AI consulting and business adoption
Who should take part in an AI consultation?
The right AI consultation team is not the whole leadership group or everyone who might touch the project. It is the smallest group able to explain the decision, real work, difficult exceptions, systems and effects on users. Use a compact core team, with privacy, security, legal or HR specialists called in only when the use case raises questions in their area.
By SVS Digital · Published
Start here
Include the decision sponsor, process owner, one or more frontline operators, the data or system owner, an affected-user or customer-support perspective, and an AI or technical specialist. One person may cover two roles, but every perspective still needs evidence. Keep privacy, security, legal, HR and sector specialists on call. Postpone decisions when an absent role owns a material fact, risk or approval.
1. Form a core team around evidence, not job titles
As a practical meeting heuristic, start with four to seven active participants; the cited guidance does not prescribe a team size. The sponsor owns the decision; the process owner explains the required result; operators show what actually happens; the data or system owner checks access; the user perspective exposes missing needs; and the AI specialist challenges technical assumptions. Invite specialists for defined questions rather than every meeting.
GOV.UK guidance recommends decision-makers, delivery expertise, back-end system knowledge and access to specialists. It also says an AI-enabled public service should have someone available who understands the technology and its impact. These are useful management practices, not private-company legal requirements.
- Decision sponsor or owner: the person able to set the objective, constraints and next decision.
- Process owner: the person accountable for the current result across handoffs.
- One or more frontline operators: include someone who handles exceptions, not only the standard path.
- Data or system owner: the person who can verify sources, permissions, quality and integration realities.
- Affected-user or customer-support perspective: someone with direct evidence of needs and difficulties.
- AI or technical specialist: someone able to explain feasibility, limitations and non-AI alternatives.
2. Ask each role to bring evidence only they can supply
Give each participant an evidence request before the meeting. A sponsor should not describe frontline work from memory, and a technical specialist should not infer user needs from system fields. If one person covers two roles, separate direct knowledge from assumptions.
NIST's Map playbook calls for documented context, users, operators, impacts, human roles, limitations and interdisciplinary participation. The AI Risk Management Framework is voluntary guidance, not a certification or legal determination. Here, it helps expose which perspective supports each claim.
- Sponsor: decision to be made, business constraint, acceptable boundary and who can approve the next step.
- Process owner: current outcome, evidence of demand or case volume, handoffs and known failure points.
- Operators: two ordinary cases, three awkward exceptions, workarounds, review effort and escalation triggers.
- Data or system owner: source inventory, access route, sensitivity, quality concerns, identifiers and verified system limits.
- User or support perspective: observed needs, complaints, support logs, accessibility barriers and consequences of error.
- AI specialist: capability assumptions, likely failure modes, test needs, human oversight and process-only or rules-based options.
3. Prevent an executive-only version of the process
Leadership can state why the work matters, but it rarely sees every correction, side spreadsheet or judgement call. Ask someone who performs the work to walk through a recent redacted case. Compare the official procedure with support logs, workflow records or observation. Disagreement may reveal two process paths.
GOV.UK discovery guidance recommends learning how users complete a task, examining support and back-office evidence, and researching people who provide the service as well as those who receive it. It is public-service guidance, not a private-company legal requirement. The transferable practice is direct observation instead of executive assumption.
- Ask the operator to narrate the last real case rather than the official procedure.
- Ask customer support what users misunderstand, abandon or need help with.
- Mark claims as observed, measured, reported or assumed.
- Record conflicting accounts as questions to investigate, not facts to average away.
4. Run a focused discovery conversation lasting 60 to 90 minutes
Send evidence requests in advance and appoint one facilitator. Stay with the current problem and decision evidence. Product demonstrations and vendor comparisons can wait. Assign each unresolved question to the relevant specialist, and do not make a recommendation if the answer could change it.
- 0–10 minutes: sponsor states the decision, boundary and reason for reviewing the work.
- 10–30 minutes: process owner and operators reconstruct one normal case and key exceptions.
- 30–45 minutes: user or support evidence identifies needs, barriers and consequences of error.
- 45–60 minutes: data or system owner checks sources, access, sensitivity and system touchpoints.
- 60–75 minutes: AI specialist tests assumptions and compares AI with process change or ordinary automation.
- 75–90 minutes: list conflicts, missing evidence, specialist referrals, owners and the next decision. At 60 minutes, capture unresolved items rather than rushing.
5. Stop when a missing role or conflict could change the decision
Do not settle a disputed fact by seniority. Record each account, the evidence needed and its owner. Pause recommendation or approval when a gap concerns authority, operating practice, affected people, data rights, security, employment impact or technical feasibility.
Bring privacy or security expertise when the use case involves sensitive data, access controls, suppliers or failures with serious consequences. Bring legal or sector expertise for regulated or rights-affecting uses, and HR when work allocation, monitoring, hiring or employee decisions may change. The consultation should identify which specialist review is needed; it does not provide that specialist’s conclusion.
- Stop if no one can authorise the problem boundary or next decision.
- Stop if frontline practice or exception handling is represented only by management assumptions.
- Stop if affected-user consequences are material but no direct or support-based evidence is available.
- Stop if data ownership, permitted access or a critical system constraint cannot be verified.
- Stop if a specialist answer could make the use unacceptable or require a different design.
6. Know what participant discovery can and cannot establish
A well-chosen group can expose assumptions and assign evidence requests. It cannot prove compliance, security, model performance, user acceptance or business value. Those questions need appropriate review and representative testing.
This guide covers who participates and what evidence each person contributes. It does not cover consultation preparation, the final discovery decision pack or the pilot brief. Keep the team small, document whose evidence supports each conclusion and expand it only for a genuine need.

