Grounded, not trained

The first question every security team asks is the right one: does this thing train on our data?

For ALIF the answer is no — and the reason is not just policy. Training on customer data is the wrong architecture for enterprise answers, even before it’s a compliance problem.

Why training is the wrong tool

A model trained on your documents blends them into weights. Three consequences follow, all bad for the enterprise:

• No provenance. The answer can’t point back to the contract, spec, or thread it came from — so nobody can check it.
• No freshness. Your data changes daily; a trained model is a photograph of last quarter.
• No boundaries. Once knowledge is in the weights, per-user permissions can’t contain it.

For a consumer chatbot these are acceptable trade-offs. For case-outcome analysis or a security questionnaire, they are disqualifying.

What grounding means instead

A grounded agent leaves your data where it lives. At question time, it retrieves the relevant sources from the systems you already run — respecting the permissions those systems already enforce — reasons over them, and answers with citations attached.

The properties reverse:

• Every answer traceable. The citation is the accountability. A lawyer can check the source; an auditor can replay the trail.
• Always current. The agent reads today’s documentation, not a training snapshot. When the spec changes, the answer changes.
• Permissions intact. Someone who can’t open a document can’t get its contents from the agent either.

This is what “accountable for every answer” means in practice. Grounding is not a feature of the agent — it is the contract between the agent and the enterprise.

The trust perimeter

Grounding handles the answers. The rest of the perimeter is conventional, and it should be: SOC 2 Type 2, ISO 27001, GDPR, SAML SSO, and regional deployment options for data residency. Security reviews are won with boring answers.

The test to run

Ask any AI vendor two questions. Where did this answer come from? — if the system can’t cite it, you can’t trust it. What do you do with our data? — if the answer involves training, the three problems above are now yours.

Grounded, not trained. It’s the difference between an answer you can act on and one you have to verify by hand — which was the work you were trying to save.