AI consulting and business adoption
What should an AI consultation deliver before a pilot is approved?
An AI consultation should leave leadership with a decision they can examine, not a presentation about possible tools. Before approving a pilot, the company needs a compact discovery pack that shows what was investigated, what remains unknown and why the recommended next step deserves time, data and budget. The pack should be readable by the business owner, users, technical team and specialists who may need to challenge it.
By SVS Digital · Published
Start here
Before approving a pilot, require six outputs: a scoped problem and current-workflow record with owners; a comparison with process change, ordinary automation and no-go options; data and system readiness findings with unresolved questions; issues affecting users and other people, plus risk and oversight questions that need escalation; a concise record of the assumptions behind each option, including costs and operational demands; and a recommendation to pilot, research, redesign or stop. Put the recommendation in the discovery decision pack. If the decision is to pilot, commission a separate next-step brief rather than treating discovery notes as a build specification.
1. Confirm the problem, workflow and decision owners
The first deliverable is a record of the problem in current-work terms. Show the trigger, main steps, handoffs, exceptions, present output, users, affected people and available baseline. Name the process owner, operational users who understand the exceptions and the executive who owns the approval decision. A proposed feature is not a problem statement.
UK Government implementation guidance says discovery should establish user needs, the problem, whether AI is suitable, the wider process, data location and responsibilities. It was written for the public sector, but these are useful management questions for private firms, not legal requirements.
- Problem, trigger, handoffs and current output.
- Baseline labelled as measured, sampled or estimated.
- Process, user and approval owners.
- Workflow unknowns assigned for investigation.
2. Compare alternatives before recommending an AI pilot
Present a short options table built around the same problem: process-only change, non-AI automation using rules or integrations, AI assistance and no-go. Compare what each option changes, its evidence and dependencies, and why it was retained or rejected. Product sourcing comes later, after the company justifies a route.
NIST's Map playbook asks organisations to examine non-AI alternatives and document purpose, human roles, context and possible impacts. UK procurement guidance recommends defining the problem and staying open to alternatives rather than writing a solution-first specification. Stopping is valid when the evidence does not show enough value or adequate control.
- Process change without new technology.
- Rules or integration automation.
- Bounded AI assistance with a human role.
- No-go option and rejection reasons.
3. Report data and system readiness, including unknowns
Distinguish evidence from assumption. For each data source, record location, owner, permitted use, sensitivity, quality and representative examples. For each system touchpoint, record access method, identifiers, permissions and whether the proposed read or write was verified. Untested integrations, inaccessible samples and unclear rights are open dependencies.
The Australian Government checklist asks about data quality and governance, integration points, current workflows, stakeholders and a decision register. UK guidance also calls for checking data location, condition, access, permission and representativeness. Private firms can use these questions to expose uncertainty; checklist completion does not prove readiness or compliance.
- Data owner, permission, quality and sensitivity.
- Representative examples available for testing.
- Access and integration claims checked.
- Unknowns assigned with evidence requests.
4. Escalate people, risk and oversight questions
Identify who uses the output, who may be affected by an error and who can detect, correct or stop it. Describe failure consequences, reviewer evidence, override authority and reporting routes. Flag workforce changes, accessibility, misuse, sensitive data and consequential decisions for the relevant security, privacy, legal, human-resources or sector specialist.
NIST Map calls for documented users, impacts, assumptions, limits, human roles and monitoring responsibilities. The Australian checklist asks about stakeholder engagement, workforce implications, governance and impact assessment. Discovery should state which reviews are required and why, not declare a proposed system safe, ethical or compliant.
- Users, affected people and failure consequences.
- Reviewer evidence, authority and escalation.
- Monitoring, incident and pause owners.
- Specialist questions without pre-judged conclusions.
5. Show the assumptions behind options, cost and operation
Use a short assumptions table for case volume, users, data preparation, integration work, licence or usage basis, employee review time, training, support, monitoring and change ownership. Give ranges or mark figures unknown. Separate one-time work from recurring operation, and do not turn hoped-for time savings into a guaranteed return.
The Australian checklist asks about expected costs, benefits, success criteria and business priority before a proof of concept. UK procurement guidance calls for multidisciplinary input, data discovery, lifecycle support and whole-life cost. These prompts reveal decision drivers. Detailed quotations and sourcing comparisons come after the route and boundary are clear.
- Business, volume and user assumptions.
- Initial and recurring costs separated.
- Review, training, support and monitoring included.
- Confidence and missing evidence shown.
6. Make one decision, then issue a separate next-step brief
Finish with one recommendation: pilot a bounded hypothesis, research a missing dependency, redesign the process or stop. Tie it to evidence, rejected alternatives, unresolved questions and conditions that could change the decision. Record the approval and any dissent. Research should name the evidence needed; stop should explain why the route is unsuitable.
After pilot approval, create a separate brief covering scope, users, test data, human review, acceptance checks, stop conditions, owners and the next decision. Discovery does not replace that brief, supplier due diligence, security review, impact assessment or legal advice, and cannot establish performance before representative testing. Its purpose is to make the next investment traceable to evidence.
- Decision is pilot, research, redesign or stop.
- Evidence, assumptions and conditions attached.
- Decision owner and register entry confirmed.
- Separate brief follows pilot approval.

