Use AI assist within its limits

What AI assist does

AI assist reads a document or pasted text and suggests values for the fields your category schema expects. It can also produce a completion checklist from a passport's schema, and a fallback carbon estimate where no declaration exists.

The workspace shows the provider, the model and the recent calls so you always know which provider answered, and it never displays the credential.

The guard rails

Three rules make a suggestion usable. First, the field must exist in your schema, including your own extensions, so a model cannot invent a field. Second, the evidence it quotes must actually appear in the document - if it does not, the suggestion is dropped and the reason is shown. Third, the result is a draft: nothing is written into a passport, and the response marks itself as requiring human review.

The model runs at zero temperature and returns a bounded number of suggestions, so the same document does not produce a different answer each time you look.

Honest degradation

If the provider is unavailable or answers with something unusable, the request does not fail and does not return an empty result pretending success. The platform falls back to its deterministic extraction, flags the response as degraded and records the provider error, so you can see that the answer came from rules rather than from a model.

A document with no readable text is reported as metadata-only with an explanation instead of a fabricated reading.

Usage and privacy

Each call is accounted with the provider, the model, latency, token counts and whether it failed or degraded - never the prompt or the response, because those contain your documents. Usage is visible per period, and calls count against the plan's hourly AI budget.

The rule to keep

AI drafts, a person decides. The platform never publishes an AI statement on your behalf, never uses a model as the compliance decision, and never lets generation stand in for a declaration you cannot support. Where a machine draft is used for translation, the same principle applies: it arrives as a draft and needs an approval.

Screen-by-screen reference

These pages were checked in the organization console on 8 October 2026 using an isolated sample organization. Controls can differ by role, plan, record state and configured providers. Each page below explains its purpose, a practical workflow, the visible controls and what to verify afterward.

AI assist (/settings/ai)

The workbench proposes category fields from stored document text or pasted text while showing provider health and usage.

Choose a stored, readable document or paste source text, then set the correct product category and document type before extraction. Compare every suggested value and citation with the original file, especially when provider health indicates a fallback or degraded call. Suggestions remain proposals until a human writes and validates passport data.

This tool writes suggestions, not passport declarations or compliance decisions.

Controls and options

• Choose a stored document or paste text, then choose category and document type.
• Run extraction and read each proposed field with its cited evidence and confidence.
• Inspect Provider and Usage to see whether a model answered or deterministic extraction handled a degraded call.
• Review Recent calls for latency and token counts; never accept a proposed value without checking the original file.