The short version: nobody, including the vendors selling them, can tell you a real total number of AI agents in existence, because anyone with a no-code builder can spin one up in under five minutes and most of them never get used past the demo. The numbers that do matter are the adoption ones: roughly 10% of organisations using generative AI have deployed an agent, Gartner expects over 40% of agentic AI projects to be scrapped by 2027, and a widely cited MIT study found 95% of enterprise generative AI pilots deliver no measurable return. If you’re a small business owner, the count is a distraction. The failure rate is the number worth your attention.
Why “how many AI agents exist” is the wrong question to start with
I get asked this a lot on calls, usually by someone who has just read a scary headline about millions of agents flooding the workplace. The honest answer is that there is no census. There is no registry. Nobody at Microsoft, Salesforce, Google, or OpenAI can hand you a total, because an agent is not a product with a serial number. It is a bit of software, sometimes just a prompt and a tool connection, that anyone can build inside a platform like Zapier, Microsoft Copilot Studio, n8n, or Salesforce Agentforce in the time it takes to make a cup of tea.
Asking how many exist is a bit like asking how many spreadsheets exist. Technically countable in theory, practically meaningless, because the number changes by the minute and most of them are duplicates or abandoned after one use. I track new agent launches and platform updates every week for my AI News This Week roundups, and in the 16 August edition alone I covered three separate companies announcing their own agent builders in the same seven days. A fortnight earlier, in the 26 July update, it was two more. The week before that, in the 19 July roundup, another platform added agent building to its existing chatbot tool. That pace alone tells you a fixed number is fiction. The count you read in any headline is out of date before the article is published.
The numbers that do exist, and what they measure
What we do have are adoption surveys, and these are far more useful than any attempt at a total headcount. Here is what the research shows:
The same thinking applies city by city; for a local worked example see AI consultant in Glasgow.
I keep a local page for Dublin with the workflows that go first there: AI consultant in Dublin.
- Capgemini’s research on enterprise AI found that only around 10% of organisations already using generative AI had deployed AI agents, but 82% said they planned to integrate agents within one to three years.
- Deloitte’s 2025 predictions put the figure at roughly a quarter of companies using generative AI launching agentic pilots or proofs of concept in 2025, rising to about half by 2027.
- Gartner predicted that by 2028, 33% of enterprise software applications will include agentic AI, up from less than 1% in 2024, with agents handling around 15% of day-to-day work decisions autonomously.
- Gartner also warned, more soberly, that over 40% of agentic AI projects will be cancelled by the end of 2027 because of rising costs, unclear business value, or weak risk controls.
- MIT’s widely reported 2025 study on generative AI pilots found that 95% of them failed to deliver a measurable return to the business that funded them.
Put those together and you get a much clearer picture than any “there are now X million agents” headline. Interest and building activity are enormous. Deployment is still low. Cancellation and failure rates are high. That gap, between how many agents get built and how many earn their keep, is the story most coverage skips past.
What counts as an “AI agent” (most vendors are stretching the word)
Part of why counting is impossible is that the word “agent” has been stretched to breaking point by marketing teams. A true agent, in the technical sense, perceives a situation, decides on a course of action, and takes it without a human clicking approve at every step, often chaining several tool calls together to finish a task. A chatbot that answers a customer question is not an agent. A Zapier automation that moves a lead from a form into a spreadsheet is not an agent, it is a workflow. A drafting assistant that writes an email for you to review and send is not an agent either, it is a copilot.
But sit through a demo from most of the major platforms and you will find all three of those relabelled as “agents” because the word sells better right now than “automation” or “assistant” did two years ago. This matters for adoption because a business owner who buys “an agent” expecting autonomous decision-making, and gets a glorified email drafter that still needs someone to press send, feels misled. That gap between the marketing word and the actual behaviour is quietly eroding trust before the technology even gets a fair test.
A story from my own client work
A client of mine, a fourteen-person recruitment agency in Kent, upgraded their CRM last spring to a plan that came bundled with eleven pre-built agents: a job-ad writer, a candidate-screener, a follow-up-email drafter, a sourcing assistant, an interview-prep bot, and six others with similarly optimistic names. It sounded brilliant on the sales call. Eleven agents for the price of one licence upgrade.
Eight months later they actively use one. The meeting-notes summariser, because their consultants were losing hours every week rewriting call notes before they could send them to hiring managers, and that one agent alone saves roughly six hours a week across the team. The other ten sit unused, quietly appearing on the monthly invoice, never switched on because nobody had the time to learn them, trust them, or connect them to the right data.
That is not an unusual story. It is close to the norm. Businesses are not short on access to agents. They are short on the time and process needed to test one, prove it works on a narrow task, and only then expand. Buying a bundle of eleven does not solve that problem, it just multiplies the number of things nobody got round to trying.
The uncomfortable part of this that most coverage avoids
Here is the bit that gets glossed over in a lot of the excitable writing about agents. The high failure and cancellation rates are not mainly a technology problem. The models are good enough for a huge number of narrow, well-defined tasks already. The failure comes from busines
Where to check the details
Related reading: how many instagram reels exist.
