Insights

Your staff installed AI agents last month. Do you know which ones?

Staff may already have AI tools on office PCs that can open files and send data out. What Microsoft announced this week, what is generally available now, and how a one week sweep finds them.

Shane Donoher · Published November 20, 2026

The short answer

Staff at small firms are already putting AI tools on office PCs. Some of those tools can open files, run commands and send data out. Most 10 to 100 person firms have no list of what is installed, who owns it, or what it can reach.

This week, at Microsoft Ignite (17 to 20 November 2026), Microsoft announced controls for those tools. One of them is generally available now: Microsoft Defender can block misbehaving local AI agents on Windows. Other important pieces are still in preview, including Entra identities for agents, Intune limits on what an agent can reach, and Purview controls for sensitive data in prompts, which Microsoft has set for a January 2027 preview. Separately, new Copilot Business purchases made through resellers are planned to turn on usage based billing by default between January and March 2027. (Microsoft Agent 365 blog, Ignite 2026; Microsoft Intune blog, Ignite 2026; Microsoft security blog, 17 November 2026; Microsoft Partner Center, November 2026 announcements)

The job for a San Diego clinic, law firm or accounting practice is practical. Build a list of the AI tools already in use, decide which ones may stay, and put security and infrastructure limits around the ones you keep. That is the AI Agent Safety Sweep.

What Microsoft announced, and what is actually available

The dates matter, because a preview is not the same as a control you can rely on today.

Generally available now, on Windows. Microsoft says Defender runtime protection for local AI agents is generally available on Windows. In plain terms, the security software already on many office PCs can now stop a local agent that starts to misbehave. Microsoft also says Defender and Entra Global Secure Access can discover more than 60 kinds of local agents, including tools such as ChatGPT Desktop, Claude Code and GitHub Copilot CLI. (Microsoft Agent 365 blog, Ignite 2026)

In preview now. Entra can give a discovered agent its own identity. Intune, with Microsoft Execution Containers, can set file and network boundaries so an agent only reaches what you allow. Intune can also mark a device running an unsanctioned agent as noncompliant and, with Conditional Access, block that device from company resources. Those are useful, but they are still in preview, so treat them as early controls, not as a finished baseline. (Microsoft Agent 365 blog, Ignite 2026; Microsoft Intune blog, Ignite 2026)

Preview planned for January 2027. Microsoft has said Purview will preview two data controls in January 2027. One is grounding protection, which Microsoft describes as a way to start Copilot without labelling every file first, based on sensitive information types. The other is data loss prevention for Copilot prompts and for local Windows agents, so sensitive data can be blocked from going into a prompt. Neither is generally available today. (Microsoft security blog, 17 November 2026)

Billing change, January to March 2027. For new Copilot Business purchases made through the Cloud Solution Provider programme, Microsoft plans to turn on usage based billing by default between January and March 2027, with a default spend limit of 4,000 Copilot Credits per user per month, which an admin can change. This applies to those new reseller purchases as described by Microsoft. Existing subscriptions may be on other terms, and no one firm's bill can be forecast from it. (Microsoft Partner Center, November 2026 announcements)

What this means for a small regulated firm

You do not need a large security team to be affected. If someone in a 20 person office installed a desktop AI tool last month, that tool may already be able to read a shared drive, a mailbox or a case file, depending on what the person's account can open and what the PC will allow.

Most of the risk is quiet. Client or patient text goes into a tool the firm never approved. A login or an API key is pasted into a prompt. A personal account on a work laptop keeps working after the person leaves. We wrote earlier about what never goes in the prompt and what to do if sensitive data goes into the wrong AI tool. This post is the step before those problems: knowing which tools are there at all.

For a wider picture of unsanctioned tools, see how to find where firm data is going.

How AI, security and infrastructure join up

An AI agent only does what the office setup lets it do. The identity of the person, and of the agent itself, decides which files and mailboxes it can see. The device, the network path and the firewall decide where it can send data. The security software on the PC decides whether a misbehaving agent is stopped. If any of those layers is missing, a new AI tool will inherit every gap already in the tenant.

So a sweep covers the AI tools and also sign ins, device management, Conditional Access and the security tools already on the laptops. On a recent security engagement, an anonymous client's Microsoft Secure Score went from 51% to 88.78%. The score is not a measure of agent use, but it is a useful picture of the identity, device and access foundations that any AI tool will sit on.

What an AI Agent Safety Sweep covers

The sweep is a one week, one time piece of work for a 10 to 100 person firm. It does not promise to find every possible tool on every personal phone. It does build a clear list of the AI tools in the Microsoft 365 tenant and on the managed Windows PCs, and it puts a decision against each one.

  1. Find what is already there. Check managed PCs and the tenant for AI tools and local agents, including the kinds Microsoft says it can now discover. Note whether each one is a firm account or a personal account, and which device it runs on.
  2. Decide what may stay. For each tool, record an owner in the firm, what work it is for, and whether it may handle client, patient or firm data.
  3. Turn on what is generally available. Confirm Defender is running on every managed Windows PC, and that runtime protection for local agents is in place where the licence allows it.
  4. Set limits on the tools you keep. Where preview controls are acceptable to the firm, give approved agents an identity, and set file and network boundaries so they cannot wander. Mark unsanctioned tools so a noncompliant device can be kept away from company resources.
  5. Write the follow up. A one page list of what was found, what was blocked, what was allowed, and what to revisit when the January 2027 Purview previews and the Copilot billing change arrive.

After the sweep, a monthly add on, Managed AI Agent Controls, can keep the list current, watch for new installs and alert on misuse. That is optional. The first job is the list.

We have not confirmed which Microsoft 365 Business Premium plans include each of these controls, so a sweep starts by checking the licences you already have, not by assuming a feature is included.

What is still unknown

As of 19 November 2026:

  • We have not confirmed, from a Microsoft source, which Business Premium plans include Defender runtime protection for local agents, Entra agent identities or the Intune agent limits. Do not treat any of those as included until the tenant and the licence are checked.
  • Preview features can change, arrive late, or be limited to certain plans. Entra agent identity, Intune file and network boundaries, and the January 2027 Purview controls should be described to staff as preview, not as a finished standard.
  • Microsoft has not, in the sources we checked, published a figure for how many small firms already have unsanctioned agents. We will not invent one. The practical question is whether your firm has a list.
  • The Copilot billing change is planned for new Cloud Solution Provider purchases between January and March 2027. Existing subscriptions may be on different terms. Check yours rather than assuming the default already applies.

We will update this post if Microsoft publishes a material change to the generally available or preview status of these controls.

How we help

The AI Agent Safety Sweep is a one week review for regulated San Diego firms. We build the list, help you decide what stays, turn on the generally available Windows protections, and set the limits the tenant can support today. If you are not sure which AI tools are already in use, an AI Safety Check is a good first step. For the identity, device and access work the sweep sits on, see our cybersecurity services in San Diego and Microsoft 365 services.

If you want the list kept current after the first week, we can add Managed AI Agent Controls on a monthly basis. Rated 5.0 on Google. Book a 15 minute call to talk it through.

FAQs

What is a local AI agent?

A local AI agent is a tool that runs on a person's PC, or alongside it, and can take actions such as opening files, running commands or sending data, rather than only answering a question in a browser tab. Microsoft's own examples of the kinds of local agents its tools can discover include ChatGPT Desktop, Claude Code and GitHub Copilot CLI. (Microsoft Agent 365 blog, Ignite 2026)

Can we block misbehaving agents today?

On Windows, Microsoft says Defender runtime protection for local agents is generally available. That is the control that is ready now, provided the PC is managed, Defender is running, and the licence covers it. The identity and Intune limit features are still in preview. (Microsoft Agent 365 blog, Ignite 2026)

Do we have to ban every AI tool?

No. The point of the sweep is to make a decision about each tool. Some tools can stay, on a firm account, with an owner and with limits on what they can reach. Tools that have no owner, or that sit on a personal account with access to client or patient data, are the ones to deal with first.

Will this stop us using Copilot?

Copilot can stay. A sweep means knowing which AI tools are in the office, including Copilot, and putting the right identity, device and data controls around them. The January 2027 Purview previews, if they arrive as described, are meant to help with prompts and file access. They are not available today.

Does the new Copilot billing start on 1 January for everyone?

No. Microsoft has said that usage based billing is planned to turn on by default for new Copilot Business purchases made through resellers, in a rollout from January to March 2027, with a default cap of 4,000 Copilot Credits per user per month. Check your own subscription and purchase route before you change any budget. (Microsoft Partner Center, November 2026 announcements)

Do we need this if we already have an IT provider?

Ask them for a written list of the AI tools on your managed PCs and in your tenant, and ask whether Defender's local agent protection is on. If they cannot show you either, the sweep is still worth doing. The list is what lets security, IT and the AI rules work together.

---

*This post is general information, not legal, tax or compliance advice. We checked Microsoft's Agent 365 and Intune Ignite 2026 blogs, the Microsoft security blog of 17 November 2026, and the Partner Center November 2026 announcements on 19 November 2026. Product availability, preview status and licensing can change, so check the current position for your tenant and talk to your own adviser about your firm's situation.*