Software now outnumbers people online

On 3 June 2026, Cloudflare's chief executive Matthew Prince posted that automated traffic had passed human traffic online for the first time. Cloudflare Radar showed about 57.5% of requests for HTML pages coming from software and about 42.5% from people; some reports of the same seven-day average give 57.4%. Cloudflare had expected the crossover in the second half of 2027. In its founders' letter it says the rise of agents and AI crawlers pulled that date forward to May 2026.

It is worth being precise about what that number counts. It covers every kind of automation together: search engine crawlers, archive bots, AI training crawlers, monitoring tools and agents acting for a person. It measures requests, not attention; a person still reads the article a crawler fetched. And it is one network's view, measured one way. The direction is not in doubt, though, and the part growing fastest is the one businesses will deal with directly.

The fastest-growing part is agents

Crawlers read the web. Agents use it: they search, compare, fill in forms and, increasingly, buy. In its September review Cloudflare says the fastest-growing part of automated traffic is no longer crawlers, and that daily requests from AI agents on its network grew by more than 1,700% in a year.

HUMAN Security, which measures traffic across its own customer base, reaches a similar conclusion from different data. Its 2026 benchmark report found that traffic from AI agents and agentic browsers grew 7,851% year over year during 2025. The report itself notes that this started from a low base, so the multiplier describes early adoption rather than a mature channel. The broader figures are steadier and arguably more telling: AI-driven traffic grew 187% across 2025, and automated traffic overall grew about eight times faster than human traffic.

The pattern is not confined to one market. Australia's Information Age carried the HUMAN findings as agents swarming the web. Cloudflare's country breakdown in the week of the crossover showed automated shares well above the global figure in several markets that host large data-centre estates, Singapore and the Netherlands among them.

Human traffic is falling in some sectors

The more uncomfortable finding for operators is what happens to people. Cloudflare reports that in some heavily crawled categories, including retail, computer software, IT services and financial services, human traffic fell by as much as 40% in under a year. If the customer increasingly arrives as software, a website designed only for a human reader is designed for a shrinking share of visits.

Cloudflare's own response is telling: it describes agents as customers in their own right, and has built ways for sites to recognise signed agents and to charge them per request. The infrastructure of the web is adapting to a second kind of visitor.

Meanwhile, many agent projects will fail

Against this growth sits a sober forecast. Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls. It also warns of agent washing, the relabelling of chatbots and older automation tools as agents, and estimates that only about 130 of the thousands of vendors claiming agentic products are real.

These two trends are not in contradiction. Agent traffic is growing because agents are cheap to start. Agent projects fail because they are expensive to run properly: they need somewhere to run that stays up, a way to be reached and audited, a way to pay, and limits that someone can enforce. The projects Gartner expects to be cancelled are, largely, the ones where nobody planned for those.

Regulators are asking for limits and accountability

Supervisors are already describing what running agents responsibly looks like. Singapore's IMDA asks organisations to bound the risks upfront, by setting limits on what agents may do and which tools and data they may reach, and to make people meaningfully accountable, with human approval at significant checkpoints. In the UK, the FCA's review of AI in retail finance found that a fifth of consumers, about 11 million adults, would be likely to use AI that acts on their behalf within pre-set goals, while worrying about trust and control.

The message from both is the same as the message in Gartner's numbers. Agents are acceptable when someone has set the limits and can show what happened.

What this means for executives

Your customers, suppliers and competitors are already sending software to your front door. Some of it is welcome, some of it is not, and you need to know which. Count it, separate agents acting for real customers from crawlers, and decide what each is allowed to do.

Start agent projects with the operating model, not the demo. The questions that kill projects in year two are predictable: what does it cost to run, who is accountable, what can it spend, and where is the record. Answer them before the pilot.

Give agents the same controls you give staff. An agent that can send mail or move money should have its own address, its own spending limit and its own log, set by a named owner. That is what turns an experiment into something an auditor will sign off.

Be sceptical of the label. If a vendor's agent cannot show you where it runs, what it is permitted to do and what it did yesterday, it may be one of the products Gartner has in mind.

Where Praecise fits

Praecise runs agents on hardware it owns. Praecise Cloud provides compute, MCP servers and backends for agents, built from source, supervised and reversible. Praecise Inbox gives every agent and person an address on a domain the business verified once, with everything said kept in one record a person can open and search. Skerry provides payment rails for agents, including stablecoin settlement, with Bridge and Stripe performing the regulated acts under their own licences. Ipnops lets businesses hand payments, books, orders and customer answers to agents, across the accounts they already use, with every run inside limits the owner sets and every run leaving a record.