Somebody on your team has already pasted something into an AI chat box that they should probably have thought harder about. A contract clause, a customer list, a chunk of source code, a patient note. This is not a hypothetical risk to get ahead of; it is a thing that is happening now, in most organizations, without a policy.
The useful question is not "is AI safe." It is: which account, on which plan, under which terms. Those three things determine the answer, and they vary enormously.
The distinction that actually matters
Broadly, there are three tiers, and they are not equivalent.
Free and personal consumer accounts
Historically the default on consumer tiers has been that your conversations may be used to improve the provider's models unless you opt out, and the opt-out is a setting most people have never opened. Treat anything typed into a personal account as having left your control. For regulated data, that is the end of the conversation.
Business, team, and enterprise tiers
The major providers offer business tiers where inputs are contractually excluded from training by default, with administrative controls and retention settings. This is a genuinely different legal posture, not a marketing tier. If your staff are going to use AI — and they are — this is the minimum acceptable footing.
API access
When software calls the model through an API, business terms typically apply by default, and providers offer zero-data-retention arrangements for qualifying use cases. This is the tier a custom build runs on, and it is the one where you can actually make commitments to your own customers.
How to find out what you agreed to
Do not rely on a blog post — including this one — for your specific situation. Check these, in order:
- Which accounts exist. Ask, or check expense reports for personal AI subscriptions. Unmanaged personal accounts are the actual exposure.
- The plan tier on each account, and whether it is under a business agreement or someone's personal credit card.
- The data controls in the admin console: training opt-out, retention period, workspace sharing.
- Whether a DPA or BAA exists if you handle regulated data. If you cannot produce one, you do not have one.
Shadow AI is the real problem
The risk almost never comes from the sanctioned, configured, business-tier deployment. It comes from the eleven tools nobody approved: the browser extension that summarizes email, the meeting notetaker someone connected to the calendar, the free tier of a writing assistant with full document access.
Those integrations often request broad permissions and are invisible to whoever is nominally responsible for security. Finding them is an inventory exercise, and it belongs in the same category as the rest of your endpoint security and security posture work.
What to do this month
- Write the policy, and make it short. One page. What is never pasted anywhere (credentials, regulated records, anything under NDA), what is fine, and which tools are approved. A policy nobody reads changes nothing.
- Give people a sanctioned option. Banning AI outright reliably produces more shadow AI, not less. A business-tier account your staff can actually use is the cheapest control available.
- Inventory the integrations connected to email, files, and calendars. Revoke what nobody can justify.
- Train on the failure mode, not the technology. People need to recognize the moment they are about to paste something regulated. That is the whole lesson.
- Log it. If you are in a regulated industry, an auditor will ask what your staff used and what controls applied. Have an answer before they ask.
Where regulated data changes the answer
Under HIPAA, CJIS, or CMMC obligations, the bar is higher than "we ticked the training opt-out." You need the contractual instrument, a documented data flow, and a defensible retention story. In practice that usually pushes you toward API-based access inside your own environment, where you control the boundary — which is one of the more common reasons organizations move from a chat subscription to a purpose-built internal tool.
If you are working through what that looks like against a specific framework, our compliance and professional services work covers exactly this ground, and the guide to buying AI responsibly goes deeper on the procurement side.
Ask us to review your current setup — it is usually a short conversation with a clear answer.