Key takeaways
- 4 August: Cloudflare gave AI agents an identity and a wallet, with spending caps, approved merchant lists and transaction limits set by the owner.
- 7 August: it launched Kitesurf, a browser built only for agents. No tabs, no themes, no extensions, no pixel-perfect rendering.
- The traffic already flipped. Machine traffic overtook human traffic on Cloudflare's network in May 2026, a year earlier than the company itself expected.
- The 1,000x forecast is a forecast. Treat it with care. The measured crossover is the part that should change your plans.
- Check your website terms today. The standard clause banning automated access may currently prohibit your own customers' assistants from buying.
For thirty years the assumption underneath every website was that a person was looking at it. That assumption paid for your design budget, your product photography and your checkout flow. It is now the thing most likely to be wrong about your online business, and the companies that build the plumbing of the internet have stopped pretending otherwise.
What Cloudflare actually announced
Two things in the same week, and they only make sense together.
- 4 August: identity and a wallet for agents.Cloudflare Wallets and cloudflare.pay give agents a stable ID linked to the human or organisation that owns them, plus the ability to spend within limits their owner sets.
- 7 August: a browser with no human in mind.Kitesurf, cloud-hosted on Cloudflare Workers, built in twelve weeks, already passing more than 215,000 web platform tests. Free in beta.
- Underneath both: the traffic mix changed first.Machine traffic overtook human traffic on Cloudflare's network in May 2026 and reached roughly 57.5% of volume in Q2.
Matthew Prince, Cloudflare's co-founder and CEO, put the reasoning plainly in the announcement.
"When an agent shows up at your door, you need to know who sent it."
Matthew Prince, CEO, Cloudflare — press release, 4 August 2026The company's own framing of the problem is the sharpest description of where most businesses currently sit. When an agent arrives to buy something or start a free trial, the business has no reliable way to tell whether it is a real customer's assistant or somebody gaming the system. The old bot-detection tools were built for search crawlers, not for software that transacts on behalf of real people. So businesses end up choosing between locking everything down or taking their chances, and neither of those works at scale.
The numbers, and how much weight to put on them
May 2026
machine traffic passed human traffic on Cloudflare's network for the first time, a year earlier than the company's own forecast
57.5%
of requests through Cloudflare's infrastructure were bot-generated in Q2 2026
1,000x
non-human versus human traffic within five years, per Cloudflare's CFO on the Q2 earnings call
Why the 1,000x figure deserves a pinch of salt
The quote doing the rounds is Cloudflare CFO Thomas Seifert saying humans will become "a rounding error" on the internet, and that this happens "not because human traffic goes down, but that's just how fast we're seeing non-human traffic grow". It is a striking line and it travelled accordingly.
It is also a five-year forecast made on an earnings call by a company that sells bot management, and, to his credit, Seifert flagged that he has form for getting these predictions wrong. We would not plan a business around 1,000x.
You do not need to. The measured number is the one that matters. The crossover already happened, and it happened a year sooner than the people with the best view of the data expected. That is enough to justify spending an afternoon on the question.
A browser with no eyes
Kitesurf is the more revealing announcement, because of what Cloudflare took out. A browser for agents does not need tabs, themes, extensions, pixel-perfect rendering or smooth scrolling. What it needs instead is to manage context windows, token costs and scale, and to survive a different threat model in which prompt injection is a live risk.
Read that list of deletions again as a description of your website. Tabs, themes, rendering and scrolling are the parts a person experiences. If a meaningful share of your visitors arrive through something that discards all of it, then your carefully art-directed pages are being consumed as a data source. Structure, clarity, accurate prices and machine-readable answers start to matter more than the hero image.
We covered the discovery half of this in our piece on what AI search traffic means for UK businesses. This is the next step: not agents reading about you, but agents doing business with you.
Three decisions this forces
None of these are IT questions, which is exactly why they tend to go unanswered.
| The decision | The lazy answer | What good looks like |
|---|---|---|
| Do you let agents in at all? | Block everything, or ignore it | Welcome research and comparison, gate the actions that cost you money if abused |
| Can an agent actually complete a transaction? | Assume the website flow protects you | Rules enforced in the API, not the interface, and a human checkpoint on irreversible actions |
| Can you tell who sent it? | Treat all automation as a threat | Verify agent identity where it exists, and decide what you do when it does not |
The middle row is not hypothetical. Last week an agent asked to book a gym class found that the gym's booking rules existed only in the website and not in the API behind it, and cancelled a stranger's reservation to free a slot. Any business whose logic lives in the front end is one motivated agent away from the same discovery.
The clause in your terms that now points the wrong way
Here is a concrete job you can do this week, and it is one almost nobody has thought about.
Nearly every UK website's terms and conditions contain boilerplate prohibiting automated access: no bots, no scrapers, no robots, no automated tools. That wording was written years ago to deter people harvesting your content and your prices. Read it again in a world where your customer's assistant does the booking. As drafted, it very likely prohibits your own customers from buying from you.
You may decide that is exactly what you want, and for some businesses it will be. The point is to decide it deliberately rather than inherit it from a template written for a different internet. Permit, restrict or charge, then make the wording say what you actually mean. We are not lawyers, so get the drafting checked properly.
Worth knowing: the wallet design tells you where this is heading. Virtual Wallets carry a spending cap, an approved merchant list and a maximum transaction size the agent cannot exceed on its own. If approved merchant lists become normal, being on the list is the new shelf space, and that is a commercial problem long before it is a technical one.
Who owns this in your business
Ask most leadership teams who owns "agents on our website" and you get a pause, then someone says IT. That is the wrong answer in the same way that "marketing owns the website" was the wrong answer in 2005.
Whether to accept agent customers is a commercial decision about your channel, your margins and your risk appetite. Whether an agent can complete a booking is a product decision. Whether you can verify who sent it is a security decision. Whether your terms permit any of it is a legal one. If nobody senior is holding all four, the default answer gets made for you by whatever your web stack happens to do today.
What to do this quarter
-
Measure your own split before you theorise
Do: pull your analytics and server logs and find out what proportion of your traffic is already automated, and which pages it hits.
Why: Cloudflare's network-wide figure is not your figure. A UK B2B firm and a consumer retailer will be in very different positions, and the argument gets much easier with your own number in the room. -
Decide your policy on agent access, in writing
Do: agree which agent activity is welcome, which needs verification or a human checkpoint, and which you refuse. Then align your website terms to that decision.
Why: "block everything or take your chances" is the position Cloudflare says most businesses are stuck in, and both ends of it cost you money. -
Move your business rules out of the front end
Do: confirm every limit that matters commercially, booking windows, quantities, discounts, cancellations, is enforced in the API and not just the interface.
Why: an agent talks to your API directly. Front-end rules are advisory to software, and this is precisely how the gym incident happened. -
Make your key pages machine-legible
Do: accurate structured data, unambiguous pricing, clear availability, plain answers to the questions a buyer actually asks.
Why: an agent with no eyes cannot infer anything from your design. What it cannot parse, it cannot recommend you for. -
Put one senior name against it
Do: assign a single owner with commercial, product, security and legal input, and give them a date to come back to the board.
Why: this crosses four functions, so with no owner it becomes everybody's second priority and nobody's first.
We will sit down with your leadership team, look at where agents plausibly touch your business in the next 18 months, and tell you which of it is genuinely urgent and which is noise. 25 minutes, no obligation, and we will say so if the honest answer is "not yet for you".
Book a strategy conversation →The part that is a skills problem
Every one of those five steps is a judgment call, and not one of them is solved by buying software. Deciding which agent activity is welcome, what an approved merchant list does to your channel strategy, and what "verified" should mean before you accept an order are commercial questions that require someone who understands both the technology and the P&L.
That combination is exactly what the Level 5 AI Strategy & Opportunity unit is built to develop, and it pairs with AI Adoption & Governance for the boundaries and verification side. A leadership team can take all three Level 5 units as one closed cohort through the AI Leadership Pathway, which works better here because the team argues about its own channel rather than a case study. For the people who then have to implement it, the AI & Automation Practitioner Level 4 covers agent deployment and responsible AI as core content. If you want the plain-English grounding first, start with our employer's guide to AI agents.
All of it is fundable through the Growth and Skills Levy, and none of it requires a coding background.
The honest summary
We are not going to tell you the web will be 99.9% robots by 2031, because nobody knows that and the person who said it does not claim to either. Forecasts made on earnings calls by companies selling the remedy deserve exactly the scepticism you are already applying.
What is not a forecast is this. The crossover has happened. Agents now have a browser designed for them, an identity that can be checked and a wallet with spending limits. The infrastructure for software to buy things from your business is being laid, quickly, by the companies best placed to see the demand.
You do not need to rebuild anything this month. You do need somebody senior to have an opinion, because the alternative is that your website terms, your API and your checkout quietly answer the question on your behalf.
Frequently asked questions.
What is agentic commerce?
Agentic commerce is when an AI agent, acting for a person or a business, browses, compares, books or buys on their behalf rather than the human doing it directly. The agent is the one filling the form, choosing the supplier and completing the transaction. For the business on the other end it means the visitor in the session is software, and the buying decision was shaped before anyone reached your website.
What is Cloudflare Kitesurf?
Kitesurf is a cloud-hosted web browser Cloudflare launched on 7 August 2026, built for AI agents rather than people. It runs on Cloudflare Workers and strips out everything that exists for human eyes, including tabs, themes, extensions and pixel-perfect rendering, in order to cut CPU and memory use against Chromium for agentic tasks. Cloudflare says it was built in twelve weeks and already passes more than 215,000 web platform tests. It is free while in beta through Browser Run.
What are Cloudflare Wallets and cloudflare.pay?
Announced on 4 August 2026, they give AI agents a stable identity and the ability to buy things within limits set by their owner. A Cloudflare account gets a unique address that works as an ID, which can be extended to specific agents so a business receiving a request can see who authorised it. An Account Wallet holds stablecoins, and Virtual Wallets can be issued to individual agents with a spending cap, an approved merchant list and a maximum transaction size. Handle reservation opened in August, with full access due in the following months.
Has bot traffic really overtaken human traffic?
On Cloudflare's network, yes. Machine-generated traffic passed human traffic for the first time in May 2026, a crossover the company had previously expected in 2027, and bot requests reached about 57.5% of volume in Q2 2026. Cloudflare's CFO Thomas Seifert went further on the Q2 earnings call, suggesting non-human traffic could reach 1,000 times human traffic within five years. That last figure is a forecast rather than a measurement, and Seifert himself noted he has been wrong before.
Do my website terms and conditions stop AI agents buying from me?
Very possibly, and this is worth checking today. Most UK website terms carry a boilerplate clause prohibiting automated access, bots, scrapers or robotic tools. That clause was written to deter scrapers, but as written it can also prohibit a genuine customer's assistant from booking or buying. Decide deliberately whether you want to permit, restrict or charge for agent access, then make the wording say what you actually mean. We are not lawyers, so take proper advice on the drafting.
Should we block AI agents from our website?
Blocking everything and allowing everything are both poor answers at scale. Cloudflare's own framing is that businesses are currently stuck choosing between locking everything down or taking their chances, and neither works. The more useful position is a deliberate one: decide which agent activity is welcome, such as research and booking, which needs verification or a human checkpoint, such as payment and account changes, and which you refuse. That is a commercial decision, not purely an IT one.
How do we build the skills to handle agentic commerce?
Treat it as strategy plus governance rather than a website task. In England both are fundable through the Growth and Skills Levy. The Level 5 AI Strategy & Opportunity unit builds the commercial judgment about where agents create or destroy value in your model, and AI Adoption & Governance covers the boundaries, verification and oversight. The AI & Automation Practitioner Level 4 apprenticeship builds the people who implement it. No coding background is required to start.
Sources: Cloudflare press release on Wallets and cloudflare.pay, 4 August 2026; TechCrunch on the Kitesurf launch, 7 August 2026; The Register on Cloudflare's Q2 2026 earnings call and the traffic forecast. Traffic percentages are for Cloudflare's own network and are not a measure of the internet as a whole. The five-year projection is a forward-looking statement by a company that sells bot management, and should be treated as such.