Your team is already pasting company data into AI tools

Somewhere in your office this week, someone highlighted a customer email thread, copied it into a free AI chatbot and typed “help me write a response to this.” Someone else pasted a contract in and asked for a summary. A third person has a free transcription tool sitting in on their calls, taking notes. Nobody asked permission because nobody thought there was anything to ask about.

This is not a prediction about where AI adoption is heading. It is a description of a normal Tuesday in a 20-person company. Survey after survey through 2024 and 2025 found large shares of employees using AI tools at work without their employer’s knowledge, with many admitting to pasting in internal or customer data. Whatever the exact percentage in your office, the practical assumption should be simple: it is already happening.

The question is not whether to allow AI tools. That decision was made for you, informally, by your team. The question is whether usage happens inside guardrails you chose or outside them.

Why this is a data problem, not a technology problem

The risk in casual AI use is not science fiction about machines. It is mundane data handling:

• Consumer-tier AI tools may retain prompts and, depending on settings and terms, use them to improve models. Data pasted into a free personal account has left your control in a way that is hard to reverse
• The accounts are personal. When the employee leaves, the chat history full of customer details leaves with them, exactly like the personal-account file sync problem
• Free transcription and meeting bots record entire conversations, including the parts where people discuss pricing, personnel or a customer’s private situation
• Pasted content is invisible to you. There is no log, no inventory, no way to answer “did any of our client data end up in that tool?” when a customer or insurer asks

If this pattern sounds familiar, it should. It is the same dynamic we described in our shadow IT and SaaS inventory post: useful tools adopted faster than anyone tracks them. AI tools are shadow IT with a bigger appetite for pasted data.

Why bans fail

The reflex response is a ban. Some businesses announce one, feel briefly responsible and move on. 3 things then happen, reliably:

1. Usage continues on personal phones, where you have no visibility at all
2. Your most productive people, the ones using the tools to move faster, learn to hide their workflow from you
3. The company forfeits real productivity gains while still carrying the data risk it was trying to eliminate

A ban converts a manageable data-governance question into an invisible one. The goal should be the opposite: make the sanctioned path easier than the shadow path.

The sane setup for a small business

You do not need an AI governance committee. For a 10 to 50 person company, the workable setup has 4 parts and takes about 2 weeks to stand up.

1. Pick 1 sanctioned tool and pay for the business tier

Choose a primary AI assistant and buy it properly. The business tiers of the major AI platforms typically run $20 to $30 per user per month and buy you the things the free tiers do not:

• Contractual terms stating your data is not used to train models
• Admin controls: you can add and remove users, which means offboarding works
• Centralized billing instead of expensed personal subscriptions
• In some tiers, audit and retention controls

Which platform matters less than the tier. The consumer free tier of anything is where “we do not know where our data went” lives. If Microsoft 365 is your backbone, Copilot has the advantage of respecting the file permissions you already set. Whatever you pick, license the people who are already using AI first, since they are your current exposure.

2. Write plain rules using 2 or 3 data classes

Skip the 12-page policy. Define what may and may not be pasted, in language people can hold in their head:

• Never paste: customer financials, employee personal information, passwords or keys, anything covered by an NDA and anything health-related
• Ask first: contracts, customer lists, pricing, internal financials
• Fair game: drafting your own emails and documents, public information, general research, code without secrets in it

That is the whole framework. The test of a good rule set is whether the office manager can recite it, not whether it would survive an audit committee.

3. Put it in writing, 1 page

A short written policy does 3 jobs: it tells employees the company has a position, it gives managers something to point to and it answers the AI questions now showing up on cyber insurance applications and customer security questionnaires. Name the sanctioned tool, state the data rules, state that personal accounts are not for company data and say who to ask when unsure. Done.

4. Teach it in 10 minutes, then revisit quarterly

Fold the AI rules into whatever security awareness rhythm you already run. One short session covering why pasting a customer’s information into a free tool is the same act as emailing it to a stranger, plus the “which class is this data?” habit. Then put a 30-minute review on the calendar each quarter, because the tools, the terms and the ways your team uses them will change faster than any policy you write today.

The meeting-bot corner case that bites first

One specific scenario deserves its own rule because it involves other people’s consent: AI notetakers joining calls. When an employee’s personal transcription bot sits in on a client call, you have recorded a client conversation into a consumer tool without asking. Some clients will care a great deal. The rule we recommend: only the sanctioned notetaker, announced at the start of the call, and never on calls where legal, personnel or sensitive client matters are discussed.

Guardrails beat gates

The businesses handling this well are not the ones with the strictest documents. They are the ones where the easy path is the safe path: a paid tool everyone knows is approved, 3 data rules everyone can recite and a culture where asking “can I paste this?” is a normal question rather than a confession.

That is a 2-week project and a modest line item, against a risk category that is otherwise completely unmeasured in most small businesses.

If you want help choosing the sanctioned tool, wiring it into Microsoft 365 and writing the 1-page policy your insurer and customers are starting to ask about, Atticus Rowan can help. We build practical guardrails for teams that are already moving fast.