Microsoft counts nearly 40 million registered agents. How many are doing work?
Microsoft's 29 July figure counts registered agents, not active ones. The three agent counts an operations lead should report, and how to collect them.
The short answer
Nobody outside Microsoft knows. On 29 July 2026 Microsoft said nearly 40 million agents were registered on Agent 365, which counts agents created, not agents running or producing results. Inside your company, report three numbers instead of one: agents registered, agents active in the last 30 days and agents producing an outcome someone accepted.
Key takeaways
- Microsoft's figure of nearly 40 million agents, given on 29 July 2026, counts agents registered on Agent 365 across tens of thousands of companies, not active ones.
- Registered, active and outcome-producing agents are three different counts, and only the last one says anything about value.
- In KPMG's survey of 204 US leaders at firms with $1bn or more in revenue (28 April to 25 May 2026), 53% were deploying agents and 26% had full real-time visibility of AI operating costs.
- Our view: an internal agent report should show all three counts, plus cost per accepted outcome, every month.
In this article
What Microsoft said
Microsoft said on 29 July 2026 that Agent 365 had nearly 40 million registered agents across tens of thousands of companies, two months after launch.1 The figure came on its fiscal 2026 fourth-quarter earnings call. It also reported more than 30 million paid Microsoft 365 Copilot seats, with net seat additions more than doubling quarter on quarter.1
Both are real numbers and both measure something narrower than headlines suggest. A registered agent is one that exists in the platform's directory. A paid seat is a licence. Neither says how often an agent runs, what it does or whether anyone used the result.
That is not a criticism of Microsoft. An earnings call reports platform adoption, and registration is the cleanest count a platform has. The problem starts when the figure is quoted as evidence that agents are doing the work of tens of millions of people.
Three counts, not one
Any organisation running agents can count them three ways, and each count answers a different question. Registered agents tell you how much building is going on. Active agents tell you how much is running. Outcome-producing agents tell you how much is useful.
From registered to useful
- Registered: exists in the directory
- Active: ran a task in the last 30 days
- Outcome-producing: a person accepted the result
The gap between the first and second counts is usually experiments. Someone builds an agent to try a feature, tests it twice and moves on; it stays registered. The gap between the second and third is usually quality. An agent that runs every day but whose drafts are rewritten or ignored is active and not useful.
Our view: the third count is the only one that belongs in a business case. Define "accepted" per agent before launch: an invoice match a clerk approved, a reply a support lead sent unchanged, a lead summary a seller used. Then count those, not runs.
What the survey data adds
Surveys of deployment run into the same counting problem. In KPMG's AI Quarterly Pulse, a survey of 204 US leaders at firms with $1bn or more in revenue (28 April to 25 May 2026), 53% said they were deploying AI agents, against 55% the quarter before.2 "Deploying" covers anything from one pilot to dozens of agents in production.
The same survey shows how few firms could produce the counts above. Only 26% of the leaders reported full, real-time visibility of AI operating costs.2 Without cost visibility, cost per accepted outcome cannot be calculated. Our note on the KPMG pulse covers the running-cost check in more detail.
Across all organisations, Gartner's 2026 CIO and Technology Executive Survey (published April 2026; data period not stated) found 17% had deployed AI agents.3 Set that beside tens of millions of registered agents and the picture is clear: a smaller share of organisations are building a very large number of agents, and nobody has published how many of them do useful work.
How do you collect the three counts?
Most of the data already exists in your platforms; the work is joining it. Registration comes from the agent platform's directory. Activity comes from run logs. Outcomes come from the system where the result lands: the ERP, the CRM or the helpdesk.
| Count | Where the data lives | Owner |
|---|---|---|
| Registered | Agent platform directory or admin console | IT or platform team |
| Active | Run logs, filtered to the last 30 days | IT or platform team |
| Outcome-producing | Approvals and edits in the ERP, CRM or helpdesk | The agent's business owner |
| Cost per accepted outcome | Model and tool spend divided by accepted outcomes | Finance with the business owner |
Our recommended monthly agent report. Field sources vary by platform.
The outcome count is only possible if each agent writes its results somewhere a person approves them. That is a design choice, and it is also the safest one: agents that end in a draft for approval are easier to measure and easier to stop. Our scoring model for the first process explains why reversibility matters when choosing where to start.
What to do with the next big number
Ask which of the three counts it is. Platform announcements usually report registrations or seats, surveys report self-described deployment, and vendors report runs. Gartner's recent forecast that agents will outnumber sellers is a count of agents too, as our note on that forecast explains.
Our view: none of these figures should change your plans much. What should change them is your own third count, month on month. Our AI automation page describes how we build agents with an exception queue and a signed-off measure, so that count exists from the first week.
