Who Is Actually Running AI Agents in Production?

TL;DR: The honest answer is that nobody outside the vendors knows the full census, and this post will not invent one. What can be done is to define production strictly, a real workload, real users or customers exposed to the output, running continuously, with a named owner who pays for its failures, and then grade the public evidence against that definition. By that standard, clearly real production use exists in narrow, supervised, high volume niches, above all customer support deflection and internal coding and operations assistance with human review. Much of what is marketed as production agents is, on inspection, a pilot, a demo with a customer logo, or an automation workflow with an agent label. The tests below let you sort any specific claim yourself.

Key Takeaways

• Production is a definition before it is a statistic. Without one, every pilot, trial, and conference demo counts, and the resulting numbers describe marketing, not deployment.
• The strongest public evidence of production use is boring: error budgets, escalation paths, pricing based on completed work, and customers who describe failure handling instead of only results.
• Support and supervised internal work are where the evidence is thickest, because the output is checkable and a human fallback exists.
• Vendor homepages are weak evidence. This site's audit found agent vendors publish demos and selected accuracy figures while publishing no error rates, which is the opposite of how mature production systems are usually described.
• Ask any vendor the failure questions: what happens on a wrong action, who is paged, what gets logged, and what the customer can switch off. The answers sort real deployments faster than any case study.

What should count as running an agent in production?

Apply 4 tests. First, real workload: the agent handles live inputs from the business, not a curated evaluation set or a sandbox the sales team controls. Second, exposure: its output reaches customers, money, production systems, or records that matter, so being wrong has a cost someone actually pays. Third, continuity: it runs as an ongoing operation across months, not a launch week pilot with a project team watching every action. Fourth, ownership: there is a named team responsible for monitoring it, handling its failures, and deciding to keep or kill it, the way any other production system has an owner.

A deployment that passes all four is production under any serious definition. A deployment that passes only the first is a pilot with good data. Much of the confusion in public discussion comes from quoting figures built on the loosest definition, any use, any trial, anyone with a licence, alongside figures built on the strictest, autonomous completion of core workflows, and treating them as measurements of the same thing. They are not, and the gap between them is where most agent marketing lives.

Where does the public evidence pass those tests?

Customer support is the clearest case. Resolution based pricing only makes sense if the system is handling real customer conversations at volume, and the public arguments about what counts as a resolution, which this site examined in what counts as a resolution, are themselves evidence of production: nobody litigates a definition that no real workload depends on. The pattern there is instructive, because support agents operate with constrained actions, heavy guardrails, and a human escalation path, and even then the industry's favourite stories include a prominent reversal, covered in this site's Hype Index entry on the Klarna customer service bet, where a headline deployment was later rebalanced toward human service. Production, in the cases that hold up, looks supervised and bounded rather than autonomous and general.

Internal assistance is the second strong case: coding agents used by working engineers, drafting, operations, and research assistants inside companies, where a professional reviews the output as part of an existing job. The review step is not a failure of the agent story; it is the reason these deployments survive. The failure analysis in how an agent deletes a production database shows the other side of the same coin: the closer an agent gets to unsupervised write access to important systems, the thinner the public record of long running success becomes, and the more the documented stories become incident reports.

Why is the rest of the evidence so weak?

Because the sellers control most of it and choose the flattering parts. In the homepage audit behind every agent demo works until you give it real work, agent vendors led with demonstrations and selected customer figures while publishing no error rates at all. A genuinely production hardened category behaves differently: mature infrastructure markets publish reliability figures, status histories, and post-mortems because buyers demand them. Agent marketing instead publishes logos, some from relationships that later turn out to have been brief trials, a pattern documented in this site's churn investigation of the AI SDR category, where expired pilots remained on display as customers.

Surveys do not rescue the picture. Vendor sponsored adoption surveys count trials as adoption and pilots as production, and independent failure statistics circulate in versions that mutate as they are quoted, as this site found when it compared the famous failure percentages against their sources. The result is a public record where the loudest numbers, adoption and failure alike, are the least checkable. A reader who wants the truth about a specific deployment has to grade primary evidence, which is slower than quoting a percentage and is the only method on offer that works.

How can you grade a specific claim yourself?

Ask for the production artefacts. A real deployment has an owner who can say what the agent did last month in operational terms, what it got wrong, how wrong actions are detected and reversed, what it costs to run per completed task including the human review around it, and why its scope is the size it is. It has logs, an escalation path, and a kill switch someone has permission to use. It also usually has a narrower scope than the brochure, because production experience trims agent scope the way it trims any system's scope: toward the tasks where the error rate and the value of automation are both understood.

Claims that dissolve under those questions are not necessarily lies. They are usually a pilot described in production language, a human team with agent assistance described as an agent with human assistance, or a workflow automation that acquired an agent label during a website refresh. Sorting them matters because the buying decision differs: a supervised assistant with a good track record is worth paying for, and an autonomous replacement for a team, at the current documented evidence level, is a bet rather than a purchase.

What does this not prove?

It does not prove that few companies run agents in production; it proves that the public evidence most people cite cannot establish how many do, and that the deployments which document themselves thoroughly enough to grade tend to be narrow and supervised. Private, well run deployments inside companies that publish nothing certainly exist, and nothing here counts them either way. No adoption percentage is stated in this post, deliberately: the sources available do not support one this site would sign, and a made up precision would be worse than the honest method, which is to grade claims one deployment at a time.