Using AI on Company Data: A Safety Checklist Before You Roll Out

Capability is not the question anymore. Whether you can roll it out without leaking is, and it fits on one page.

By Yigit Gok · Updated

Key takeaways
  • Six pre-rollout checks: permission-aware retrieval, written no-training terms, encryption, a content-blind verifiable audit log, a residency answer, an oversharing review.
  • The checklist is go or no-go, not a scorecard: a strong audit log does not compensate for retrieval that ignores permissions.
  • No-training commitments are now standard and inspectable: Anthropic, OpenAI, and AWS all publish them for commercial use.
  • Run the oversharing review before connecting sources, because AI does not create exposure, it industrializes whatever exposure already exists.

It is safe to use AI on company data when six checks pass before rollout: retrieval is permission-aware, the model provider commits in writing not to train on your content, data is encrypted in transit and at rest, access is logged content-blind and verifiably, residency is answered contractually, and an oversharing review has been run on the sources. Fail one, fix it first.

Is it safe to use AI on company data?

Safe is a property you verify, not a feeling you develop after two weeks without incident. Using AI on company data is safe when specific controls hold, and reckless when they are assumed. The canonical how-to is using AI on company data without leaking; this page compresses it into the checklist a team runs before switching anything on.

Treat it like a launch review: someone owns each row, each row produces evidence, and any failing row stops the rollout for its scope until fixed. The list is short because the failure modes concentrate: most incidents trace back to retrieval or to permissions, not to exotic attacks.

What is on the pre-rollout checklist?

Six checks, each with a concrete pass condition. The order matters less than the rule that all six pass. If a vendor cannot produce the evidence for a row, that row fails; a roadmap answer is a no for today, whatever it becomes next quarter.

The pre-rollout safety checklist for AI on company data
CheckWhat to verifyPasses when
Permission-aware retrievalRetrieval is scoped to the asker before the model sees anythingA low-clearance test account cannot surface a named sensitive doc
No-training termsThe model provider's written commitment covers your product and tierThe clause is in the contract or published terms you can cite
EncryptionIn transit and at rest, with key management statedThe vendor documents both, and key custody is answerable
Content-blind auditAccess is logged as actor, object, and rule, not document textYou can review who read what and verify the log's integrity
Data residencyWhere content and embeddings physically liveThe location is contractual, not a support-ticket answer
Oversharing reviewSource permissions match intent before connectionDrifted shares are fixed or excluded before go-live

What do written no-training guarantees look like?

Like plain sentences in commercial terms, which is the point: you are looking for quotable commitments, not marketing pages. Collect the exact clause for every provider in your stack, at your tier, and file it with the rollout decision. The biggest providers all publish theirs, so this check takes minutes.

Start with Anthropic, whose commercial terms provide that "Anthropic may not train models on Customer Content from Services." OpenAI publishes the API-side equivalent in its API documentation: "data sent to the OpenAI API is not used to train or improve OpenAI models (unless you explicitly opt in to share data with us)." Read the scope lines carefully: commitments attach to specific products and tiers, and a consumer plan is usually not one of them.

Which leak paths does the checklist close?

Leaks follow one of four paths: retrieval overshares, the provider trains on inputs, prompts and logs retain sensitive text, or an agent holds broader access than its task needs. The checklist closes each path with a control, and what makes it secure explains the architecture behind the first and hardest one.

The oversharing path deserves its own pass before go-live: an oversharing review walks the sources you plan to connect and fixes permission drift first. Everything after rollout is easier if that review happens before it, because a summary already delivered cannot be revoked.

What happens when a check fails?

A failing check stops the rollout for the affected scope, not necessarily for everything. Connect the sources that pass while the failing ones get fixed, and keep the decision written down: which check failed, what evidence would flip it, who owns the fix, and when it gets re-tested. Scope discipline keeps the project alive without inheriting the risk.

The standing process this checklist samples is setting up access governance; run it continuously once you are live. AIVM Brain was built to make the six rows checkable by default, with plans with SSO and audit published openly; it comes from AIVM and is free to start.

Questions, answered

Is it safe to use AI on company data?

It is safe when verified controls hold: permission-aware retrieval, written no-training commitments, encryption in transit and at rest, a content-blind audit log whose integrity you can check, a contractual residency answer, and an oversharing review of connected sources. Absent those checks, safety is an assumption, not a property.

What should you check before putting AI on company data?

Six things: retrieval scoped to each asker, the provider's no-training clause at your tier, encryption with stated key management, an access log recording actor and object rather than content, where data and embeddings physically live, and whether source permissions match intent. Each check must produce evidence.

How do you stop AI from leaking company information?

Close the four leak paths. Permission-aware retrieval stops oversharing; written no-training terms stop your content improving someone else's model; content-blind logging keeps sensitive text out of audit records; and least-privilege, per-agent identities stop an assistant from holding broader access than its task requires.

Does using AI on company data mean training on it?

Not under current commercial terms. Anthropic, OpenAI, and AWS all publish commitments that customer or API content is not used for training, though scope varies by product and tier. Consumer plans often differ. Collect the exact clause for your stack and keep it with the rollout record.

Give your team and agents one brain they can trust.