How AI is used, limited, measured and switched off.

What AI does here

Two things only: it drafts a one-paragraph synopsis of a finished report, and it copies quotes out of a document you uploaded for a document review. In both cases a validator checks the output against the sources before anyone sees it: a synopsis sentence that cites nothing, or a quote that is not in the document word for word, is discarded and counted.

What AI never does

It never decides a result. Screening outcomes, match tiers, rule hits and case outcomes come from fixed, published rules and from your analysts. No alert is closed by a model. Nothing a model writes is presented as a decision or as a basis for one; every report says so in its limitations.

Disclosure on every report

Every report and audit bundle carries one line, "AI involvement", stating what a model did in that report, or that no model was used. It is part of the evidence record and is checked by the offline verifier like every other field.

Your switch

An administrator can turn AI assistance off for the whole organisation in Settings. Reports are then produced by rules alone, and the disclosure line says so.

Data handling

Only the case's own evidence is sent to the model, with customer-supplied text and web text placed as data, never as instructions. We do not train models on your data. The model provider operates under a contract with no retention of request content; the provider is named in the trust centre on request, not in marketing.

Prompt injection

Text found in documents or web pages that tries to steer an AI reader is detected, kept out of the reading, and flagged for a person to look at.

Measured, not advertised

We publish the method and results of our matching and reading accuracy, including their limits, rather than a headline percentage. Agent-style features that act on alerts are not offered; if we add them, they will start in shadow mode, be measured against analyst decisions, and be opt-in.

Report a concern

If a report contains AI text that is wrong or unsupported, use the report's feedback control or write to support@inack.io; each report is reviewed and the validator updated.