Responsible AI, as a control rather than a principle.

Transparency

The person is told, and the record says so.

Each AI reply segment produced in a governed call carries the required disclosure, in the language and at the position the applicable jurisdiction requires. The disclosure runs inside the conversation, not in a policy document filed somewhere. In controlled testing.

The interaction leaves a compliance receipt carrying the input hash, the output hash, the link to the previous receipt and its Merkle inclusion. Anyone can recompute the first three without a key; the fourth, inclusion in the Merkle root, is checked against the published root with the inclusion proof from the call's Article 50 report. Together with them sits an authentication code issued under a key held by Audact. Later alteration or reordering is detectable by recomputation against a retained commitment.

Human oversight

Someone is named, and can stop it.

A deployment requires a named human operator with the authority and the tooling to intervene, pause or end any conversation. Escalation paths to a person are configured per use case before the first call, not designed after the first complaint.

Robustness

The check runs before the words, not after the transcript.

Calls that would breach a disclosure or jurisdictional rule are stopped before dispatch. Brand-governance policy validates pricing, claims and commercial commitments against what the customer approved. Incident detection runs continuously against the evidence chain rather than waiting for someone to look.

Fairness

We govern dialogue policy; we do not write it.

Audact does not author customer-facing dialogue. Customers configure their own deployments and take structured impact assessments through the platform, with prompts for risks to protected characteristics under the AI Act and national equality law.

That division matters: the party who decides what an agent says is the party who has to answer for it, and the tooling should make that obvious rather than blur it.

Frameworks

What we work against.

The EU AI Act is the binding text and takes precedence in any conflict. Alongside it we use the NIST AI Risk Management Framework and the OECD AI Principles as reference frameworks for our own practice.

Audact AI Ltd signed the European Commission's AI Pact on 10 June 2026 and appears on the Commission's signatory list. The Commission is explicit that these pledges are not legally binding: it is a voluntary public commitment, and it is not a certification, an audit or an approval.