A client contract gets pasted into a personal ChatGPT account because someone needs a summary before a meeting. The employee solves the immediate problem, and nobody is at fault: pasting a document into an assistant to save an hour of reading is what the tool is good at. The gap is that the company never decided what its rules are, so each employee is running their own.
AI privacy at a company isn't a compliance document. It's four paths data leaves by, a short list of data categories, and one paragraph of rules that say which material can travel which path. None of it needs a compliance team, and the first version can fit into an afternoon.
The four paths data leaves by
Most AI tools expose company data through four common paths, and each path needs a different control.
Typed prompts. The words an employee types around the question carry the company material: client names, unreleased numbers, the details of a dispute. It's the path people picture, and the easiest to govern, because the approved workspace handles it.
Uploaded files. A contract, a spreadsheet of customers, a screenshot of an internal dashboard: one upload carries more than a prompt does. The casual "summarize this" habit does most of its damage here, since a 40-page PDF hands over more than anyone would type.
Connectors and integrations. Link an AI tool to your email, calendar, or document drive and the tool gains standing access: it reads what arrives, not just what someone pastes. That's convenient for a daily-brief workflow and worth treating as a decision rather than a checkbox clicked during onboarding.
Settings. The quiet path. Whether content trains models by default, how long the provider keeps it, and who inside the tool can read conversations: settings decide what happens to everything that traveled the other three paths.
Name the data categories first
Rules attach to categories of data, not to tools. Tools change; the categories stay. Four buckets cover most companies:
- Public: marketing copy, published pricing, the website.
- Internal: process docs, org charts, meeting notes. Embarrassing if leaked, not existential.
- Confidential: client lists, financials, contracts, source code, the roadmap.
- Regulated: health records, payment data, background checks. Data with laws attached.
The buckets turn a vague instruction ("be careful with data") into an assignable rule: public and internal material can go into the approved tool freely, confidential material goes in only through the business workspace once its settings are confirmed, and regulated material needs written approval from a named person.
Put the business defaults in place
Two defaults decide most of the privacy picture: training and access. We'd confirm both before rollout, not after an incident.
On personal plans, the consumer products differ. OpenAI's personal plans have data sharing enabled by default, with a setting to turn it off (OpenAI, 2026). Anthropic lists model training as opt-out across its Claude plans (Anthropic, 2026). Business workspaces start from the other side: OpenAI doesn't use ChatGPT Business workspace content, inputs or outputs, for training by default (OpenAI, 2026), and both vendors' business plans add the controls a company needs either way: admin roles over members, centralized billing, and usage visibility (OpenAI, 2026).
The practical move: if confidential categories are in play, the approved tool runs on a business plan, and an admin confirms the training setting, the retention policy, and who can see conversations before anyone pastes client material. How to Evaluate AI Tools for Your Team walks the full pre-purchase check.
Write the one-paragraph policy
A policy people can't remember won't guide their next decision. One paragraph people have seen works better. Copy, edit, send:
Use your company-provided [tool] account for work with AI. Public and internal material is fine. Confidential material (client names, financials, contracts, code) goes in only through the company workspace, and never through a personal account. Regulated data (health, payment, background) stays out of AI tools unless [owner] approves it in writing. Not sure which category you're holding? Ask [name] before you paste.
The paragraph works because it names three things: the categories, the boundary, and the person.
Assume the tools are already in use
Adoption happened through browsers, one employee at a time, usually before anyone wrote a policy. A ban with no alternative reads as one more restriction and tends to get quietly ignored; an approved account with the right defaults gives people an easier path than the forbidden one. The ask: find who's already pasting work into personal accounts, hand them approved ones, and show the workspace difference. Skip the lecture.
Revisit when something changes
The settings you verified hold until the vendor changes them. OpenAI is retiring custom GPTs on December 11, 2026 and replacing them with plugins (OpenAI, 2026); vendors repackage features without asking. A quarterly ten-minute recheck of the training setting, the retention policy, and the connector list can catch the drift. Add one trigger outside the calendar: any new tool or new data category starts the check again.
When you need the pieces
- How to Evaluate AI Tools for Your Team: the pre-purchase version of the settings check, seat pricing included.
- Is ChatGPT Safe to Use? What Beginners Need to Know: the personal-side settings (history, temporary chats, the training toggle).
- How to Use AI for Customer Research: the privacy step for customer data specifically, including the indirect-identifier check.
FAQ
What is shadow AI?
Employee use of AI tools the company didn't approve, usually personal accounts on personal plans. The fix is an approved path with better defaults, not detection software.
Does ChatGPT train on my business data?
Not on a business workspace by default: OpenAI doesn't use ChatGPT Business workspace content, inputs or outputs, for model training. Personal plans are the opposite, with data sharing enabled by default and a setting to turn it off. Anthropic treats model training as opt-out across its Claude plans.
Can employees use personal AI accounts for work?
They can, and the terms are the problem: a personal account has no admin oversight, no centralized controls, and, on OpenAI's personal plans, training enabled by default. Company material typed into a personal account also leaves with the employee who typed it.
Do we need an enterprise plan for AI privacy?
Not for a small team. The self-serve business plans carry the defaults that matter: no training on workspace content, admin controls over members, centralized billing. Enterprise plans add single sign-on, compliance certifications, and dedicated support, which start to matter at larger headcounts or in regulated industries.
Who should own AI privacy at a small company?
One named person, and at the smallest size that is the founder or owner. The job is three questions: which data categories can go where, is the workspace setting confirmed, and who answers when someone isn't sure. It fits in an afternoon a quarter.