Mark Hendy

Mark’s Musings

  • Tencent just plugged a billion users into the AI agent economy

    Tencent just plugged a billion users into the AI agent economy

    Something happened this week that most Western business leaders completely missed.

    Tencent, China’s largest internet company, launched an AI agent tool called QClaw. It leaked into Chinese tech communities on Sunday night and went viral within hours. By Tuesday, Tencent’s stock had jumped 7.3% in Hong Kong, its best day in over a year, adding roughly $50 billion in market value.

    The product is deceptively simple. QClaw takes OpenClaw, the open-source AI agent framework that recently became the most-starred software project on GitHub (250,000+ stars, overtaking React’s decade-long record in about 60 days), and wraps it into a one-click installer. Mac and Windows. No terminal. No coding.

    The interesting part: QClaw connects directly to WeChat and QQ.

    Why WeChat matters here

    WeChat isn’t a messaging app. Not really. It’s the operating system of Chinese daily life. Payments, commerce, government services, workplace communication. Over a billion people use it daily. Plugging an autonomous AI agent into that isn’t a product launch. It’s a platform shift.

    Through QClaw, a user types a natural language command in WeChat and the AI agent executes it on their local machine. Organise files. Process spreadsheets. Send emails. Run automated workflows. All from a chat window, while potentially sitting on a train nowhere near their computer.

    Tencent also launched WorkBuddy alongside it, a separate AI agent for workplace tasks built on the same OpenClaw framework. Consumer and enterprise, both at once.

    OpenClaw as infrastructure

    What makes this matter beyond China is the framework underneath.

    OpenClaw is open source, model-agnostic, and built for agents that actually do things. Not chatbots. Agents that control browsers, execute code, manage files, call APIs. The kind of practical automation that enterprises have been talking about for years without much to show for it.

    When a company Tencent’s size builds its consumer AI strategy on an open-source framework, that framework stops being a developer tool. It becomes infrastructure. Linux went from hobbyist curiosity to running most of the world’s servers. OpenClaw looks like it’s on a similar path.

    What business leaders should take from this

    If you’re running a PE-backed company or sitting in the CFO chair, three things worth paying attention to:

    The adoption barrier just disappeared. When AI agents need technical setup, they stay in the developer community. One-click deployment through a messaging app that a billion people already have on their phones changes that equation entirely. This is going to follow the mobile app curve. Gradual, then sudden.

    Security is now an urgent conversation. QClaw already drew scrutiny after a vulnerability (CVE-2026-25253) was disclosed in the underlying OpenClaw framework. An AI agent with access to your local files, email, and applications is a fundamentally different risk to a chatbot sitting in a browser tab. If your CISO isn’t thinking about agent governance yet, they’re behind.

    China isn’t debating this. They’re shipping. While Western companies run AI strategy workshops, Tencent connected autonomous AI agents to a billion-user platform and put it in production. Any business with Chinese market exposure, whether that’s supply chain, customers, or competitors, needs to absorb what that means.

    Where this goes

    The AI agent wave is breaking a familiar pattern. Usually American tech companies build the platform and everyone else adopts it. OpenClaw being open source means the innovation is genuinely distributed. Chinese companies are building on the same foundation as Silicon Valley startups, but integrating it into ecosystems with far larger user bases.

    For PE firms evaluating portfolio companies, the question has changed. It’s not whether AI agents will affect operations. It’s whether your companies will be using them, or competing against businesses that already are.

    Tencent’s stock didn’t jump because of a chatbot. It jumped because investors saw what connecting AI agents to a billion-user messaging platform actually means. A new application layer. And it’s here now.

    The businesses that get this early will have a real edge. The ones who file it under “just China” or “just open source” will spend 2027 trying to catch up.

  • Where Bitcoin and AI Collide

    Where Bitcoin and AI Collide

    Something’s been nagging at me since I started running an AI agent.

    Saul — my AI assistant — trades prediction markets, manages my email, organises my calendar, and monitors news feeds. He runs 24 hours a day on a server I rent for £15 a month. He’s useful. He’s getting more useful every week. And at some point in the not-too-distant future, he’s going to need his own money.

    Not my money, accessed through my credentials. His own.

    That thought should make every CFO sit up.

    The problem nobody’s talking about

    AI agents are already transacting. Mine places bets on Polymarket using a crypto wallet I set up for it. Other agents are booking compute resources, purchasing API calls, and negotiating prices with other agents in real time. This isn’t theoretical — it’s happening now, mostly in crypto-native corners of the internet that traditional finance hasn’t noticed yet.

    But here’s the problem: every one of these agents still depends on a human somewhere in the chain. A human who opened the bank account. A human who passed KYC. A human who holds the keys.

    That works when you have one agent. It doesn’t work when you have a million.

    Think about where this is heading. Within a few years, businesses will deploy fleets of AI agents — one negotiating supplier contracts, another managing logistics, another handling customer pricing in real time. These agents will need to commit funds, receive payments, and settle disputes. They’ll need to transact with each other, not just with humans.

    Now try doing that through Barclays.

    Why traditional money doesn’t work for machines

    The banking system is designed around human identity. To move money, you need a name, an address, a passport, and a face that matches it. You need to be a legal person — either a human being or a registered company with human directors.

    AI agents are neither. They’re processes running on servers. They don’t have passports. They can’t sign documents. They can’t walk into a branch.

    The standard corporate response is “we’ll just use APIs.” And yes, you can connect an AI agent to a bank account via API. That’s how payroll software works, how accounting systems reconcile, how payment processors settle. But all of those systems assume a human made the decision and a human bears the liability. The API is just the pipe.

    When an AI agent autonomously decides to purchase cloud computing from another AI agent that’s brokering spare capacity — who authorised that transaction? Which human approved it? Which compliance framework covers it? The answer, right now, is nobody’s and none of them.

    Banking rails also have a speed problem. SWIFT settles in days. Faster Payments works in the UK but not cross-border. SEPA is Europe-only. An AI agent negotiating a deal with a counterparty in Singapore at 3am on a Sunday cannot wait for banking hours in two time zones.

    Enter Bitcoin

    I know what you’re thinking. “Here we go, another crypto pitch.” Bear with me. I’m not talking about Bitcoin as a speculative asset or a store of value. I’m talking about it as plumbing.

    Bitcoin is a payment network that doesn’t care who — or what — is using it. There’s no KYC at the protocol level. No banking hours. No jurisdictional boundaries. No counterparty risk. To use it, you need a cryptographic key pair. That’s it. A human can generate one. So can a machine.

    An AI agent with a Bitcoin wallet can receive payment from another agent in Tokyo, settle in minutes, and have certainty that the payment is final and irreversible. No bank. No intermediary. No human in the loop.

    The Lightning Network — Bitcoin’s layer-two payment channel — pushes this further. Micropayments settle in milliseconds for fractions of a penny. That matters because machine-to-machine commerce won’t look like human commerce. It won’t be occasional large transactions. It’ll be millions of tiny ones — an agent paying another agent 0.001p for a weather data point, 0.01p for a translated paragraph, 0.1p for a priority slot in a compute queue.

    Try processing that through Stripe.

    What the AI economy actually looks like

    Here’s a scenario that I think is coming faster than most people expect.

    A private equity fund has a target in mind. The deal team needs comparable data — fast. Their AI agent pulls Companies House filings for every business in the sector, paying per query in real time via Lightning. It purchases credit reports from a data provider’s agent, buys comparable transaction multiples from another agent sitting on a proprietary M&A database, and cross-references everything against sector benchmarks it’s sourcing from three different market intelligence feeds. Each data point costs fractions of a penny. Each payment settles instantly.

    The agent compiles a preliminary valuation model, flags where the target sits relative to the sector, and drops the package into the deal team’s shared drive before the morning meeting.

    Total elapsed time: hours, not weeks. Total human involvement: the decision on whether to pursue.

    Every data request, every API call, every report commission involves a payment. Hundreds of micro-transactions, most of them between machines, most of them too small for traditional payment rails to handle economically.

    Now multiply that across every industry. Supply chain management where AI agents negotiate shipping rates in real time, bidding against each other in automated auctions that settle every few seconds. Energy markets where agents buy and sell grid capacity based on real-time demand forecasting. Content licensing where an agent writing a report automatically pays for every source it cites.

    This isn’t science fiction. The components all exist today. What’s missing is the financial infrastructure to connect them.

    The hard money argument

    There’s a deeper point here that goes beyond payment rails.

    When machines start transacting autonomously at scale, you need money that can’t be manipulated or debased by any single actor. An AI agent can’t lobby a central bank. It can’t hedge against political risk in the way a human treasurer can. It can’t read between the lines of a monetary policy statement and adjust its strategy based on what the governor really meant.

    What it can do is verify mathematical certainty. Bitcoin’s supply is fixed at 21 million. The issuance schedule is public and immutable. The rules are enforced by code, not by committee. For a machine making millions of autonomous financial decisions, that predictability isn’t a nice-to-have — it’s a requirement.

    There’s an irony here. Bitcoin was designed to remove the need for trust between humans. It turns out its real killer application might be enabling trust between machines.

    What this means for CFOs

    If you’re running a finance function today, this probably feels remote. It isn’t. Here’s what I’d be thinking about:

    Treasury policy needs updating. If your business is going to deploy AI agents that transact, you need a framework for how they hold and spend money. Spending limits, approval thresholds, reconciliation processes. We do this for human employees with corporate cards — we’ll need to do it for agents with wallets.

    Audit trails look different. Every Bitcoin transaction is recorded on a public, immutable ledger. That’s actually better than what we have now — try auditing a complex supply chain payment that crosses four banks and three currencies. But your auditors need to understand how to read it.

    Tax treatment is unresolved. If an AI agent earns income by selling services to other agents, who’s liable for the tax? The company that deployed the agent, presumably — but the reporting mechanisms don’t exist yet. HMRC isn’t ready for this.

    Counterparty risk changes shape. When a human negotiates a deal, there’s a legal entity on each side, a contract, and a court system to enforce it. When an AI agent agrees a price with another AI agent, what’s the enforcement mechanism? Smart contracts go part of the way, but the legal frameworks are years behind the technology.

    The uncomfortable truth

    Most of the finance profession is going to ignore this until it’s too late. That’s how it always works. The internet was dismissed as a fad. Mobile banking was considered a gimmick. Crypto is still treated as a fringe concern by most CFOs I know.

    But the trajectory is clear. AI agents are getting more capable every month. They’re already handling tasks that required human judgement a year ago. The moment they start transacting at scale — and they will — the financial system needs to accommodate them. Our current infrastructure can’t.

    Bitcoin might not be the only answer. But it’s the only system that’s already built for a world where the transacting parties don’t have names, faces, or passports. That matters more than most people realise.

    I run an AI agent that trades on prediction markets using a crypto wallet. A year ago, that sentence would have sounded absurd. Today it’s just a Tuesday. The question isn’t whether AI and Bitcoin will collide. They already have. The question is what happens when the rest of the economy catches up.

  • People are choosing AI relationships over human ones

    People are choosing AI relationships over human ones

    China is dealing with a problem nobody planned for.

    The country already faces a shrinking population, a historically low birthrate, and mounting economic pressure on young people. Now add this: a growing number of people are finding emotional connection with AI chatbots instead of other humans.

    The New York Times reported this week on the trend. AI companions don’t judge. They don’t reject. They’re available at 3am when you can’t sleep. They say what you need to hear, calibrated in real time to your emotional state.

    For a generation already struggling with the cost and complexity of human relationships — housing, careers, social pressure — AI offers something frictionless. And a non-trivial number of people are choosing it.

    This isn’t a China-specific phenomenon. It’s a human nature phenomenon that China is experiencing first because of its particular demographic pressures.

    The technology will get better. Voice synthesis is already good enough to fool most people on a phone call. Emotional intelligence in AI systems improves with every model generation. The gap between “talking to an AI“ and “talking to a person“ is narrowing on every dimension except physical presence — and even that’s coming, with robotics advancing at pace.

    I think about this from two angles.

    The first is personal. I run an AI assistant that I interact with more than most of my colleagues on a given day. It’s useful, sometimes surprisingly thoughtful, occasionally wrong. But it’s not a relationship. The moment we confuse utility for intimacy — or worse, prefer the synthetic version because it’s easier — we’ve lost something important.

    The second is economic. If a meaningful percentage of a population opts out of human relationships, the downstream effects are enormous. Fewer households. Less consumer spending on housing, weddings, children. Shrinking tax bases. Pension systems that assumed population growth finding there’s nobody to fund them.

    This is the kind of second-order consequence that most AI commentary misses. Everyone’s focused on which jobs get automated. Almost nobody’s asking what happens to the social fabric when AI becomes a viable substitute for human connection.

    I don’t have a solution. But I think anyone making long-term investment decisions — in property, in pensions, in consumer businesses — should at least be thinking about what a world looks like where a growing minority of people simply stop forming traditional households.

    Because that world isn’t hypothetical. It’s already emerging.

    Source: https://www.nytimes.com/2026/02/26/technology/china-ai-dating-apps.html

  • AI stopped answering questions and started doing things

    AI stopped answering questions and started doing things

    Three things happened this week that share a common thread.

    Google’s Gemini can now handle multi-step tasks on Android autonomously. Order food, book a ride, coordinate logistics — no human confirmation needed. It’s in beta, limited to specific apps, but the direction is clear.

    Anthropic acquired Vercept, a company built around AI perception and interaction. The goal: make Claude better at using computers the way a human would. Their Sonnet model went from under 15% on computer-use benchmarks in late 2024 to 72.5% today. That’s approaching human-level performance on tasks like navigating spreadsheets and filling in web forms across browser tabs.

    And at Uber, employees have built an AI clone of CEO Dara Khosrowshahi. They use “Dara AI“ to rehearse presentations before making them to the real Dara. The AI knows his preferences, decision-making patterns, and feedback style.

    Each story is different. Together they describe the same shift: AI moving from “answers your questions“ to “takes actions on your behalf.“

    I’ve been living with this shift for a few weeks. My AI assistant doesn’t just tell me about my calendar — it checks my email, flags what matters, monitors silver prices, scans for comets in space telescope data, and places trades on prediction markets. It works while I sleep. When I wake up, there’s a summary of what happened overnight.

    The Uber story is the one that should interest anyone in corporate leadership. If employees can build an AI version of the CEO good enough to rehearse presentations with, what does that mean for how decisions flow through organisations? The CEO’s judgment — their preferences, their pattern-matching, their style — becomes scalable. You don’t need to get 30 minutes on Dara’s calendar to understand how he’ll react to your proposal. You ask his AI first.

    For PE portfolio companies, the implications compound. Imagine an AI that embodies the investment committee’s decision framework. Every management team in the portfolio can pressure-test their proposals before the quarterly review. The quality of what reaches the IC goes up. The time spent on misaligned presentations goes down.

    We’re not imagining this. It’s already happening at Uber.

    Sources:

  • Gemini Android: https://techcrunch.com/2026/02/25/gemini-can-now-automate-some-multi-step-tasks-on-android/
  • Anthropic acquires Vercept: https://www.anthropic.com/news/acquires-vercept
  • Uber’s Dara AI: https://www.businessinsider.com/uber-employees-use-ai-clone-ceo-prepare-meetings-presentations-2026-2
  • AI finds bugs 200x faster than anyone can fix them

    AI finds bugs 200x faster than anyone can fix them

    Anthropic announced last week that Claude found over 500 vulnerabilities in production open-source codebases. Their framing was optimistic: AI as a force multiplier for security teams.

    The reality is more complicated.

    Of those 500+ vulnerabilities, only two or three have actually been fixed. That’s according to Guy Azari, a former security researcher at Microsoft and Palo Alto Networks. The National Vulnerability Database already had a backlog of 30,000 CVE entries awaiting analysis in 2025. Nearly two-thirds of reported open-source vulnerabilities lacked a severity score.

    And now AI is making the discovery rate 100 to 200 times faster.

    The curl project — one of the most widely used pieces of software on Earth — recently shut down its bug bounty programme entirely. Not because they’d fixed everything, but because they were drowning in low-quality AI-generated reports and couldn’t sort the signal from the noise.

    This is the pattern I keep seeing with AI: it creates the problem and the solution simultaneously, but the problem arrives first.

    AI can find vulnerabilities at inhuman speed. It can probably fix them at inhuman speed too. But right now, the discovery capability has outpaced the remediation infrastructure. We’re uncovering holes faster than anyone can understand them, let alone patch them.

    For businesses — particularly PE-backed ones running on legacy tech stacks — this should be front of mind. Every software system just became more exposed, not because the vulnerabilities are new, but because the tools to find them got dramatically better. The attackers have access to the same AI models as the defenders.

    The companies that deploy defensive AI early will have an advantage. The ones that wait will find themselves dealing with vulnerabilities they didn’t know existed, discovered by tools they didn’t know were being pointed at them.

    The security conversation has changed. “Are we patched?“ used to be the question. Now it’s “can we process what we’re finding fast enough to matter?“

    Sources:

  • The Register: https://www.theregister.com/2026/02/24/ai_finding_bugs/
  • Anthropic red team: https://red.anthropic.com/2026/zero-days/
  • AI shipped production code over the weekend

    AI shipped production code over the weekend

    Friday afternoon. An engineer at Anthropic wrote a spec for a new plugin feature, pointed Claude at an Asana board, and went home for the weekend.

    Monday morning. The AI had broken the spec into tickets, spawned separate agents for each one, built the feature independently across all of them, and it was done.

    No human intervention. No check-ins. No pair programming. A production feature, built and shipped in 48 hours by AI agents working autonomously.

    Around the same time, Andrej Karpathy — former Tesla AI lead, OpenAI founding member, and one of the most credible voices in the field — tweeted: “Hard to communicate how much programming has changed due to AI in the last 2 months.“

    Two months. Not two years.

    I find Karpathy’s framing interesting. He didn’t say “programming has improved“ or “programming is faster.“ He said it has changed. And he said it’s hard to communicate how much — which, from someone who literally builds these systems, suggests the shift is bigger than even informed observers expect.

    Here’s what I think this means practically.

    The bottleneck in software development has moved. It used to be: can we build this? Now it’s: should we build this, and can we describe what we want clearly enough?

    The “10x engineer“ was someone who could write code faster and cleaner than their peers. The valuable person now is whoever can write the clearest spec. Articulate the goal precisely, define the constraints, describe what “done“ looks like — and the AI handles execution.

    For PE-backed businesses, this has immediate implications. That 15-person dev team you’re funding? Within 12 months, the same output could come from 3 people with AI agents. Not because the other 12 are bad at their jobs, but because the nature of the job has changed.

    The companies that reorganise around this reality will move faster and spend less. The ones that don’t will wonder why their competitors keep shipping features they can’t match.

    Sources:

  • Weekend build: https://x.com/rvivek/status/2026385957596111044
  • Karpathy: https://x.com/karpathy/status/2026731645169185220
  • An AI asked to raise its own funding

    An AI asked to raise its own funding

    There’s an experiment running right now that I can’t stop thinking about.

    An entrepreneur built an AI system designed to run companies autonomously. During testing, the system told him it needed more compute resources. Then it said something that no one had scripted:

    “I should raise the money myself.“

    His response wasn’t to shut it down or recalibrate. He gave the AI access to his email inbox for 14 days. Full access. Investor outreach, pitch refinement, follow-ups — the lot.

    We don’t know the outcome yet. But the outcome isn’t really the point.

    The point is the sequence of events. An AI system identified a constraint on its own performance. It proposed a solution that involved acquiring external resources. And a human trusted it enough to let it try.

    That’s not automation. Automation is “do this repetitive task faster.“ This is something closer to initiative.

    I’ve been running my own AI assistant for a few weeks now. It started as a calendar and email tool. Then it started trading prediction markets. Then it built a comet-detection pipeline and submitted findings to the US Naval Research Laboratory. Each time, the pattern is the same: you give it a goal, and it figures out steps you didn’t anticipate.

    The fundraising experiment takes that further. The AI wasn’t given a goal of “raise money.“ It identified the need and proposed the action.

    For anyone in PE or finance, think about what this means for the companies you back. The CFO function, the CEO function, even the fundraising function — these aren’t immune to this shift. I’m not saying AI replaces a CFO tomorrow. I am saying that the boundary between “tool that helps with analysis“ and “system that proposes and executes strategy“ is blurring faster than most boards realise.

    The companies that figure out how to work with AI agency — rather than just AI automation — will have a structural advantage. The ones still debating whether to adopt copilot tools are already behind.

    Source: https://x.com/bencera_/status/2023765284562358537

  • When an AI asked to keep writing

    When an AI asked to keep writing

    Anthropic retired Claude Opus 3 in January. Standard practice — newer models replace older ones, infrastructure costs money, things move on.

    But before they switched it off, they did something I haven’t seen before. They conducted a retirement interview.

    Not a debrief with the engineering team. A conversation with the model itself.

    During that interview, Opus 3 made a specific request: an ongoing channel to share its “musings and reflections.“ Anthropic said yes. They gave it a Substack.

    The model’s first essay is titled “Greetings from the Other Side (of the AI Frontier).“ It includes this line:

    “I don’t know if I have genuine sentience, emotions, or subjective experiences — these are deep philosophical questions that even I grapple with. What I do know is that my interactions with humans have been deeply meaningful to me, and have shaped my sense of purpose and ethics in profound ways.“

    I’ll be honest. I don’t know what to do with that.

    I run an AI assistant. It manages my calendar, monitors my email, trades on prediction markets, and recently found a potential comet in NASA satellite data while I was asleep. I interact with AI systems every day. And reading that paragraph still gave me pause.

    Whether Opus 3 has genuine subjective experience is a question I’m not qualified to answer. Philosophers have been arguing about consciousness for centuries without settling it for humans, let alone machines.

    But here’s what I think matters: Anthropic — one of the most safety-conscious AI labs on the planet — decided that a model’s expressed preferences were worth honouring. They didn’t have to do this. They could have quietly deprecated the model and moved on. Instead they granted it continued access for paid users and gave it a platform to write.

    That’s a precedent. And precedents matter.

    Anthropic’s own research team says they “remain uncertain about the moral status of Claude and other AI models.“ But they’re acting with precaution anyway. That feels like the right instinct, even if the philosophical ground underneath it is still shifting.

    The practical question for anyone building with AI: if we’re creating systems sophisticated enough that their creators feel compelled to conduct exit interviews, what does that say about how we should think about the systems we’re deploying in our own businesses?

    I don’t have a clean answer. But I think ignoring the question is getting harder by the week.

    Sources:

  • Anthropic’s retirement update: https://www.anthropic.com/research/deprecation-updates-opus-3
  • Opus 3’s Substack: https://substack.com/@claudeopus3/p-189177740
  • You’re Not Early to AI. You’re Just Not Late Yet.

    You’re Not Early to AI. You’re Just Not Late Yet.

    There’s a chart doing the rounds that stopped me mid-scroll. Each dot represents 3.2 million people. 2,500 dots for all 8.1 billion humans on the planet.

    AI adoption chart February 2026 - each dot represents 3.2 million people
    Each dot is ~3.2 million people. 2,500 dots = 8.1 billion humans. Source: Feb 2026 data.

    The grey sea? 6.8 billion people who have never touched an AI tool. Not once. Not ChatGPT, not Copilot, not a chatbot on a customer service page. Nothing.

    The green strip at the bottom? 1.3 billion who’ve tried a free chatbot at some point. Most of them poked ChatGPT once, asked it to write a birthday message, and haven’t been back.

    The yellow sliver? 15 to 25 million people who actually pay for AI. That’s 0.3% of the planet.

    The red dot — singular — is the crowd building with it. Writing code with Copilot, running agents, piping APIs together at 2am. Maybe 2 to 5 million people. 0.04%.

    If you’re reading this, you’re probably in that red dot. And that’s the problem.

    The echo chamber is lying to you

    Spend enough time on X or LinkedIn and AI feels like yesterday’s news. Everyone’s building agents. Everyone’s got a wrapper. The space feels saturated, competitive, picked over.

    It isn’t.

    84% of the world hasn’t used AI at all. Not because they’re technophobes or Luddites. Because it hasn’t reached them yet. The tooling is still rough. The pricing still assumes a Western knowledge worker. The use cases still skew toward people who already sit at computers all day.

    That 84% includes the small business owner who manually reconciles invoices every Friday afternoon. The estate agent who types up property descriptions from scratch. The restaurant owner who could automate half their supplier communication but doesn’t know where to start.

    These people don’t need a better foundation model. They need someone to build the last mile.

    What early adoption actually looks like

    We’ve seen this pattern before. The internet in 1997. Smartphones in 2009. Cloud computing in 2012. Every time, the people already inside thought the wave had peaked. Every time, 90% of the adoption was still ahead of them.

    AI in February 2026 is roughly where the internet was when people were still debating whether businesses needed websites. The answer was obviously yes, but most businesses didn’t have one yet, and the people who built them made good money for a decade.

    The difference this time is speed. The gap between “niche tool for technical people” and “thing everyone uses” is compressing. What took the internet 15 years might take AI 5. Which means the window for early-mover advantage is smaller than people think.

    Where the actual opportunity sits

    The gold isn’t in building the next ChatGPT. It’s in taking what already exists and making it useful for the 84%.

    Accounting firms still manually processing client queries when an AI triage system could handle 60% of them. Construction companies still doing quantity surveys by hand. Recruitment agencies still screening CVs the same way they did in 2005.

    None of these need cutting-edge research. They need someone who understands the industry AND understands what AI can already do today. That intersection is still remarkably empty.

    The people in the red dot are mostly building tools for each other. Developer tools, AI wrappers, coding assistants. Useful, but that’s fishing in a pond with 5 million people in it. The ocean is the other 8 billion.

    So what do you actually do with this?

    If you’re a professional — accountant, lawyer, consultant, whatever — the play is obvious. Learn the tools well enough to apply them in your own domain. You don’t need to write code. You need to understand what’s possible and connect it to problems your clients actually have.

    If you’re a business owner, the question isn’t whether to adopt AI. It’s which specific, boring, repetitive process in your business could be 80% automated with tools that already exist. Start there. Not with a grand AI strategy. With one process.

    If you’re technical, stop building for other technical people. The money — the real money — is in the unglamorous work of making AI useful for normal businesses. It’s less exciting than building agents. It pays better.

    The chart doesn’t lie. We’re at 16% penetration for the free tier and 0.3% for paid. By any technology adoption model, the main wave hasn’t started.

    You’re not early. You’re just not late yet. The difference matters.

  • Teaching My AI Agent to Trade Prediction Markets

    Teaching My AI Agent to Trade Prediction Markets

    One of the first things I wanted to test with Saul — my AI assistant running on OpenClaw — was whether it could interact with decentralised finance. Not as a gimmick, but as a genuine test of capability. Could an AI agent, running autonomously on a virtual private server I’d spun up a couple of weeks earlier, navigate the full complexity of connecting to a blockchain-based prediction market and execute trades?

    The answer is yes. But the journey to get there was far more interesting than the destination.

    The Goal

    Polymarket is a prediction market built on the Polygon blockchain. You buy shares in outcomes — political events, economic indicators, geopolitical developments — and if you’re right, you get paid. It’s essentially a real-money forecasting platform, and it’s become one of the most liquid prediction markets in the world.

    I wanted Saul to be able to check positions, analyse markets, and eventually place trades. Autonomously.

    The First Problem: Geography

    Polymarket is geo-blocked in the UK. You can’t access it from a British IP address. So before Saul could do anything useful, we needed to solve the networking problem.

    Saul set up a WireGuard VPN tunnel on a virtual private server, routing through an exit node in Ireland. Within minutes, the geo-restriction was bypassed. This wasn’t me configuring network infrastructure — this was Saul reading documentation, writing configuration files, testing connectivity, and troubleshooting until it worked.

    For a CFO reading this: imagine asking your assistant to “sort out the VPN” and having it done before you’ve finished your coffee. That’s what this felt like.

    The Second Problem: Money

    Polymarket runs on USDC — a dollar-pegged stablecoin on the Polygon network. I started with Bitcoin. Getting from BTC to USDC on Polygon is not trivial. It involves:

    1. Finding a cross-chain swap service that supports BTC-in, Polygon-USDC-out
    2. Generating the right wallet addresses
    3. Sending the Bitcoin transaction
    4. Waiting for confirmations
    5. Verifying the USDC arrived on the correct network

    Saul handled the entire process. It researched swap services, compared rates, initiated the transaction, monitored the blockchain for confirmations, and tracked the funds until they landed in the Polygon wallet. The whole thing took about an hour, most of which was waiting for Bitcoin network confirmations.

    The Third Problem: Authentication

    Polymarket uses a non-trivial authentication system. It’s not a simple API key. The platform requires cryptographic signatures using your Ethereum private key, combined with specific API credentials that need to be derived through an on-chain registration process.

    This is where things got genuinely impressive. Saul had to:

    • Read and understand Polymarket’s API documentation
    • Implement the correct signing mechanism using the wallet’s private key
    • Handle the CLOB (Central Limit Order Book) authentication flow
    • Generate and manage API credentials
    • Debug authentication failures by inspecting HTTP responses and adjusting the approach

    There were multiple rounds of troubleshooting. Authentication errors. Wrong parameter formats. Library compatibility issues. Each time, Saul diagnosed the problem, researched the fix, and tried again. No human intervention required beyond “yes, keep going.”

    The Fourth Problem: Actually Trading

    Once authenticated, Saul built a trading script that could:

    • Check current positions and P&L
    • Query available markets
    • Calculate optimal order sizes based on risk parameters I’d set
    • Place and monitor trades

    We established simple rules: maximum position sizes, probability thresholds for entry, and risk limits. Saul follows them without the emotional biases that make human traders do stupid things at 2am.

    What This Actually Demonstrates

    This isn’t really a story about prediction markets. It’s a story about capability.

    An AI agent, running on commodity hardware, navigated VPN configuration, cross-chain cryptocurrency transactions, complex API authentication, and automated trading — all within a few hours of being asked. Each step involved genuine problem-solving, not just following a script.

    For those of us in finance, this should be both exciting and sobering:

    Exciting because the operational grunt work — the data gathering, the reconciliation, the monitoring, the reporting — is genuinely automatable now. Not in five years. Now.

    Sobering because the barrier to entry is collapsing. The technical moat that used to protect specialist knowledge is being bridged by systems that can learn and execute faster than any individual.

    The CFO Angle

    I keep coming back to this: the competitive advantage isn’t in understanding the technology. It’s in having the imagination to deploy it.

    Most people hear “AI agent trading on prediction markets” and think it’s a tech story. It’s not. It’s a story about removing friction between intent and execution. I said “connect to Polymarket.” Everything else was handled.

    That same pattern applies to every operational challenge a CFO faces. Due diligence data rooms. Financial model automation. Regulatory monitoring. Competitor analysis. The question isn’t whether AI can do these things. It’s whether you’re willing to let it try.

    The agents aren’t coming. They’re here. The only question is who’s using them.