How we use AI in our work
We use AI in a limited way. A named person reproduces every finding, sets every severity and signs off every report. Client material never goes to a third-party model service. For how an engagement runs, see how we work.
What AI is used for
- Searching for violations of written invariants, such as balance conservation or lock and unlock symmetry.
- Drafting and summarising.
- Research and drafting of public material, such as the content of this site. That work uses a commercial model provider and never involves client material. Details are available on request.
What AI is never used for
- Deciding that a finding is confirmed.
- Setting severity.
- Signing off a report.
- Handling keys, secrets or credentials.
- Acting on a client system. No AI tool or agent runs against client systems. Testing is done by the named reviewer, inside a written scope list.
- Sending anything to a client.
What a named person always does
- Reproduces every finding. A finding is reported as confirmed only after the named lead has reproduced it. Suspicions are labelled as suspicions.
- Sets every severity, with a one-line reason, as in every report.
- Signs off every report. The principal is the named lead on every engagement.
- Only the principal runs AI tools on client material. No one else does, and no contractors are involved.
Where client material goes
- Client material is never sent to a third-party model service.
- When a model is used on client material, it is a local model. Details of the local model are available on request.
- Client material means code, configuration, architecture documents, findings, scope notes, correspondence and any extract of these.
- Prompts, outputs and files from client-related AI work are held with other client material. They are deleted with it at the end of the engagement, or on request.
- For how we handle data in general, see the privacy page.
No AI on your engagement
A no-AI engagement is offered on request, with no change to scope. Model names and versions are available on request.
Questions
Do you use AI?
Yes, in a limited way. We use it to search for violations of written invariants, and to draft and summarise. Reviews are done by people.
Does AI replace the reviewer?
No. A named person reproduces every finding, sets every severity and signs off every report. AI is a tool inside the work, not the product.
Is client code sent to an AI provider?
No. Client material is never sent to a third-party model service. When a model is used on client material, it is a local model. Client material means code, configuration, architecture documents, findings, scope notes, correspondence and any extract of these.
Can we ask for no AI on our engagement?
Yes. A no-AI engagement is offered on request, with no change to scope.
How does this affect our audit or regulator evidence?
A deebug report is evidence for your own compliance work, not an audit opinion, certification or regulatory approval. The use of AI does not change that.
Want to know more before a scoping call?
If you want more detail on how we use AI, ask. We answer in writing.