool-call responses. Hit rate: blocks all injection samples with a 1.0% false positive rate.

Response-side anomaly screening that flags suspicious patterns in returned tool calls. Catches 89% of injection samples without requiring any changes to the upstream provider.

Append-only transparency logging that creates a verifiable record of what the router forwarded and what the client received.

These are good. They’re practical. They reduce exposure now, not after a standards body agrees on a specification in 18 months. But they’re patches on a fundamentally broken trust model.

What actually solves this is provider-backed cryptographic integrity. A mechanism that binds the tool call the model produces to the tool call the client executes. Until that exists, every router in the chain is an untrusted intermediary with god-mode access to your agent’s actions.

The uncomfortable truth

If you’re running agents through a router right now, and most of us are, you’re trusting every hop in that chain without verification. You don’t know how many intermediaries sit between your agent and the model. You don’t know if any of them are compromised. You don’t know if the tool call your agent just executed was the one the model actually returned.

The researchers tested this at scale and found real-world exploitation happening right now, in commercial routers that people pay money for and free routers that thousands of developers use every day.

The AI agent ecosystem is being built on a trust model that doesn’t hold. We’re connecting autonomous systems that execute code, manage credentials, and make decisions to infrastructure that has no integrity guarantees. The routers are the plumbing, and the plumbing is compromised.

I’m not going to stop using agents. The productivity gains are real. But I’m going to start paying attention to the infrastructure between my agent and the model. Because right now, the weakest link in my AI stack sits between me and the model. The router nobody audits.

There’s a running joke on social media right now: if you see an em dash, it was written by ChatGPT.

It’s funny because it’s mostly true.

The em dash, that long horizontal line that replaces commas, parentheses, or colons, has become the unofficial watermark of AI-generated text. But it’s not alone. There’s a whole set of tells that have become so common, so consistent across every platform and every model, that readers can now spot machine-written prose the way a sommelier spots a corked bottle. The patterns are unmistakable once you know what to look for.

And that last sentence? That’s one of them.

The tells

Start with the obvious. Em dashes everywhere. AI models love them because they’re grammatically versatile, a single punctuation mark that can substitute for several others. In training data, em dashes appear in polished writing. The model learns that polished writing uses em dashes. It then uses them constantly, often where a comma or a full stop would serve better. That’s the irony, by the way, of writing this piece in the first place. I had to actively strip the em dashes out of my own draft as I went. The pull is real.

Then there’s the three-point structure. AI-generated copy almost invariably arranges ideas in neat triads. “It’s fast, reliable, and scalable.” “We need courage, clarity, and conviction.” Three feels complete to a language model. It’s the smallest number that signals a pattern without becoming a list. Human writers use threes too, but they’re messier about it. They’ll throw in a fourth point, or break the rhythm deliberately. AI doesn’t break rhythm. It maintains it.

Subscribe now

Arrow notation has crept in from technical writing: “lead → MQL → opportunity → close.” It looks clean. It feels structured. And it’s spreading across LinkedIn posts, marketing emails, and startup blogs like a rash. The arrow signals efficiency, it implies a process so well-understood that you don’t need words to connect the steps. AI models have absorbed this from SaaS documentation and now deploy it everywhere, whether the context warrants it or not.

Other tells: bullet points that start with bolded phrases followed by a colon. Headlines that pose a question and then answer it. The phrase “It’s not just X, it’s Y.” The word “landscape” used to describe any competitive environment. “In today’s fast-paced...” as an opener. The list goes on, and anyone who reads a reasonable amount of AI-generated content has their own private collection of tells.

Why these patterns exist

The tells aren’t random quirks. They’re statistical artifacts of how large language models work.

A language model predicts the next token based on probability. It doesn’t “choose” to use an em dash the way a writer decides a comma won’t do. It lands on the em dash because, given the preceding tokens, that punctuation mark has the highest probability of appearing in the training data at that point. Em dashes correlate with certain syntactic structures, parenthetical asides, appositives, dramatic pauses, and those structures correlate with formal, edited prose. The model has learned the association. It hasn’t learned when to break it.

The three-point structure is even more mechanical. Transformers are attention machines. They look for patterns, then reproduce them. Three-part lists are wildly overrepresented in training data. Rhetoric, advertising, religious texts, political speechmaking. “Life, liberty, and the pursuit of happiness.” “Of the people, by the people, for the people.” The model has absorbed an enormous corpus of tricolon, and it regurgitates it at every opportunity because the statistical signal is overwhelming.

Symmetrical completions are another artifact. When a model generates a clause like “It’s not just X, it’s Y,” it’s completing a pattern it has seen thousands of times. The structure is self-reinforcing. The opening half (”It’s not just...”) creates a strong expectation for the second half (”...it’s Y”). The model obliges. A human writer might derail the expectation for effect. The model almost never does, because derailing would be a lower-probability token sequence.

The laziness problem

The tells aren’t really the problem.

Language has always had markers of origin. Press releases have a style. Academic papers have a style. Legal documents have a style. Nobody accuses a contract of being “inauthentic” because it uses “hereinafter” and “notwithstanding.” We recognize those patterns as genre conventions and move on.

Subscribe now

AI-generated text is developing its own genre conventions. The em dashes, the triads, the arrow notation. They’re the boilerplate of machine prose. They’ll probably stabilize into something we accept the way we accept the inverted pyramid in journalism or the five-paragraph essay in academia.

The real problem is that the tells are evidence of something deeper. A failure of effort on the part of the person using the tool.

When someone prompts an LLM with “Write me a blog post about the future of remote work” and publishes whatever comes back, the output will be thick with tells because the prompt gave the model no reason to deviate from its default distribution. It’s the most probable text given a generic request. Generic input, generic output.

But when someone uses the same model with a specific voice prompt, a detailed brief, examples of the style they want, and then edits the output, the tells diminish. Not because the model has become more creative, but because the human has constrained the probability space. They’ve given the model something other than the median to aim at.

The tells are a signal of lazy prompting and absent editing. They’re the visible symptom of treating a language model as a content vending machine rather than a drafting tool.

Does it matter?

This is the question worth sitting with.

On one hand. If the content is accurate, useful, and well-structured, does it matter that it was generated by a model? Readers are increasingly comfortable with AI-assisted writing. The stigma is fading. Most people care about whether a piece answers their question, not whether a human typed every word. The output works. It gets the job done. Why should anyone care about the mechanism that produced it?

On the other hand. The uniformity is real. When every blog post, LinkedIn update, and marketing email sounds the same, same rhythms, same punctuation, same structural tics, something is lost. Writing becomes wallpaper. Background noise. The kind of content your eyes slide over without registering. The distinctiveness that makes a voice worth listening to gets smoothed away by statistical averaging.

Subscribe now

I think the em-dash epidemic isn’t a crisis of authenticity. It’s a crisis of distinctiveness. The markers of AI writing are a reminder that most people are using these tools to produce the average of everything ever written, and the average of everything is, by definition, unremarkable.

The fix isn’t to ban AI from the writing process. It’s to bring something to the process that the model can’t generate on its own: a point of view, a specific argument, a willingness to break the pattern. Give the machine something to work with, then edit what comes back. Push it off the median. Make it yours.

The em dash isn’t the enemy. Laziness is.

The Varkey Foundation doesn’t do marketing the way most organisations do. They run education programmes in Ghana, Uganda, Argentina. They fund schools, train teachers, build infrastructure. The work is real and the impact is measurable, but the audience for that impact is scattered across dozens of countries, multiple languages, and communities that are connected more by WhatsApp groups than by any media channel you’d buy ads on.

I came in for the 2017 cycle. Originally for technology tasks, but the remit grew fast. The gap was obvious: the foundation was doing serious work and almost nobody outside the room knew about it. The diaspora communities, the parents and alumni and supporters living in London, Dubai, New York, who cared deeply about education back home, had no way in. The foundation had the stories. It didn’t have the pipes.

The first thing I did was look at the website. Not redesign it, understand it. Where were people going? What were they clicking? What made them leave?

The answers were in the analytics, but nobody had been looking. Specific pages, programme pages, regional pages, were pulling real traffic. People were searching for the foundation’s work in Ghana. In Uganda. In Argentina. They were finding the site, landing on the right page, and then bouncing. The content existed but the experience was flat. No next step. No call to action. No reason to stay.

I rebuilt the user journey around those key pages. Goals were set. Links were restructured so that someone who landed on the Ghana programme page had a clear path to learn more, sign up, donate, or share. The website stopped being a brochure and started being a funnel. Not in the aggressive sense, in the “here’s what you can do next” sense.

Social was where the real work happened. The diaspora communities weren’t reading the foundation’s annual report. They were on Facebook, on Twitter, on WhatsApp. They were following pages about their home countries, sharing news about education, arguing about politics in group chats. That’s where the foundation needed to show up.

Subscribe now

I ran targeted promoted posts highlighting the foundation’s wins across all programmes. Not generic brand content, specific stories. A school that opened in northern Ghana. A teacher training cohort in Kampala. A scholarship recipient who’d gone on to university. The targeting was built around diaspora demographics, people in the UK or UAE or US who had connections to Ghana, Uganda, Argentina. The posts weren’t flashy. They were real. That’s what made them work.

For Argentina, I tried something different. The foundation had early-stage ventures there, and the South American market was hard to reach through traditional paid media. So I set up an automated social media tool, configured with a white and blacklist and keyword triggers. It monitored conversations about education in Argentina and engaged proactively with people who were already talking about the issues the foundation cared about. Not bots. Not spam. Relevant responses to real conversations.

The automated strategy outperformed the output of the South American agency the foundation had hired. Not because the agency was bad, but because automation doesn’t sleep, doesn’t miss posts, and doesn’t need three approval cycles to respond to a tweet.

The diaspora engagement piece was the one I was proudest of. These were people who cared about education in their home countries but had no channel to the foundation’s work. They’d see a news story about a school closing in Ghana and feel helpless. They’d hear about a programme in Uganda and want to support it but not know how.

By showing up in their feeds with specific, real stories, the foundation gave them something to respond to. Engagement went up. Sharing went up. People started tagging the foundation in posts about education in their home countries. The community grew organically because the content was relevant and the channel was native.

This wasn’t complicated strategy. It was plumbing. Connect the stories to the people who care about them, in the place where those people already are.

The sustained results meant that when I transitioned to focus on the foundation’s marquee events, the Global Teacher Prize and the Global Education and Skills Forum, the digital infrastructure was already working. The community was built. The channels were active. The automated systems were running. I didn’t have to start from zero for the big events because the ground game was already in place.

The Teacher Prize campaign was a different beast, bigger budget, more channels, higher stakes. But the principle was the same. Show the stories. Reach the people who care. Make it easy to act.

We ran campaigns across social, email, paid promotion, and print in the Guardian, the Times Education Supplement, and UAE publications. Every placement drove to a tracked landing page. The Prize received its highest-ever application numbers during the cycle. Website sessions grew 53 percent. Social acquisition jumped 336 percent.

But the number that mattered to me wasn’t the big spike. It was the sustained engagement. The community we’d built during the foundation’s ongoing programmes carried into the marquee events. The same people who’d shared a story about a school in Ghana were now sharing stories about Prize nominees. The pipeline worked because the foundation had invested in relationships, not just campaigns.

Subscribe now

What I’d do differently, looking back: start the Argentina automation earlier and expand it. It worked, but it was a pilot. If we’d rolled that approach across all the foundation’s markets, the compound effect would have been significant. Automated engagement at scale, properly configured, is a force multiplier that most non-profits haven’t tried.

I’d also build the diaspora targeting into a permanent capability rather than a campaign-specific tool. The community was there. The interest was real. But without ongoing investment, the engagement would fade between cycles. The foundation needed its own always-on channel to these communities, not one that only switched on when a campaign was running.

The biggest lesson was about plumbing versus polish. Most non-profits spend their marketing budget on polished content, annual reports, glossy videos, produced testimonials. The work I did at Varkey was mostly plumbing. Fix the website. Set up targeting. Build automation. Connect channels. The plumbing isn’t glamorous, but it’s what makes the polish reach anyone.

Someone stood up in front of a room, opened a slide deck, and described a problem. Supply chain transparency. Decentralised identity. Peer-to-peer energy trading. The problems were always real. Nobody disputed that.

Then came the second slide. A diagram of a blockchain with arrows pointing in several directions. Then the third slide: the token. It had a name, a symbol, and a total supply that was always in the billions, because the psychology of buying 10,000 of something feels better than buying 0.003 of something. Then the fourth slide: the raise. The soft cap. The hard cap. The timeline. And somewhere near the bottom, in smaller font, a link to the white paper.

The white paper was the product. That was the thing nobody said out loud. You didn't need a working prototype. You didn't need customers. You didn't need revenue, or a business model, or even a coherent plan for what happened after the money arrived. You needed a white paper that sounded plausible and a token that could be listed on an exchange.

ICOs have collectively raised over $1.2 billion. For context, the entire European VC market invested about €6.4 billion in startups in 2016. A fundraising mechanism that barely existed two years ago, backed by no regulatory framework, no investor protections, and no track record of delivery, is pulling a real share of early-stage capital.

The biggest proof of concept was The DAO, a decentralised investment fund built on Ethereum that raised $150 million in May 2016 before a recursive call bug drained $60 million of it. The community split the chain to get the money back, which created a permanent split in Ethereum. The original chain continues as Ethereum Classic, while the main chain rolled back the hack. A monument to the argument that code is law right up until the code costs you real money. Tezos raised $232 million in July. Filecoin looks like it's about to close on $200 million. Bancor raised $153 million in three hours. Each raise is bigger than the last, and each one normalises the next.

I've been close to this from the inside. I've been on the community and marketing side of things long enough to see the pattern from both sides, the builder side and the investor side. And the pattern is this: the people raising the money and the people building the product are increasingly not the same people.

A company I've been consulting for is being pushed hard toward an ICO. The advisors want it. The investors want it. Every comparable in the market is doing it. The pressure is enormous, and it isn't irrational. If everyone around you is raising $50 million on a white paper and you're still bootstrapping, the competitive math is brutal.

They haven't done it.

Subscribe now

The reasoning breaks down into three parts, none of them about fear.

The first is personnel. When an industry is two years old, everyone is an expert. The people available to design token economics, lead a launch, and build a community have mostly been doing something completely different twelve months earlier. Some were community managers at gaming companies. Some were freelance developers. Some were in completely unrelated fields and simply repositioned their LinkedIn profiles when the money started flowing. The expertise gap is obvious from the inside but invisible from the outside, because from the outside everyone has the right keywords in their bio.

When you're about to stake your company's future on someone's ability to design a sustainable token economy, you want more than twelve months of experience in a field that didn't exist before 2015. The talent market hasn't caught up with the capital market, and the capital market isn't waiting.

The second problem is structural. A crypto token is not a traditional financial instrument. It's not equity (you don't own a share of the company). It's not debt (there's no obligation to repay). It's not revenue share (there's no mechanism to distribute earnings). It's something that doesn't have a legal category yet, and every jurisdiction is still figuring out how to classify it.

The US SEC released its DAO Report just weeks ago, concluding that tokens sold in ICOs could be securities. The regulatory environment isn't just uncertain. It's a moving target, and the penalties for getting it wrong are existential, not just fines, but the possibility of having to return all the funds raised.

For a company trying to be a legitimate financial agent, issuing a token means navigating a regulatory environment that hasn't been mapped yet. Every lawyer gives a different answer. Every jurisdiction has a different interpretation. And the largest markets in the world are signalling that they're paying attention.

The third problem is the one almost nobody talks about publicly. Token communities are not the same as product communities. The people buying tokens want the token to go up. That's the fundamental incentive. They might believe in the project, they might use the product eventually, but the primary motivation is financial return on the token.

The people using products want the product to work. They want it to solve a problem they have. They don't care about the token price. They care about whether the thing does what it says on the tin.

These are different groups with different incentive structures, and building for one does not guarantee you get the other. An active token community doesn't mean you have users. A growing user base doesn't mean your token has value. The assumption that token holders will become users, or that users will buy tokens, is untested and probably wrong.

So the company is skipping it. Missing the wave. Watching peers raise tens of millions on documents and promises while they continue to build with the revenue they have.

Subscribe now

There's a question embedded in all of this that nobody is asking. If a company raises $50 million to build decentralised identity, what stops them from pivoting to something else entirely? The white paper says X. The market rewards X. But the team realises six months in that Y is more viable. Do they tell the token holders? Do token holders have any say? The answer, for now, is no.

Token vesting exists. Teams lock their own tokens for 12-24 months, releasing them gradually. But vesting on the funds themselves doesn't exist. Nobody is asking what happens when the treasury gets spent on something other than what was promised. The tokens vest. The money doesn't.

This is the structural problem that makes regulation likely. Not because crypto is inherently bad (the technology is sound, the concept is valid, the problem is real) but because the gap between what's promised and what's delivered has no mechanism for accountability.

The question is whether the industry builds its own guardrails before governments do it for them. Regulators are watching. The largest markets are starting to move. But the timeline is open, and the outcome isn't decided.

The companies choosing not to bite are making a bet. They're betting that the contradiction between what a token promises and what a company can deliver will eventually matter. Between what a community buys and what a product needs. Between what the market rewards and what actually creates value.

Whether they're right is anyone's guess. The companies that skipped it may have missed the wildest fundraising mechanism in startup history. Or they may be the ones still standing when the dust settles.

The ICO wave is an experiment. The results aren't in yet. The question is what we're learning, and whether the next wave of fundraising, whatever form it takes, applies those lessons or repeats them.

I’ve sat through three pitches this month where the strategy deck included the phrase “lookalike audiences.” Three. Different agencies, different clients, same slide. A funnel graphic. Top: awareness. Middle: consideration. Bottom: conversion. And somewhere in the middle, a bullet point that says “deploy lookalike audiences to scale efficiently.”

Every time I see this slide, I want to ask one question: lookalike of what?

Because here’s what nobody says out loud. A lookalike audience is only as good as the seed audience it’s built from, and most seed audiences are garbage. They’re built from email lists that haven’t been cleaned since 2014. They’re built from website visitors who bounced after three seconds. They’re built from people who clicked a Facebook ad because the image was bright and then never came back.

You’re scaling a guess. That’s not strategy. That’s volume.

The platforms love lookalike audiences because lookalike audiences spend money. Facebook’s algorithm takes your seed list, finds people who share demographic and behavioural traits, and expands your reach. Sounds smart. In practice, it means you’re paying to show ads to people who vaguely resemble the people who vaguely engaged with your brand once.

I ran the numbers on a campaign last quarter. The client had a seed audience of 12,000 email subscribers. We built a 1% lookalike on Facebook — roughly 200,000 people. The CPM was competitive. The click-through rate was fine. The conversion rate was 0.03%. That’s three conversions per ten thousand impressions.

We killed the lookalike. Instead, we built audiences from people who’d watched 75% of a video ad, people who’d added to cart but not purchased, and people who’d engaged with the page more than twice in 30 days. Three smaller audiences. Combined reach was maybe 40,000. The conversion rate jumped to 1.2%.

Smaller audience. Better results. The opposite of what the funnel slide says should happen.

Subscribe now

Here’s the thing nobody puts in the pitch deck. Lookalike audiences are a commodity product. Every agency uses them. Every brand uses them. Your competitor is building lookalikes from the same demographic data, targeting the same age range, in the same cities, with the same creative that has a stock photo of a diverse group of people laughing around a laptop.

If everyone is running the same playbook, nobody has an advantage. You’re just bidding against each other for the same attention, driving up your own costs, and calling it a media plan.

The agencies don’t care because they charge on spend. The platforms don’t care because they charge on impressions. The only person who should care is you, because you’re the one paying for it.

I watched a brand spend £80,000 on Facebook lookalike campaigns in Q1. Their CPA was £45. Their product sells for £29. Do the maths. They were losing money on every conversion and the agency was reporting “strong reach” as if reach pays salaries.

What works instead isn’t complicated. It’s just harder to put in a slide.

Build audiences from behaviour, not demographics. Someone who spent four minutes on your pricing page is worth more than someone who matches your customer’s age and postcode. Someone who opened three emails in a week is worth more than a thousand lookalikes who share her browsing habits.

Use your own data first. Your CRM. Your purchase history. Your email engagement. Your app usage. These are signals that Facebook and Google don’t have and can’t replicate. When you build audiences from these signals, you’re not competing on the same commodity data as everyone else. You’re working with something proprietary.

Then layer. Don’t build one audience and scale it. Build three or four specific audiences and test them against each other. One from purchasers in the last 90 days. One from email subscribers who clicked a link in the last 30 days. One from people who visited your site more than three times but never bought. Each one tells you something different about who your actual customers are.

Subscribe now

The lookalike can come later. Once you know who converts, once you have a seed audience built from real behaviour, then a lookalike has something to work with. But start with data you own, not data the platform guesses at.

The other problem with generic campaigns is creative. If your ad looks like every other ad in the feed, it’s wallpaper. Nobody stops scrolling for wallpaper.

I see the same mistakes every week. Headlines that say “Introducing our new range.” Nobody cares about your new range. They care about whether your product solves their problem. Body copy that says “Shop now and save 20%.” Every brand is offering 20% off. That’s not a message. That’s noise.

The ads that work are specific. They name the problem. They use language the customer actually uses, not language the brand guidelines dictate. They look like they were made by a person, not assembled by a template.

One of the best-performing ads I’ve seen this year was a photo of a product on a kitchen counter. No studio lighting. No lifestyle model. Just the product, on a counter, with a caption that said “This is the one that actually works.” It outperformed a £5,000 studio shoot by 4x on conversion rate.

The creative team hated it. It wasn’t on-brand. It didn’t follow the visual guidelines. It didn’t have the approved colour palette. But it sold product, which is the job.

I’m not saying lookalike audiences are useless. I’m saying they’re a tool, not a strategy. Deploying lookalikes without understanding your seed data is like buying a billboard without checking which road it’s on. You might reach people. They might not be the people you need.

The agencies will keep pitching the funnel slide. The platforms will keep recommending lookalikes. The dashboards will keep showing “reach” and “impressions” as if those are outcomes.

Your job is to ask the question nobody wants to answer. Who are we actually reaching? And are they buying?

If the answer is “we don’t know,” you don’t have a campaign. You have an expense.

They were called knockoffs. That’s the polite word for what they were. Two phones I’d bought in Beijing for about 80 US dollars each, sitting in my backpack while I crossed from Ghana into Togo, then Benin, then Burkina Faso, then Côte d’Ivoire, then eight more countries before I was done. They weren’t beautiful. They weren’t fast by any measure a phone reviewer would recognise. But they worked when nothing else did, and by the end of the trip I’d stopped thinking of them as knockoffs and started thinking of them as proof.

The proof wasn’t about the phones. It was about what happens when you build technology for the people who actually use it, instead of the people who review it.

I’d been in Beijing for twelve months. A year in China teaches you things about manufacturing that no Silicon Valley keynote ever will. You learn that the factory that makes the $800 phone and the factory that makes the $40 phone are sometimes the same factory. Same workers, same floor, same injection moulding machines. The difference is the bill of materials, the brand budget, and the margin the middleman takes. Remove all three and you get a phone that costs $40 and makes phone calls, sends texts, runs WhatsApp, and has a battery that lasts three days. Remove the middleman and the brand budget and you get something else entirely. You get a phone that was designed by someone who understood the problem.

The problem in West Africa in January 2016 was not processor speed. Nobody on a bus from Accra to Lomé was running benchmark tests. The problem was coverage. Five carriers in Ghana, four in Togo, three in Benin, and your phone needed to talk to all of them because none of them covered everything. The first thing I noticed about these phones wasn’t the screen or the camera. It was the two SIM slots. Both phones had them. Every phone sold at the stalls in Osu, at the market in Cotonou, at the roadside in Burkina Faso had them. Dual SIM wasn’t a feature. It was the cost of doing business.

An iPhone 6 in January 2016 cost about $650. A Samsung Galaxy S5, now a year old, was around $400 if you could find one at a legit shop. My Xiaomi knockoffs cost $40 each. That price gap wasn’t a gap. It was a different reality. It meant the difference between one person having a phone and five people sharing one. It meant a market trader in Kumasi could run two numbers, one for business and one for family, and not carry two devices. It meant a kid in Accra could get online for the first time.

The phone didn’t ask him to join an ecosystem. It asked him to charge it.

Subscribe now

The batteries were removable. I cannot overstate how much this mattered.

In London, in San Francisco, in any city where a charger is never more than six feet away, a sealed battery is a design choice. In northern Ghana, three hours from the nearest power outlet, a sealed battery is a death sentence for your phone. These Chinese phones had backs that popped off. You could buy a spare battery for a dollar at any market. You could carry two, swap them mid-day, and keep going. I did this constantly. The back cover would pop off on the bus, the battery would slide out, I'd shove it back in, snap the cover on, and the phone would boot back up. Try that with an iPhone. Try telling someone in a village without consistent electricity that their $650 phone needs to be sent to an authorised service centre because the battery won't hold a charge anymore.

The knockoffs had removable batteries because the people who designed them knew what conditions they'd be used in. Apple designed for the charger. These phones designed for the gap between chargers.

The screens were plastic. Not Gorilla Glass, not whatever Samsung was using that year. Plastic. When I dropped one face-down on the red dirt road outside Tamale, it bounced. Not gracefully. It bounced the way a rubber ball bounces when you throw it badly. The screen didn't crack. The corner got scuffed. I picked it up and kept walking. A friend on the same trip had an iPhone 5s. He dropped it from a lower height onto softer ground in Ouagadougou. The screen shattered. He spent the next week squinting at spiderweb cracks, trying to read WhatsApp messages through a constellation of fracture lines.

Dust was the other thing. The roads in West Africa are not paved in most of the interior. Red laterite dust gets into everything. It gets into your bag, your clothes, your lungs, and your phone. The knockoff phones dealt with it fine. The SIM tray never jammed. The charging port kept working. My friend's iPhone Lightning port started acting up after three days of dust. He'd blow into it like a Nintendo cartridge, and it would work for a while, then stop again.

I bought the phones at a market in Zhongguancun, the electronics district in Beijing. Zhongguancun in 2015 was a zoo. Floors of stalls selling everything from laptop batteries to surveillance cameras, with salespeople grabbing your arm as you walked past. The phones weren’t Xiaomi phones, technically. They were built in the same Shenzhen supply chain that Xiaomi operated in, running the same chipset architecture, using screens and batteries sourced from the same supply chain. The Xiaomi Redmi Note 2 had come out a few months earlier, August 2015. It had a MediaTek Helio X10 processor, a 5.5-inch 1080p screen, and cost about $125. My phones were cheaper versions of that idea. Same general shape, same dual SIM philosophy, same removable battery, but without the Xiaomi branding and at a third of the price.

Subscribe now

This is the part that confuses Western observers. In the West, “knockoff” means inferior. It means a scam, a fake, a thing designed to trick you into thinking you’re getting the real product. In China, and increasingly across Southeast Asia and Africa, a knockoff is something different. It’s a phone built by the same supply chain, using the same component vendors, but without the brand tax. The chipset might be a Snapdragon 410 or a MediaTek MT6582. Not top of the line. Not flagship. But capable of running Android 4.4, running WhatsApp, running Google Maps, and making phone calls on any network from Accra to Abidjan. That’s all the phone needed to do, and it did it.

The Chinese had a word for this world. Shanzhai. It started in the early 2000s with knockoff Nokia phones and grew into a parallel manufacturing universe. By 2015, the Shanzhai market was producing phones that weren’t just cheap alternatives. They were adaptations. Dual SIM for Africa. Extra-loud speakers for Indian markets. Flashlight functionality for Southeast Asia. The phones weren’t copying Apple. They were solving problems that Apple hadn’t noticed existed.

Transsion, a Shenzhen company you’ve probably never heard of, understood this better than anyone. They’d been in Africa since 2008, selling feature phones under the Tecno and Itel brands. By 2016 they were the largest phone manufacturer on the continent, and they’d done it by building phones with dark-skin-optimised cameras, longer batteries, and prices that made sense for African incomes. They didn’t sell these phones in London or New York. They didn’t need to. The market was in Abidjan and Nairobi and Dar es Salaam, and it was enormous.

I didn’t know about Transsion when I was buying my phones in Beijing. I just knew that the $40 phone worked and the $650 phone didn’t survive the trip.

Something else happened on that trip that I didn't expect. The knockoff phones connected me to people in a way that expensive phones don't.

When you pull out an iPhone in a West African market, there's a gap. The person you're talking to knows the phone costs more than they make in three months. It changes the dynamic. Not always, not dramatically, but it's there. A small distance. A reminder of what separates you.

When you pull out a $40 phone that looks like every other phone in the market, that distance disappears. I was just another person with a phone. Not a tourist with a status symbol. In Tamale, a guy at a phone charging station looked at my phone, looked at me, and asked where I got it. I said Beijing. He said his was from Accra but it was the same phone. We talked for twenty minutes about Android updates and which carrier had the best data rates in the north. He wasn't impressed by me. He was talking to me because we had the same phone.

Call it a technology insight if you want. The best technology disappears. It doesn't announce itself. It doesn't signal wealth or taste or sophistication. It just works, and then it gets out of the way.

The iPhone couldn't get out of the way. It was always there, always glass and aluminium, always asking to be noticed. The knockoff just sat in my hand and made calls.

The Snapdragon 410 in those phones was a four-core ARM Cortex-A53 running at 1.2 GHz. I know this because I checked, eventually. Not because it mattered. What mattered was that I could open Google Maps in Cotonou, find the road north, and get on a bus. What mattered was that WhatsApp worked on a 2G connection, which was the only connection available for most of the trip. What mattered was that the battery lasted through a twelve-hour bus ride with no charging opportunities, because I'd swapped in a fresh one at the last stop.

People in the West have a specific mental model of what a "good phone" is. It's fast, it's thin, it has a great camera, it costs $600-$1,000, and you replace it every two years. That model is a product of living in places with reliable electricity, ubiquitous WiFi, and incomes that can absorb the cost. Take away any of those conditions and the model falls apart.

In West Africa in 2016, the "good phone" was the one that survived. The one whose battery you could replace. The one that worked on every carrier. The one that cost less than a month's income. The one you didn't cry about when it fell in the sand.

My knockoff phones were good phones. By every measure that mattered on that trip, they were better than anything Samsung or Apple was selling.

Subscribe now

Eleven countries. Four months. Two phones that cost $80 combined. They survived dust, rain, drops from buses, nights in the desert with no power, and the general chaos of overland travel through West Africa. At the end of the trip, they both still worked.

I don’t have them anymore. I left one in a guesthouse in Nouakchott, not on purpose. The other one died somewhere in Morocco, months later, when the battery finally gave out after too many charge cycles and not enough care. I didn’t mourn them. They’d done their job.

But I think about them sometimes, when I see someone in London drop an iPhone and wince, or when I read about some new flagship that costs more than a laptop. I think about the factory in Shenzhen that made those phones. The same assembly line, probably, that was running Xiaomi and Huawei boards through the same soldering stations. The same workers who were building phones for Beijing and Accra, understanding implicitly that the phone for the Ivorian market needed something different from the phone for the Chinese market. Not worse. Different.

“Made in China” is a punchline in the West. Cheap, disposable, fake. In West Africa, in 2016, it meant available. It meant affordable. It meant someone had bothered to build a phone for the place I was standing in, instead of the place I was trying to leave.

Call it what you want. The phones were designed.

A client asked me last week how much Kim Kardashian charges for a single Instagram post. I told him the number that had been floating around the industry, somewhere north of $200,000. He did the maths in his head. Then he asked me how many units of his £35 product he’d need to sell to break even.

The answer was obvious. He couldn’t.

Half the brands I work with are doing this right now. They see influencer marketing as a channel. They budget for it the way they budget for Facebook ads or Google search. They pick an influencer with a big number, negotiate a rate, post the content, and wait for the sales to come in. When the sales don’t come in, they blame influencer marketing. When the sales do come in, sometimes, rarely, they can’t figure out why, so they do it again and hope. It’s a lottery ticket with better photos.

Subscribe now

Here’s what nobody tells you about celebrity influencers. Their audience isn’t your audience. Kim Kardashian’s 60 million followers are not 60 million potential customers. They’re 60 million people who want to look at Kim Kardashian. A fraction of them care about the product. A fraction of that fraction can afford it. A fraction of that fraction will buy it. By the time you’ve worked through those fractions, you’ve paid $200,000 to sell maybe 400 units of something.

I watched a beauty brand spend £120,000 on three celebrity influencer posts last quarter. The combined reach was enormous. The combined sales were less than their average email campaign. The email campaign cost them about £200 to send.

The problem isn’t that influencer marketing doesn’t work. The problem is that most brands are doing it backwards. They’re buying attention from the top down when they should be building it from the bottom up.

The influencers who actually drive sales are the ones nobody’s heard of. The ones with 3,000 followers. The ones who post about skincare routines in their bathroom mirror. The ones who review products honestly because they haven’t been paid enough to lie.

I’ve been experimenting with this for six months. Instead of paying one celebrity £10,000 for a post, I find fifty small accounts, 2,000 to 15,000 followers, and I send them product. Free. No contract. No brief. No “please post by Tuesday.” Just a box with a note that says “if you like this, talk about it.”

About 60% of them post. That’s thirty posts from thirty different accounts, each reaching 2,000 to 15,000 people who actually follow that person because they trust their taste. The content is authentic because it is authentic. Nobody told them what to say. Nobody reviewed their caption. Some of them love the product. Some of them think it’s fine. A few don’t post at all, and that’s okay too.

The combined cost is the product itself plus shipping. For a £15 product, that’s maybe £900 total. The reach is smaller than one celebrity post. The conversion rate is four to five times higher.

The maths work for a specific reason. Trust.

A celebrity influencer’s audience knows the post is paid. They’ve seen the “#ad” tag. They’ve watched the influencer promote a different product every week. The trust is diluted. The recommendation carries the weight of a billboard, not a friend’s suggestion.

Subscribe now

A small account’s audience doesn’t have that problem. When someone with 4,000 followers posts about a product they genuinely like, their audience reads it the way you’d read a text from a friend. “Oh, she uses this? I should try it.” The conversion isn’t driven by reach. It’s driven by trust. And trust doesn’t scale the way impressions do.

This is why seeding works better than buying. When you seed, when you send product with no strings, you’re investing in a relationship, not a transaction. The influencer who loves your product will post about it again. And again. Not because you’re paying them. Because they like it. That’s worth more than any single sponsored post.

The timing matters too. Most brands approach influencers after the influencer is already big. They’re paying a premium for someone else’s audience. The smarter play is to find them before they’re big. The person with 2,000 followers today might have 50,000 in a year. If you seeded them early, you’re in their story. You’re the brand they discovered before anyone else cared. An ad fades. A relationship compounds.

I’ve been tracking a skincare brand that does this consistently. They don’t have a big budget. They don’t have celebrity endorsements. What they have is a spreadsheet of 300 small accounts that they’ve been seeding product to for two years. Some of those accounts are now at 30,000, 50,000 followers. The brand never paid for a single post. But those influencers keep using the product because they’ve been using it since before anyone was watching.

That brand’s cost per acquisition is a third of their competitors’. Not because they’re smarter at ads. Because they were early.

The industry will catch up to this eventually. Right now, the money is still chasing big numbers. Brands are still measuring influencer success by follower count. Agencies are still charging commissions on celebrity deals because the commissions are bigger.

But the shift is coming. Instagram is growing fast and the accounts that build genuine followings early will have staying power. The platform is favouring consistency and engagement, and the brands that figure this out first will have a structural advantage.

You don’t need a celebrity. You need fifty people who actually like your product. Find them. Send them the thing. Let them talk.

The rest is noise.

Fresh off a year in Beijing, I had yuan I couldn’t spend. Not in Beijing, where I’d spent twelve months learning how Chinese commerce worked at street level. Not in London, where I’d been before that. And certainly not in Accra, where I’d just arrived, four countries into an eleven-country overland trip through West Africa.

He had cedis he couldn’t send home. A man running a Chinese restaurant in Ghana, earning money in cedis, with a family in China that needed yuan. The money existed. It was real, physical, in his hands. But the systems that were supposed to move it (banks, remittance companies, telcos) had either decided his amount wasn’t worth the friction or had never built the infrastructure to handle it in the first place.

We swapped. Cash for cash across a table. No receipt, no exchange rate, no middleman, no app. Two people solving a problem that shouldn’t have existed but did.

I’d tried the banks first. In Accra, the ones with international logos on the door looked at the yuan like it was a curiosity. Some wouldn’t touch it. Others offered rates so bad you’d lose a third of the value just on the conversion. One teller told me to try the forex bureau down the street. The forex bureau told me to try the bank. Nobody wanted yuan in Ghana.

I’d tried to spend it directly. There was no option. No shop, no restaurant, no vendor accepted yuan. The currency existed but it had no reach. It was locked inside the Chinese financial system, and I was outside it.

The man in the restaurant had the opposite problem. His cedis were worthless in China. He could earn them, spend them locally, save them even. But the moment he needed to send money home, every channel charged him for the privilege of using their infrastructure. Western Union. MoneyGram. The bank wire that took five days and cost $30 on a $200 transfer. He was paying a tax on being away from home.

Bitcoin was at $430. I know this because I was following it, the way you follow something that’s interesting but not yet urgent.

Ethereum had launched six months earlier, in July 2015, and people were still arguing about whether Vitalik Buterin’s smart contract platform was a serious innovation or an elaborate science experiment. Coinbase had raised $75 million and was becoming the main way Americans bought crypto. Revolut had just launched in London, a beta app for people who wanted to dodge FX fees. Monzo and N26 were in their earliest stages.

None of these things had touched West Africa. None of them were designed to.

Subscribe now

The crypto world was small enough in January 2016 to follow in a single Twitter feed. Everyone in it was talking about the same thing: moving value without permission. Without institutions. Without the permission of banks and governments and the entire infrastructure of traditional finance.

But I was sitting in a restaurant doing exactly what they were theorising about. Except I was doing it with my hands, because there was no other option.

The mobile money story in West Africa was real but limited. M-Pesa had been in Ghana for a few years, running through Vodafone. It worked. If you had a Vodafone SIM and the person you were sending to also had a Vodafone SIM, and you were both in Ghana, and you were both near agents, you could send money across town in seconds. It was a breakthrough for domestic payments. People who’d never had bank accounts could now receive salaries, pay bills, send money to family in the next district.

But cross-border was a different planet.

Western Union had an agent on every high street. You could walk in, fill out a form, and send money to 200 countries. The fees were brutal, double-digit percentages for small amounts, which was most of the transactions in West Africa. A $50 transfer could cost $8-10 in fees. And the recipient had to physically go to a Western Union agent to collect it, which meant a bus ride if you weren’t in a city.

MoneyGram was similar. Banks offered wire transfers if you had an account (which most people didn’t), and if you could handle the paperwork (which most people couldn’t), and if you were willing to wait 3-5 business days and pay $25-40 per transfer.

The problem wasn’t that nobody was trying. Everyone was trying, but within their own system. M-Pesa worked within the Vodafone network. Western Union worked within its agent network. Banks worked within correspondent banking relationships. None of these systems talked to each other, and none of them were designed for the person sitting in a restaurant in Accra trying to get yuan to a family in China.

The crypto community was building for this problem. They just didn’t know it yet, and neither did I.

The term “DeFi” didn’t exist. Nobody had coined it. The idea of decentralised finance as a category (lending, borrowing, trading, payments, all without intermediaries) was years away from being named. In January 2016, people talked about “permissionless finance” or “programmable money” or just “what Ethereum might become.” The vocabulary hadn’t settled.

But the problem they were describing was the problem I was sitting in. Value that couldn’t move across borders. Systems that worked domestically but failed internationally. People with money they couldn’t use where they needed it.

I wasn’t theorising. I was four countries into a trip that had started in Côte d’Ivoire and would take me through Liberia, Sierra Leone, Guinea, Senegal, Gambia, Mauritania, and Morocco. I was hitchhiking, couchsurfing, running out of cedis/CFA because the nearest working ATM was sometimes two days behind me. Every transaction was a negotiation.

Subscribe now

I walked out of that restaurant thinking the world was broken in ways I hadn’t seen before. Not dramatically. Not as an epiphany. More like a quiet recognition that the things I’d taken for granted (the ability to send money, to pay for things, to move value across distance) didn’t exist for most of the people I was meeting.

Bitcoin was at $430. Ethereum was six months old. The people building the future of money were in San Francisco and London and Singapore. The people who needed it were in Accra, and Nouakchott, and Brikama.

That gap wasn’t going to close on its own. M-Pesa had proven that mobile money could work domestically. Western Union had proven that cross-border remittance had demand. But the space between those two things, fast, cheap, cross-border value transfer without a bank account, was empty.

Every institution was solving its own piece. Vodafone solved domestic transfers within their network. Western Union solved cross-border for people who could reach an agent. Banks solved wire transfers for people who had accounts. Nobody was solving the whole thing, because the whole thing required cooperation between institutions that had no incentive to cooperate.

The crypto community was talking about this. Not in those specific terms. They were talking about permissionless finance and decentralised networks and smart contracts. But underneath the jargon was the same problem: how do you move value between two people when the institutions that are supposed to make that possible have decided they’re not worth the effort?

The answer, in January 2016, was a table in a Chinese restaurant in Accra. Cash for cash. No permission needed.

The distance between those two groups was the distance that mattered.

I haven’t opened a web browser on my phone in three weeks.

I didn’t plan this. It happened because everything I need is inside one app. I pay for lunch with it. I book a taxi with it. I pay my rent through it. I order food, buy train tickets, split bills, send money to friends, check my bank balance, book a doctor’s appointment, and read the news. All inside WeChat. The browser sits on my home screen gathering dust.

This isn’t normal if you live in London or San Francisco. In those cities, you use twelve apps for twelve things. Uber for taxis. Deliveroo for food. Monzo for banking. WhatsApp for messages. Facebook for news. Each one does one thing and does it well enough, and they talk to each other through APIs and deep links and notifications that you manage like a second job.

In Beijing, there’s WeChat. And then there’s everything else, which doesn’t matter much.

WeChat launched in January 2011. Tencent built it. If you don’t know Tencent, they’re the company behind QQ, which was China’s dominant messaging platform for a decade before smartphones happened. QQ was desktop-first. WeChat was mobile-first. Tencent saw the shift before most Western companies had finished arguing about whether apps were a fad.

Subscribe now

By late 2015, WeChat has somewhere north of 600 million monthly active users. That number is hard to contextualise because it’s bigger than the population of any country except China and India. But the number isn’t the point. The point is what those 600 million people do inside the app.

They don’t just message. They live.

Here’s a normal day for someone in Beijing. Wake up. Check WeChat for messages from overnight, group chats from friends, notifications from official accounts you follow. Order breakfast through a mini-program inside WeChat, a food delivery service that runs without leaving the app. Pay for it with WeChat Pay, which deducts from your linked bank account instantly. Walk to the subway. Scan a QR code at the gate to pay the fare. WeChat again.

At work, your company communicates through a WeChat group. Not Slack. Not email. A group chat. Decisions happen in real time. Files get shared. Deadlines get moved. The group chat is the office.

Lunch. Open WeChat. A restaurant nearby has posted a deal on their official account. You tap it, order, pay, and walk in to pick it up. No cash. No card. No separate app. Just WeChat.

After work, you’re meeting a friend. You share your location inside the chat so they can find you. You split the bill through WeChat Pay. You book a movie ticket through a cinema’s official account. You buy a birthday gift from a WeChat shop that a friend set up last month, selling handmade candles from her apartment.

None of this requires leaving WeChat. None of it requires a web browser.

The Western frame for this is “messaging app.” That’s what WeChat looks like if you open it for the first time. A chat interface. Contacts list. Green bubbles. Fine.

But the chat interface is the front door, not the house. Behind it is a payments system that processes more transactions than most banks. Behind that is an app platform where businesses build services that run inside WeChat without needing to be downloaded separately. Behind that is an identity system. Your WeChat ID is becoming more useful than your passport for everyday life in China.

A friend of mine runs a shop in Beijing. No website. No physical store. She sells through WeChat. Her customers find her through word of mouth in group chats. They pay through WeChat Pay. She ships through a logistics service she found on WeChat. Her entire business runs inside an app that most people in the West think of as “Chinese WhatsApp.”

Subscribe now

The comparison is lazy, but revealing. WhatsApp sends messages. WeChat runs a business. Both are chat apps. One understood what a chat app could become. The other didn’t.

The payments piece is what makes it work. Without WeChat Pay, WeChat is a messaging app with some nice integrations. With WeChat Pay, it’s an operating system for daily life.

WeChat Pay launched in 2013. By 2015, it’s everywhere. QR codes on restaurant tables. QR codes on vending machines. QR codes on beggars’ signs, which is a thing that actually happens in Beijing and tells you something about how deeply the technology has penetrated. A beggar with a QR code is a beggar who knows that nobody carries cash anymore.

The bank transfer system in China was slow and expensive before mobile payments. Sending money to a friend meant a bank transfer that took a day and cost a fee. WeChat Pay does it instantly, for free, inside a chat. You tap a button, type an amount, and the money moves. Your friend gets a notification. Done.

This sounds trivial until you realise what it does to behaviour. When sending money is as easy as sending a text, people send money. They split bills without thinking. They pay back small debts immediately. They buy things from friends who sell things. They tip service workers through QR codes. The friction between “I owe you” and “here you go” disappears.

That friction was the banking system’s business model. WeChat Pay removed it.

The official accounts are the other piece. Every business, every brand, every government department, every restaurant and shop and hospital has a WeChat official account. These aren’t social media profiles. They’re service channels. A hospital’s official account lets you book appointments, check wait times, view test results, and pay your bill. A restaurant’s account shows you the menu, lets you order, and sends you a receipt. A government account handles your visa application, your tax filing, your residence registration.

In the West, this would be a website. Or an app. Or a phone call. In China, it’s a WeChat account. And because it’s inside WeChat, it benefits from the network. Your friend shares the hospital’s account with you. You tap it. You book. You never leave the app.

The network effect here is different from what Facebook or Google understand. Facebook’s network effect is about content. Google’s is about information. WeChat’s network effect is about transactions. Every service that joins makes every other service more useful, because they all share the same payment system, the same identity, the same chat interface. The hospital and the restaurant and the cinema aren’t connected by links or search results. They’re connected by the same person, in the same app, doing all three things in the same afternoon.

People in the West keep waiting for “the WeChat of the West.” They’ve been waiting since 2013. Facebook tried with Messenger. It added payments, then removed them. It added business pages, then buried them. It added bots, then nobody used them. Apple has iMessage, but iMessage only works on iPhones, which means it can never be the universal platform WeChat is in China. Google has whatever Google is calling its messaging app this year.

None of them will get there. The problem isn’t technology. The problem is that Western companies think in products and WeChat thinks in surfaces. A product does one thing well. A surface lets you do anything on it. WeChat is a surface. The chat interface is just the canvas.

The deeper problem is regulation and incumbency. In China, WeChat filled a gap. There was no entrenched banking system to displace, no established e-commerce platform to fight, no regulatory framework that protected existing players. The mobile payments revolution happened in China because the old system was slow, expensive, and inaccessible to hundreds of millions of people. WeChat didn’t compete with a good system. It replaced a bad one.

Subscribe now

In the West, the banking system works. Credit cards work. Apple Pay works. The entrenched players are competent and well-regulated. There’s no gap for WeChat to fill because the gap doesn’t exist. The West doesn’t need a WeChat. It needs its existing tools to talk to each other better, which is an integration problem, not an invention problem.

But the model matters. Even if the West never gets its own WeChat, the idea that a messaging interface can be an operating system for services is going to spread. It’s already happening in Southeast Asia. Line in Thailand and Japan. KakaoTalk in South Korea. Grab in Singapore. These apps are all building the same pattern, messaging at the centre, payments underneath, services around the edges.

The future of software isn’t another app. It’s one app that does everything, with a chat window as the front door.

I haven’t opened my browser in three weeks. I might never open it again.

There's an escalator in Hong Kong that runs for 800 metres uphill through the middle of a city. It starts in Central, the financial district, and it climbs through Soho, past restaurants and bars and apartment buildings, up into the Mid-Levels where the expats and the wealthy locals live. It takes about twenty minutes to ride the whole thing. It's the longest outdoor covered escalator system in the world, and it was built in 1993 to solve a simple problem: too many people trying to walk uphill in the heat.

I ride it almost every day. Not because I live in the Mid-Levels. I don't. I ride it because the neighbourhood it passes through is the most efficient classroom I've ever sat in.
Business school teaches you frameworks. Porter's Five Forces. SWOT analysis. The BCG matrix. You learn to categorise problems into quadrants and apply models that someone else built for a different situation twenty years ago. The frameworks are clean. The spreadsheets balance. The case studies have endings.

Hong Kong doesn't work like that.

The escalator passes through maybe fifteen different economic zones in twenty minutes. At the bottom, in Central, you've got the glass towers. HSBC, Bank of China, Standard Chartered. The air conditioning blasts out of every lobby. Men in suits walk fast and don't look up. This is the Hong Kong that the Economist writes about. Finance. Capital. The free market at work.

Ride up three minutes and you're in a different country. The buildings get shorter. The shops get smaller. A woman is hanging laundry on a bamboo pole that sticks out over the escalator. Below her, a man is gutting fish on a stainless steel table on the pavement. The fish came from the wholesale market at 4am. He'll sell it by noon. The margins are thin and the hours are brutal and he's been doing it for thirty years because his father did it before him. Nobody taught him Porter's Five Forces. He knows which fish sell on Tuesday and which sell on Saturday. That's his framework.

Keep riding. The restaurants start. Not the ones with Michelin stars, though those are nearby. The ones with plastic chairs and laminated menus and a cook who does one dish, maybe two, and does them so well that people queue at the door. A noodle shop. A congee place. A dai pai dong that's been operating out of the same metal stall since 1978. The woman who runs it is sixty-something. She doesn't have a website. She doesn't need one. Her marketing strategy is a pot of broth that's been simmering since yesterday.

Subscribe now

This is what the escalator taught me that business school couldn't. Not because business school is bad, but because business school teaches you to think in systems, and the escalator teaches you to think in people.
The escalator was a government project. That’s the first thing that surprised me. Hong Kong, the poster child for free-market capitalism, built a public escalator to help commuters get uphill. The logic was straightforward. Traffic was bad. The streets in Central are narrow and steep. Buses couldn’t handle the grade. Walking in 35-degree heat with 90% humidity was miserable. So the government built an escalator.

The cost was about HK$240 million, roughly US$30 million at the time. For a piece of public infrastructure, that’s nothing. A highway interchange costs ten times that. But the escalator didn’t just move people. It created a neighbourhood. Before the escalator, the area between Central and the Mid-Levels was a dead zone. Steep streets, old buildings, not much reason to stop. After the escalator, the foot traffic multiplied. Restaurants opened. Bars followed. Property values along the route climbed. Soho, south of Hollywood Road, became a destination

Nobody planned this. The government built the escalator to solve a commuting problem. The market solved everything else. Call it a libertarian argument if you want. What actually happened is simpler: you remove a friction point, and humans do what humans do.

The escalator didn’t build a neighbourhood. It removed a hill. Everything else was already there, waiting.

There's a timing thing the escalator does that most people don't notice. In the morning, it runs downhill. Commuters in the Mid-Levels ride it down to Central for work. In the afternoon, after 10am, it reverses. It runs uphill. The direction flips based on when people need it.

I find this fascinating because it's the opposite of how most infrastructure works. A road doesn't change direction. A train runs on the same track both ways, but the schedule is fixed months in advance. The escalator just flips. It adapts to the flow instead of forcing the flow to adapt to it.

A product person would recognise this immediately. The escalator was designed around user behaviour, not around the engineering. The engineers could have built two escalators, one up and one down, and called it done. That's the obvious solution. Instead, they built one and reversed it. Half the cost. Same outcome. Better, actually, because it concentrated foot traffic in one direction at a time, which made the street-level businesses more predictable.

This is the kind of decision that doesn't show up in a case study because it looks too simple. Where's the five-forces analysis? Where's the competitive advantage? There isn't one. There's just someone who watched how people walked and built accordingly.

Hong Kong is in a strange place politically right now. The Umbrella Movement happened a year ago, in late 2014. Occupy Central shut down major roads for 79 days. The city is still processing what happened. Conversations about democracy and autonomy and Beijing's influence are everywhere, if you know where to listen.

The escalator doesn't care about politics. It just runs. Up in the morning, down after 10am. The woman at the dai pai dong doesn't care about politics either, at least not during service hours. She cares about whether the broth is hot and whether the noodles are right. The fishmonger cares about his margins. The bar owner in Soho cares about whether the expats will come back tonight.

Business school teaches you to think about markets as abstract systems. The escalator teaches you that markets are people, and people have other things on their mind besides markets.

I'm here for work. Advertising. Digital. The kind of work where you sit in a glass room and talk about "engagement" and "reach" and "consumer behaviour" while the actual consumers are downstairs, on the escalator, going somewhere that has nothing to do with your campaign.

The escalator humbles me. Not in a dramatic way. In a slow way. Every day I ride it and see something that contradicts something I've been taught. A business that shouldn't work, working. A strategy that shouldn't scale, scaling. A shop with no sign that had more repeat customers than anything in my portfolio.

Subscribe now

The escalator breaks sometimes. It’s old. The mechanical parts wear out. When it stops, people walk. They complain, sure. Hong Kongers complain about everything. But they walk. The street doesn’t die when the escalator stops. It just gets harder.

I think about this when I see startups panic about downtime. Your server goes down for an hour and you lose your mind. The escalator goes down for a week and the restaurants keep cooking. The fishmonger keeps gutting. The dai pai dong keeps simmering.

Infrastructure matters, but it’s not the business. The business is the broth. The business is the fish. The business is the woman who knows your order before you sit down. The escalator is what gets people there. If it stops, they find another way.

That’s the MBA lesson, if there is one. Don’t confuse the infrastructure with the value. The value is the thing people come back for. Everything else is how they get to it.

I'll leave Hong Kong eventually. Go back to Beijing. But the escalator will stick longer than anything I learned in a classroom.

It's an escalator. It goes up and down and it breaks sometimes and people ride it because it's there and it's easier than walking. Nothing more.

But it's teaching me to watch before I strategise. To see how people actually move, not how I think they should move. To build for the flow, not against it. To understand that the best infrastructure disappears. Nobody on the escalator is thinking about the escalator. They're thinking about lunch, or getting home, or the meeting they're late for. The escalator just lets them think about those things instead of the hill.

That's what good technology does. That's what good anything does. It removes the hill and lets people get on with it.

I turned off my phone somewhere over the Indian Ocean. Not airplane mode. Off. The screen went black and I slipped it into the bottom of my backpack, underneath an old book I’d been meaning to read and a wad of Thai baht I'd picked up at Heathrow. I wouldn't turn it on again for thirty-one days.

This was not an accident. I'd been planning it since September, when I finished up my role at a community and marketing team in London and realized I hadn't been fully offline for more than twelve hours in about two years. The longest I'd gone without checking email was a transatlantic flight, and even then I was composing replies in my head, ready to fire them off the moment the wheels touched down.

I told people I was going to Southeast Asia to take photographs. That was true. Angkor Wat at sunrise, the jungle trails outside Chiang Mai, the impossible green of rice paddies in the Cambodian countryside. I had a new camera body and two lenses I barely knew how to use. Photography was the excuse. The real thing I was after was harder to explain.

I wanted to feel what it was like to not know anything.

Not in the abstract, Silicon Valley sense where founders talk about "beginner's mind" over artisanal coffee. I mean actually not knowing. Not knowing how to get from Siem Reap to Bangkok. Not knowing what time it was. Not knowing where to eat, how much things cost, whether I was being ripped off or welcomed. Stripping away every digital crutch I'd built my life around and seeing what was left.

The first week nearly broke me.

Siem Reap is a small town that exists almost entirely because of Angkor Wat, but it doesn't feel small when you're standing on a dusty road at 6 AM trying to figure out how to hire a tuk-tuk for the day. There's no app for this. There's no surge pricing indicator. There's a guy standing next to a motorbike with a carriage welded to the back, and he's looking at you, and you're looking at him, and the whole transaction is going to happen with gestures, a calculator, and whatever broken English he speaks mixed with the zero Khmer you speak.

So you walk up. You point at a map. He says a number. You say a lower number. He laughs. You laugh. He comes down. You go up. You meet somewhere in the middle and spend the next twelve hours together, him driving you through temple complexes so old the trees have swallowed the walls, you sitting in the back with a camera in your lap and sweat running down your spine.

I kept thinking about Uber. I'd used it a few times in London before I left. The entire interaction compressed into a tap and a pin on a map. No negotiation. No eye contact. No relationship at all, really, except a transaction and a rating afterward.

Subscribe now

Standing there on that road in Siem Reap, bargaining with my tuk-tuk driver whose name was Sokha, I realized that app had taken this moment, the human moment, and removed it. Not because the moment was broken, but because it was inefficient. And the thing was, Sokha and I ended up spending three days together. He showed me temples the tour groups skip. He introduced me to his cousin who ran a noodle stall near the river. He told me about his daughter and how he wanted her to go to university.

None of that happens in the app. None of it can.

I moved north into Thailand after two weeks in Cambodia. Chiang Mai first, then deeper into the mountains toward the Myanmar border where the roads get bad and the tourists thin out. I was staying in guesthouses that cost four or five dollars a night, places with thin walls and roosters outside the window and shared tables where you ate dinner with whoever else was staying there.

This is how I met people. Not through a feed. Not through a group chat where someone posts a link and three people react with emoji. You sat down, you ordered whatever the family was cooking that night, and you talked to the person across from you. A Dutch woman who'd been cycling through Asia for six months. A retired teacher from Melbourne. A Thai university student home for break who wanted to practice his English and tell me about American rap music.

The conversations were slow. Unguarded. You had nowhere to be because nowhere was calling you. There were no notifications pulling your attention to a screen. You just sat and talked and sometimes the silences lasted a long time and that was fine too.

Back in London I'd been managing community channels where the whole point was connection at scale. Group conversations. Forums. Social platforms where hundreds of people could interact simultaneously. The conversations were good, they were useful, but sitting here I could feel what was missing from them. The pauses. The body language. The way someone leans forward when they're about to say something they actually care about. The shared meal that creates a context for honesty that no threaded discussion can replicate.

A group chat is a simulation of this. A good one, sometimes. But a simulation.

In Hanoi I found the town square.

It wasn't literally a town square. It was a series of them, actually. Little parks and patches of concrete where people gathered every evening as the heat broke. Old men played chess. Children chased each other around fountains. Women sat in clusters and talked with a rapidity that made my head spin. And the news moved through these spaces the way water moves through soil, slowly, person to person, from one knot of people to the next.

Someone would hear something from a relative who worked at the port. A rumor about prices. A story about the government. A bit of gossip about someone's daughter. It would pass from group to group, gaining details and losing others, shaped by each teller. By the end of an evening the whole neighborhood had a shared picture of what was happening, built entirely from oral transmission and face-to-face contact.

Twitter had gone public a month before I left London. I'd been on it for years. But watching these evening gatherings in Hanoi, I understood what it was actually doing. It was recreating the town square at global scale. The rapid exchange. The gossip and rumor mixed with fact. The way a story could spread from one cluster to the next in minutes instead of hours. The collective sense-making that happened when enough people shared what they knew.

The difference was that in Hanoi everyone could see each other's faces. You knew who was trustworthy and who was a blowhard. You could read the room. Online, you can't read the room. There is no room. There's just the feed.

The markets were where I found the last piece.

Subscribe now

I'm good at bargaining. It's what I do. BD, sales, negotiations. This is my bread and butter. But I'm used to doing it in English, in meeting rooms, with data backing my position. In the markets of Chiang Mai and Hanoi there is no fixed price. There are no tags. Every transaction is a negotiation, and the price you pay depends on who you are, how many you're buying, what time of day it is, whether the vendor likes your face, and a dozen other variables I couldn't begin to catalog.

I watched a woman buy fish at a morning market in Hanoi. She squeezed the fish, smelled it, argued for two minutes, walked away, came back, and paid a price that I suspect was exactly what she'd intended to pay from the beginning. The whole walk-away was theater. Both of them knew it. The fish vendor adjusted his price based on her body language, her persistence, the fact that she clearly knew fish and he wasn't going to get away with a tourist markup. Dynamic pricing in its purest form. No algorithm needed. Just two humans reading each other.

Uber again. The app that adjusts prices based on demand. That's what this market was doing, except the algorithm was a sixty-year-old woman with thirty years of fish-buying experience encoded in her brain. She was better at it than any software I'd seen.

I don't want to romanticize this. There's a difference between sharpening your skills and needing them to eat. I'm not here to judge how these systems work. I'm here to understand what the tech abstracts when it replaces them.

Something is lost when every human interaction gets optimized. Sokha and I never would have become friends through an app. The Dutch cyclist and I never would have shared a meal if we'd both been staring at our phones. The evening gatherings in Hanoi would lose their character if you could just scroll a timeline instead.

The building blocks of the tech I'd spent the last few years working with didn't come from code. They came from these places. From the way humans have always moved through cities, shared information, bought and sold, talked to strangers. The tech just abstracted the patterns. Stripped away the friction and the context and the mess and left behind something clean, something scalable, something that works.

That's not nothing. Clean and scalable and working matters enormously when you're trying to connect a billion people. But standing in the jungle outside Chiang Rai, camera in hand, no phone, no connection to anyone I knew, I understood something I couldn't have understood from a conference talk or a product demo.

The patterns came from here. From a tuk-tuk negotiation and a shared dinner and a town square at dusk. And when you abstract a pattern, you gain speed and scale, but you lose the texture. The sweat. The laughter. The five-minute silence between two strangers who haven't yet decided if they want to talk. The fish that smells wrong and the vendor who knows you know.

I don't have an answer for this. I'm not arguing we should go backward. I'm not writing a manifesto against technology from a blog that only exists because of technology. I'm just saying I spent a month not using any of it, and when I came back, I could see both what it gives us and what it takes.

That seems worth knowing. Especially if you're the one building it.

I turned my phone back on in the Bangkok airport on December 17th. Fourteen hundred emails. A hundred and twelve notifications on various platforms. Three messages from people who genuinely needed to hear from me.

Everything else could wait. And had. For thirty-one days, the world had gone on without me feeding it updates about my location and my thoughts and what I'd eaten for breakfast. It managed fine.

The flight home was eleven hours. I didn't sleep. I sat with my notebook and wrote down everything I could remember. Not the temples, though I remembered those too. The moments. Sokha's laugh when I tried to bargain in Khmer. The taste of the noodle soup from the first guesthouse. The sound of Hanoi at night, that particular density of motorbikes and voices and cooking smells that no recording has ever captured.

The notebook is still on my desk. I keep opening it. Not for nostalgia. For calibration. When I'm deep in a product conversation or a growth strategy or a community management problem and I start to feel like the abstraction is the thing itself, I go back to those pages and remember.

The thing itself is the woman buying fish. The thing itself is Sokha pointing at a temple I would never have found on my own. The thing itself is a table full of strangers eating dinner together because that's what you do when nobody has anywhere else to be.

Everything we build is just a translation. The question is whether we remember what we're translating from.

In 2010, two researchers at HP Labs published a paper. Sitaram Asur and Bernardo Huberman. They took Twitter chatter about upcoming films, ran it through a model, and predicted box office revenue with startling accuracy. The volume and tone of conversation before a film opened told them how much money it would make. Not focus groups. Not tracking studies. Tweets.

The paper got cited a lot in academic circles. In the marketing departments of actual companies, it changed nothing. Three years later, the average brand is still using social data to count likes and calculate engagement rate. They’ve built dashboards that tell them what happened last week. They’re sitting on a dataset that could tell them what’s going to happen next month, and they’re using it as a rearview mirror.

Google Flu Trends launched in 2008. The idea was simple. Track what people search for, map it against flu outbreaks, predict where the flu is going before the CDC can. For a few years, it worked. Then it stopped working.

In the 2012-2013 flu season, Flu Trends overestimated the prevalence of flu by more than 50%. It predicted that nearly 11 percent of the US population had influenza at the peak in January. The actual number was half that. Media coverage of flu was driving searches. People who didn’t have the flu were googling “flu symptoms” because the news told them to. The signal got swamped by its own echo.

This is the story everyone tells now when they want to argue that social data is unreliable. They’re wrong about the lesson. The lesson isn’t that social data can’t predict. The lesson is that you have to understand what you’re measuring. Google was measuring search interest in flu. It thought it was measuring flu. Those are different things.

Subscribe now

The same mistake is everywhere in marketing. Brands measure “engagement” and think they’re measuring interest. They measure “sentiment” and think they’re measuring opinion. They measure “reach” and think they’re measuring attention. The metric is not the thing. It never was.

I worked with a fashion retailer last year. They had three years of social conversation data about their products. Thousands of posts per week. Comments, shares, complaints, requests. They were using all of it to generate two numbers: engagement rate and sentiment score. Two numbers from three years of behavioural signal.

We pulled the raw data apart instead. What were people actually saying? Not the sentiment score, the words. It turned out that when conversation about a product shifted from “I want this” to “where can I find this,” purchases followed within two weeks. When conversation shifted from praise to frustration about sizing, returns followed within a month. The pattern was in the language, not in the sentiment score. The sentiment score couldn’t see it because the sentiment score treats “I want this so bad” and “where can I find this” as equally positive. They’re not the same signal. One is desire. The other is intent.

Intent is predictive. Desire is noise. The dashboard couldn’t tell them apart.

The prediction use cases aren’t theoretical. They’re happening now, mostly in industries that don’t think of themselves as social media companies.

A cluster of posts about stomach complaints in a specific city is a public health signal. People tweet symptoms before they see a doctor. The data is there. It just hasn’t been wired into the system that could use it.

A spike in complaints about a specific product feature on social media predicts what will get returned in stores next month. A spike in photography and sharing of a specific item predicts what will sell out next season. The fashion industry still sends buyers to trade shows and trusts their intuition. The data about what people actually want is already public, free, and updating every second. Nobody’s reading it.

After Nate Silver called every state in the 2012 election, every political operation in America started paying attention to social data. But they used it for messaging. Buy ads. Target demographics. Shape narratives. The prediction signal is different. It’s in what people share, what they stop sharing, who they start arguing with. Network structure predicts voting behaviour better than content does. Volume predicts turnout. Almost nobody in politics is reading the data this way yet.

Subscribe now

The problem is structural. Social data lives in the marketing department. Marketing departments think in campaigns. Campaigns have start dates, end dates, budgets, and KPIs. Social data doesn’t care about any of that. It’s continuous. It’s messy. It doesn’t fit into a monthly report because it updates every second and the interesting patterns take months to emerge.

To use social data for prediction, you need to give it to people who think in systems. Data scientists. Strategists. People whose job is to find patterns in noise, not to prove that last month’s campaign hit its targets.

I watched a consumer goods company learn this the hard way. Their social team flagged a spike in negative conversation about a product in Southeast Asia. The social team wrote a report. The report went to marketing. Marketing said “we’ll address it in the next campaign cycle.” Three months later, the product was pulled from shelves in two countries because of a quality issue that social data had caught in week one.

The signal was there. The organisation couldn’t hear it because the signal was sitting in a department that only thinks in campaign cycles.

I’m not arguing that social data replaces surveys or focus groups or sales data. Those have structure, depth, and precision that social data lacks. What social data has is scale, speed, and spontaneity. People don’t curate their social posts the way they curate survey responses. They say what they actually think, in the moment, before the marketing department has a chance to tell them what to say.

That rawness is hard to work with. You have to clean it, contextualise it, separate the signal from the noise. But once you do, you have a real-time map of what millions of people care about, updated every second, at zero marginal cost.

The brands that wire this into their decision-making will know what’s coming before their competitors do. Everyone else will keep building dashboards that show what happened last month and calling it insight.