Governance & responsible AI
How we build AI systems for clients
- A model proposes; a policy-checked service commits.We do not build systems in which a model writes directly to a system of record. The distinction between proposing and committing is the whole audit story, and it is architectural rather than procedural.
- The decision boundary is written before it is coded.Which decisions are automated, which are recommended, and which stay human, agreed with an accountable owner and versioned alongside the code that enforces it.
- Every decision leaves reconstructable evidence.Inputs, policy version, output and the human action. The test is whether a decision can be reconstructed two years later, not whether events were logged.
- We decline automations without labelled history.Without a record of past outcomes there is nothing to evaluate against. We say so before the engagement rather than during the pilot.
- We name what should not be automated.The assessment deliverable includes the honest list. It has ended engagements, and that is the correct outcome when it does.
Machine-assisted drafting, and this site
Our position, stated publicly because we are asked (D-D): machine assistance is permitted for drafting, never for claims, and every page carries a named practitioner who reviewed it. A firm selling AI transformation that quietly publishes unreviewed generated material under its own name has answered the question about its judgement.
Concretely: no figure, client reference, capability claim or certification status reaches this site without a human verifying it against a primary record. Where a fact is unverified, the page says so — which is why phrases like to confirm and permission pending appear throughout rather than being tidied away.
Editorial review
Every page in the reference library passes a named editorial owner and a practitioner reviewer before publication (D-C). Volume is not a licence to publish unreviewed material under the firm’s name, and the cadence is set by review capacity rather than the other way round.
Governance of what we build
Administrative keys stay with the client
We do not retain the ability to change a client’s deployed contracts after handover. Where an upgrade path exists, the client holds it under an independent approval set.
No capability claim without the running software
Platform pages are verified against the software before publication, and a claim is removed if the software changes.
No client named without written permission
Fresh permission, not a permission granted for a different purpose years ago.
The related-party relationship is disclosed everywhere it applies
At body weight, on screen, without interaction — including on every platform page that touches the agent runtime.
Security posture, certifications and data handling are published separately on the trust centre.