When the next six billion internet users are agents

By Michael Stothard, Lorcan Delaney and Adriana Vitagliano

The first 6 billion internet users were humans. The next 6 billion are going to be AI agents. Most people will have a general-purpose personal agent like a Muse or an Instinct doing their bidding, or maybe a fleet of specialist ones, and that will reshape sectors from ecommerce and travel to social media and online dating.

At the same time, the B2B versions of these agents are already changing how pretty much everyone works. Most of the hyperscalers and thousands of startups are trying to own the context layer of the business, connecting to its tools and systems of record and building "AI workers" on top to do the jobs.

The big question, for us at firstminute, is not whether the above predictions will come to pass. We think they will. The more far-sighted question is what happens next, once machines are doing the clicking. What follows is our attempt to map the second-order consequences of six billion agents.

Online commerce

For twenty-five years online shopping has been built for a human eye and a human thumb. Product pages, search rankings, review stars and the small dopamine hit of the "Buy now" button all exist to persuade a person to click. That person is about to be clicking a lot less.

Much of the research for shopping already happens inside an AI assistant, a shift that has happened frighteningly quickly since ChatGPT launched in late 2022. By one estimate, ChatGPT handles around 50 million shopping queries a day. Agents are now beginning to be allowed to complete the purchase as well, and McKinsey reckons agentic commerce could be worth $3 to 5 trillion globally by 2030.

Generative engine optimisation, the business of getting your products recommended by AI, is the first layer to change. Companies like Peec AI and Profound, and many others, are offering to help brands appear in ChatGPT when people type queries like "what cool shoes can I buy?"

Beneath it sit startups providing the underlying infrastructure for the ecommerce sites themselves, for example making product catalogues machine-readable so an agent will consider them more favourably (or at all). FERMÀT and Firmly are two examples.

Others on the infrastructure side give brands their own sales agents so they keep the customer relationship instead of handing it to the big model companies (Swap, Daydream). A third group proves who an agent is when it turns up to spend (Skyfire, Nekuda).

The CTO of a $40bn+ listed US ecommerce retailer told firstminute it was optimising its catalogue for AI agents, which it sees as a potential "new channel", and improving its own "agent concierge" on its site to keep hold of the customer.

Another interesting sub-question is what happens to advertising. Search ads, display, sponsored listings, influencer deals and affiliate links all assume a human is looking. What happens to the more than half a trillion dollars of annual digital ad spend? Does it die or go somewhere else?

The end of the apathy economy

A surprising number of businesses make money from people not paying attention. There's the gym membership nobody uses, and the subscription that renews because cancelling means a phone call. There's the savings account paying a fraction of the market rate because moving is a hassle. There's the insurance renewal priced higher for loyal customers, a practice so common in UK home and motor insurance that the FCA banned it in 2022.

Agents don't get bored or forget, and they don't mind waiting on hold. We have entered a world where you can tell a personal agent to keep your bills as low as possible and it will cancel, switch, claim refunds and contest fees across every account you have, every month.

For consumers that means money back. For companies whose unit economics depend on inertia, it's a problem. Some incumbents will fight back with more friction: blocking agents, designing cancellation flows to defeat automation, and writing terms that forbid it.

The startup opportunity here is less obvious. One route is treating apathy-based businesses as targets for disruption. Another is becoming the company that does the switching, like Nous, the agentic company for managing household bills being built by Greg Marsh, who previously founded onefinestay.

Marketplaces

The marketplace may be the most defensible business model technology has produced. eBay, Amazon, Airbnb, Uber, DoorDash and the Alibaba complex all run the same machine. They gather enough buyers and sellers in one place, match supply with demand, take a cut, and let network effects lock everyone in. NFX, a fund built around network effects, reckons about 70% of the value created in technology since 1994 came from businesses with network effects at their core.

Agents put pressure on that logic. If an agent can assemble the market itself, checking every hotel, freelancer or expert, what is the marketplace for? Since Muse launched, the consumer internet has had its own SaaSpocalypse moment. In September, Booking's shares were down 18%, Expedia's 13% and Airbnb's 15%.

Are marketplaces doomed? It's too early to tell, but we think the pure aggregators, whose value is a searchable list, are exposed, because the agent now does the searching. Price comparison sites are the obvious example.

The ones built on a live network (Uber has to find a driver near you within four minutes), on trust (Airbnb's reviews and guarantees) or on proprietary supply may well be fine. So may those with taste, which stand for something in how they serve customers. All this is probably good for Trustpilot and similar companies, because how else do agents find out what to trust? Maybe an agent-first Trustpilot is a good idea, though.

Consumer: who owns the customer?

The logic of the personal agent will likely reshape the consumer internet one vertical at a time. One big question, though: will the universal agents absorb everything, or is there room for specialists?

Our view is that the universal agents will own the front door and most of the long tail. Reordering groceries, booking a restaurant and paying a parking fine don't need a dedicated company. Specialists win where the job needs something the general agent can't get alone. That could be a network of other people, proprietary supply, regulatory permission, or someone willing to carry the liability when things go wrong.

Dating is an interesting example. Hinge, Tinder and Bumble built big businesses by gathering supply and demand and then monetising the swipe. A new wave of startups, including Fate, Iris and The Relationship Company (backed by Elad Gil), lets you talk to an agent that finds you a date. So no swiping, just good intros to potential partners. But the fundamental need to have lots of people on the platform does not change.

Travel is another example. An agent can already book a weekend in Lisbon. Planning a multi-stop safari through Namibia is harder, because it involves small operators with hard-to-access websites and plenty can go wrong. Expedia, which bought firstminute portfolio AI travel company Layla, could be killed by universal agents, or it could be made stronger by building its own agents on top of the verified supply it has spent years building up.

Money, health and education also have huge potential as categories. An agent that refinances your mortgage when rates fall, rebalances your portfolio, reads your wearable data or tutors your child at exactly the right pace would be enormously valuable. These areas are also either regulated or depend on a kind of trust that takes years to earn. We expect specialists with the right licences and data to lead in some categories, with the universal agents perhaps plugging into them.

Agents talking to agents

Most of this piece assumes one agent dealing with businesses. The stranger consequences come when personal agents deal with each other. If your Instinct and mine can talk, they can find a time to meet without either of us opening a calendar, make an introduction when they spot a shared interest, or settle who owes what after a holiday.

They'll also need rules about what to share. My agent might tell yours I'm free on Thursday without saying why, and tell my family's agents more than a stranger's.

That points to a social graph between agents, with trust levels and permissions attached. Whoever builds that layer, deciding which agents can vouch for each other and what they're allowed to reveal, could end up holding a version of the network effect that agents threaten elsewhere. It's also where we'd expect the most surprising behaviour, because nobody has designed social norms for machines before.

See below Michael's semi-viral LinkedIn post about his Instinct talking to his wife, which was also picked up by Business Insider.

B2B: AI workers

The business counterpart of the personal agent is the AI worker, software sold as a member of staff. Almost every B2B startup we meet now pitches some version of a unified context layer for a company's operations, with agents on top that do the work.

These companies come in three broad shapes.

General coworkers for small and medium-sized businesses, like the firstminute-backed Fieldwork or the Accel-backed Viktor, handle tasks from media buying and cash-flow forecasting to triaging email.

Function-specific agents target a single department. Examples include HR (CellCog), IT service management (Serval, Console), marketing and sales (Rose, Artisan), recruiting (Juicebox) and customer support (Sierra, Decagon).

Vertical agents go deep into one industry's workflows. Examples include ecommerce (FERMÀT), healthcare and pharma (Vocca, ClarityCare), insurance (Sprout AI), field services and manufacturing (Movable Voice, Avoca) and travel (Otto).

One interesting second-order consequence of all this is what work becomes. Our favourite take is the slightly, but only slightly, tongue-in-cheek one from Yoni Rechtman, Partner at Slow Ventures, who said the only jobs left will be:

• Slop cannons: people using AI to produce content like code, Excel models, slides, social media posts and emails.

• Infra and security people: people who can orchestrate all the AI agents and keep them secure.

• Hot people: by which Yoni means people who can talk to other people.

• Adults: people who can manage the other three.

Infrastructure: identity and security

If agents are going to buy, hire and post on our behalf, someone has to build the infrastructure to identify, secure and contain them (one of the four jobs left!).

Start with identity. Anyone an agent deals with needs to know who it is and whom it acts for. "Know your agent" has gone from an academic phrase to a funded category in about a year. Skyfire, Basis Theory, Nekuda and Vouched are building the identity layer, while Visa and Mastercard race to build it into the payment rails.

Bot detection has become harder and more valuable at the same time. Once anyone can send an agent to buy concert tickets or book restaurant tables, telling good bots from bad ones becomes an everyday problem for every business with a website.

Then there's security. In late 2025 Anthropic disclosed what it described as the first largely autonomous cyber-espionage campaign, with AI carrying out an estimated 80 to 90% of an intrusion effort against around thirty targets.

Attackers that never sleep and move at machine speed need defenders that do the same, and the money is following in a big way. Cisco bought Astrix, which secures AI agents and machine identities, for a reported $400m in May. In July, Cyera agreed to acquire Oasis Security for about $1bn.

Money: paying and insuring the agents

Agents will need to pay for things, and there's already a fight over how.

Visa (Intelligent Commerce) and Mastercard (Agent Pay) want agent transactions on the card networks. Banks and open-banking players want account-to-account payments. Stripe is trying to sit across all of it, building agentic checkout with OpenAI and buying into stablecoins and wallets through Bridge and Privy.

The x402 protocol and the crypto crowd want something native to the internet, and a cluster of new companies is betting that crypto rails are the only practical way to handle the tiny, constant payments agents will make to each other for data, compute and API calls.

Large companies will mostly decide this contest, but every rail needs tooling, and some of that will come from startups.

A related problem is that agents stall when they hit a paywall or need credentials. Solid is one of the companies giving agents their own isolated environments, bank accounts and programmable spending limits, so they can keep working without a human approving every transaction.

Then there's insurance. When an agent gets something wrong, someone has to pay. Since January 2026, insurers have been able to attach new standard generative AI exclusions to commercial liability policies, and a growing number do.

A wave of specialist insurers has moved into that gap, including the startups Mount, Klaimee, Corgi and Armilla, with Munich Re writing cover for smaller businesses. To price the risk, they have to measure how reliable an agent actually is. That means their underwriting is becoming the first real safety standard for agents, well before any regulator writes one.

Agentic internet basics: browser and email

Chrome was built for people to look at. Agents don't need the tabs, the pixels or the extensions, and running thousands of copies of a human browser at once is slow and expensive. Two kinds of company have spotted this. Browserbase, Steel and Browserless run Chrome in the cloud at scale so agents can drive it remotely, mostly through the Chrome DevTools Protocol that developers already used to automate it. Lightpanda goes further and is building a headless browser from scratch for machines.

Email is the other piece of plumbing agents borrow from us. They can work through our inboxes, but AgentMail, backed by General Catalyst, is building an email provider whose users are AI agents.

Why can't we just give agents Gmail? According to founder Haakam Aujla: "Gmail was never designed for developers or automation. There are hard limits in the API around creating inboxes, and strict rate limits on sending and even reading."

He adds: "OAuth, the standard for authorizing app access, requires a browser to refresh tokens, when ideally everything is an extension of the agent through an API. And if you create a new Gmail account for your agent, Google bans you."

What other basic internet tools need to be built fresh for the agentic era? There are probably dozens.

At firstminute we back European founders at inception building global tech businesses. If you're building in any of the categories above, or can think of other tools the agentic internet needs, please email us at michael@firstminute.capital or lorcan@firstminute.capital. We would love to hear from you.