AI transparency
We build with AI, so we owe you a plain account of where it is used, what a human checks before it reaches you, and what we will never do with your data. This is a voluntary disclosure. It is not a certification, and we hold none.
Where we use AI
Drafting and summarisation
Language models draft and summarise text — descriptions, summaries, and structured extracts from documents a user has supplied. Output is a draft, and a draft is never the final word on a decision that affects a person.
Matching and ranking assistance
Models help surface and order candidate matches against a stated requirement. Ranking is an aid to a human reviewer, not an automated selection, and a human can reorder or discard the ranking entirely.
Structured data extraction
Models convert unstructured submissions into structured fields. Extracted values remain editable by the person they describe and are not treated as verified facts merely because a model produced them.
Agent-facing endpoints
We publish a machine-readable interface so external AI agents can query public information directly. Those endpoints return the same facts a human reader gets — we do not serve different content to agents.
What a human reviews
The rule we hold ourselves to is narrow and testable: a model can propose, a person disposes.
- No model output is delivered to a customer as a final decision without a human in the loop.
- No adverse decision about a person — rejection, removal, or withdrawal of access — is made by a model alone.
- A human can override any model output, and the override is what ships.
- Where a model output is shown as a draft, it is presented as a draft rather than as a conclusion.
What we do with your data
Your data does not train third-party models
We do not submit customer or applicant personal data to third-party model providers for training, and we do not consent to provider-side training on our submissions. Where a provider offers a training-opt-out or zero-retention path, we take it.
You can reach a human
Every AI-assisted surface has a human contact path. You are never required to resolve a dispute with an automated system, and asking for a human is not treated as an escalation.
Retention is bounded
Prompts and model outputs tied to an identifiable person follow the same retention limits as the underlying record they belong to. They are not retained separately or indefinitely as a training corpus.
How we organise this
We organise our own AI governance around the four functions of the NIST AI Risk Management Framework. This is a self-directed choice of structure. The NIST AI RMF is voluntary, it confers no certificate, and there is no body that certifies against it — so we make no compliance, conformance or assessment claim of any kind here.
Govern
A named human owns AI-use decisions. Where a model touches a person-affecting workflow, the decision to deploy it is recorded rather than emergent.
Map
We enumerate where models are used and which of those uses could affect a person, so the person-affecting subset is small, known, and reviewable.
Measure
Model-assisted outputs are checked against human judgement on the surfaces where a human already reviews them. We do not currently publish quantitative model-performance metrics, and we will not publish any until they are measured.
Manage
Human override is the standing control. Where a model behaves unacceptably, the remedy is to remove it from the path rather than to tune it in place while it continues to affect people.
What we do not claim
- We hold no AI-related certification, and no third party has audited or assessed our AI governance.
- We do not publish model-performance, accuracy or bias metrics, because we have not measured and validated them. When we do, we will publish the method alongside the number.
- We do not claim our models are unbiased. We claim that a human reviews the outputs that can affect a person, which is a control, not a cure.
Questions about how a specific output was produced: [email protected].
Related: AI policy · Privacy · Accessibility