Microsoft moved Agent 365 from preview to general availability in early May 2026. The product exists because enterprise IT departments lost track of where AI agents were running, what data they touched, and who was responsible for them. Microsoft calls the gap "shadow AI" and sells the answer at $15 per user per month, with prerequisite licensing in the Microsoft 365 E7 suite.

If you run a small business, that product is not for you. The problem absolutely is.

Shadow AI in a small business looks like this. Your salesperson pastes a customer's contract into ChatGPT to summarize it before a meeting. Your marketing lead builds a Zapier workflow that fires AI-drafted replies to inbound leads, with no log of what was sent or to whom. Your bookkeeper uses Copilot to draft variance commentary that goes straight onto a board deck. Nobody asked permission. Nobody can audit it. Most of the time it works fine, until the time it does not.

The 20 MSP looked at small businesses earlier this year and found at least three unsanctioned AI tools in active use at every company they surveyed. None had a written policy. Most owners did not know what their team was already doing.

You will not solve this with a $15-per-seat enterprise product. You solve it with a single page of writing and three small process changes. The page is cheaper. The process changes are what actually move the needle.

What the one-page policy needs to say

Skip the legalese template you found on a vendor blog. A working policy for a 5-to-50-person business answers four questions and stops. Anything longer will not get read.

One. Which tools are approved. Not "AI tools generally." Specific names. ChatGPT Team is fine. The free version of ChatGPT is not, because it trains on conversations by default. Claude.ai is fine. Random AI summarizers your team found on Product Hunt last week are not.

Two. What data never goes into an AI tool. The obvious entries: customer financial records, anything covered by an NDA, employee personal information, anything you have not explicitly published. The less obvious one: full client communications, including emails you wrote. People paste those in constantly without thinking about it.

Three. Who has to approve AI-generated work before it goes out. Not all of it. The categories that matter: anything sent to a client, anything that becomes a financial commitment, anything that goes on the website or social media. Everything else can ship without a second pair of eyes.

Four. How a new tool gets added to the approved list. Make this easy. If the process to request a new tool is hard, your team will skip it and you are back to shadow AI. A Slack message to the owner counts. A two-line form is fine.

That is the policy. One page. Updated quarterly. The version that exists beats the version that is still being drafted.

The three process changes that catch what the policy misses

A policy stops the things people would have stopped if they thought about them. It does nothing about the things people do not realize they are doing. That is what the process changes are for.

First: turn off training on every paid AI account you use. ChatGPT Team, Claude.ai for Work, and Microsoft Copilot all let you opt out of having your conversations used as training data. The setting is buried. Find it. The free tiers do not give you that switch. Pay for the seats your team actually uses and turn off the toggle.

Second: route AI-generated client work through a shared channel. Not as a gate. As a record. A #ai-client-output Slack channel where anyone using AI for a client-facing deliverable drops the prompt and the output before sending. This costs nothing, creates an audit trail, and turns shadow AI into observable AI. The team learns from each other's prompts at the same time.

Third: do an AI inventory once a quarter. Ask everyone, in writing, what AI tools they have used for work in the last 90 days. You will be surprised. The first inventory will look like a list of every tool that launched on Product Hunt. The second will be shorter. By the fourth, your approved list and your actual usage will start to line up.

Where this connects to building things, not just buying them

Most of what we just covered is operational. It costs nothing and does not require a developer.

The part where development matters comes when shadow AI is hiding a real workflow that should be a real product. The salesperson summarizing contracts in ChatGPT is doing a job that should probably live inside your CRM, with the right model, the right prompt, the right audit log, and access controls that match your existing permissions. That is a focused AI integration project, not a per-seat governance platform. The marketing automation built on a Zapier-and-AI duct-tape stack is usually a sign you need proper third-party integrations with logging built in.

The pattern we see most often: a small business catches one piece of shadow AI, panics about the rest, and starts shopping for an enterprise governance product they cannot afford and do not need. The honest answer is that you need a one-page policy, three process changes, and a short list of the workflows that have grown big enough to deserve real software. We covered that last set of decisions in more detail in what your business should actually automate with AI.

Microsoft built Agent 365 because Fortune 500 companies have an average of 12 agents running, often spanning three clouds, and nobody knows which identity owns what. Your small business has the same problem at a fraction of the scale. The fix scales the same way. Write the page. Make the three changes. Buy the software for the workflows that earn it. Skip the rest.