Disclosure is the starting point, not the control

The UK Government's Consultancy Playbook recognises the growing use of AI in consultancy and allows contracting authorities to ask bidders about intended AI use in the procurement process and in producing advice or outputs. PPN 017 provides optional questions covering AI use in bid preparation and service delivery.

A yes-or-no declaration does not show whether the output is dependable. Buyers should connect disclosure to an evidence schedule, acceptance criteria and named accountability.

Separate tender AI from delivery AI

Ask two questions. First, was AI used to draft, summarise, score or substantiate the bid? Second, will AI be used to research, analyse, recommend, write or quality-assure the contracted deliverables?

The risks differ. Bid use can affect the accuracy and originality of claims. Delivery use can affect client confidentiality, factual reliability, professional judgement, intellectual property and the reproducibility of advice. Record each use separately.

The 12-item buyer evidence schedule

  1. Purpose and boundaries. The tasks AI may support and the decisions it must not make.
  2. Tool and model register. Provider, product, model or version, configuration and material dependencies.
  3. Source and provenance register. The evidence used, where it came from and how citations were checked.
  4. Data-handling record. Client data entered, retention, training use, location and access controls.
  5. Named human approver. A competent person accountable for the final deliverable.
  6. Review method. Defined checks for facts, calculations, reasoning, bias, omissions and instructions.
  7. Validation evidence. Test cases, comparison or sampling proportionate to the consequence of error.
  8. Limitations statement. Known uncertainty, assumptions, exclusions and areas requiring specialist judgement.
  9. Confidentiality and IP controls. Rights to inputs and outputs, protected material and leakage prevention.
  10. Subcontractor and subprocessor map. External parties, locations, responsibilities and change controls.
  11. Incident and change record. Errors, corrective action, tool changes and notification thresholds.
  12. Retention and handback. Evidence retained for audit, deletion duties and usable client records at exit.

Make human review observable

“Human in the loop” is too vague to be an acceptance criterion. Name the reviewer, their competence, the checks performed, the source material inspected and the outcome. For higher-consequence advice, require independent review rather than approval by the person who prompted the system.

The consultant remains accountable for the contracted output. AI assistance should not blur who formed the opinion, checked the evidence or accepted the limitations.

Use acceptance gates for material advice

Set gates before work starts: approved tools and data conditions, an agreed evidence format, draft review, factual validation, limitation sign-off and final acceptance. A deliverable should pause where material claims cannot be traced, confidential information was handled outside approved controls or the named reviewer cannot explain the basis of the recommendation.

This aligns with government procurement guidance calling for lifecycle testing, ongoing evaluation, accountable governance and go/no-go decisions at meaningful stages.

Questions to put into the procurement

  • Where will AI be used in the bid and the service, and for what purpose?
  • Which tools, models and third parties will process our information?
  • How will sources, calculations and material claims be verified?
  • Who is professionally accountable for each deliverable?
  • What limitations, uncertainty and excluded uses will be disclosed?
  • What evidence will accompany delivery, and what triggers notification or re-approval?

Translate the accepted answer into the contract, delivery plan and acceptance record. A tender statement that disappears after award is not an operational control.

Frequently asked questions

Should consultants disclose AI use?

Buyers should ask suppliers to disclose material AI use in preparing tenders and delivering advice, including the purpose, tools, data handling, controls and accountable human reviewer.

Is human review enough for AI-generated advice?

No. Human review needs a named competent reviewer, defined checks, source verification, documented limitations and clear accountability for the final deliverable.

What evidence should accompany AI-assisted advice?

The evidence should cover provenance, tools and versions, data handling, human review, validation, limitations, confidentiality, subcontractors, changes, incidents, retention and acceptance.

Turn the declaration into evidence

Pair this schedule with the supplier AI-use verification method and the procurement go/no-go pack. For help applying the controls to a live procurement, talk to AI Act Ready.

Last reviewed 16 November 2026. This article provides general information, not legal advice.