AI consulting and business adoption
How to write an AI pilot brief before development starts
An approved AI idea is not yet a build brief. If the team starts configuring tools before it agrees who will use the result, which records are permitted and how failure will be detected, the pilot can produce an attractive demo without answering the business question. A short pilot brief gives the process owner, users and technical team one testable scope before development begins.
By SVS Digital · Published
Start here
Write the brief around one workflow and one decision. Name the current problem, intended users, exact step AI may support, inputs and prohibited data, expected output, human reviewer, integrations, test set, baseline, acceptance thresholds, failure response, owner and end-of-pilot decision. Keep production rollout, extra departments and untested actions out of scope. Approve the brief with the process owner, a representative user, the technical owner and any required security, data-protection, legal or sector specialist before connecting live data or allowing real side effects.
1. Start with the decision the pilot must support
State the business problem in current-work terms: for example, service enquiries wait in a shared mailbox and staff re-enter the same fields into the CRM. Then write the decision due at the end of the pilot: expand to a controlled production trial, revise the approach, collect more evidence or stop. “Explore AI” and “improve efficiency” are themes, not decisions.
Name one process owner and the people affected by the result. NIST's AI RMF says context, intended purpose, users, affected groups, benefits, costs and foreseeable impacts should be mapped before risk is measured and managed. It also treats a no-go decision as a valid result. The pilot brief should therefore expose the assumptions that the test is meant to confirm or reject.
- One current workflow and one observable problem.
- A named process owner with authority to accept or stop the pilot.
- Users and affected people identified, not only the project sponsor.
- A written end decision: expand, revise, research or stop.
2. Draw the boundary around one task and its handoffs
Describe the trigger, the step AI will support, the expected output, who checks it and where the work goes next. Separate assistance from automation. An assistant may draft a reply that an employee approves; an automated system may classify a request and update a queue. Those designs need different permissions, review capacity and failure handling.
The UK AI Knowledge Hub recommends breaking work into workflows, tasks and steps, and notes that AI is inconsistent across tasks. It advises testing with known examples rather than assuming that success on one task transfers to another. List explicit exclusions such as pricing decisions, employment decisions, legal conclusions, refunds or unsupervised customer messages unless those actions have been separately assessed and approved.
- Trigger, input, AI-supported step, output and next recipient.
- In-scope user group, channel, language and case type.
- Human decision points and escalation route.
- Out-of-scope actions written in plain language.
3. Specify data, systems and side effects before choosing a tool
List each data source, who owns it, which fields are necessary and whether representative test examples may be used. Record sensitive or personal data that must be removed, masked or kept out of the pilot. Do the same for outputs: where they are stored, how long they are retained, who can view them and how a correction is made.
Map integrations separately from model behaviour. Reading a knowledge base, looking up a customer and writing to a CRM are three different permissions. For the first test, prefer a sandbox or recognisable test records. If a live write is necessary, define authorization, duplicate protection, logging, rollback and a person who verifies the final business record. A spoken or displayed confirmation is not proof that an external action succeeded.
- Permitted sources and fields, with an owner for each.
- Representative test data without unnecessary personal information.
- Read and write permissions listed per connected system.
- Retention, access, correction, logging and rollback defined.
4. Define acceptance checks before anyone sees the demo
Create a test set from ordinary cases, frequent exceptions and rare cases with serious consequences. For every case, record the expected output or decision, permitted variation and reviewer. Measure more than answer quality: include completion of the workflow, review time, corrections, failed integrations, user adoption and operating cost where they affect the decision.
The UK AI Knowledge Hub recommends planning evaluation early, establishing a baseline and considering quality, efficiency, user satisfaction, usage and cost. NIST likewise calls for documented test sets, metrics and acceptable limits that match the deployment context. Avoid a single score that hides a critical failure. A pilot may pass routine drafting accuracy and still fail if it sends an unauthorized message or omits an important exception.
- Baseline from the current process, with the source and period recorded.
- Known examples covering routine work, exceptions and failure paths.
- Separate quality, workflow, review, adoption and cost measures.
- Critical failures that cause an immediate pause.
5. Assign human review, training and specialist checks
Name the reviewer, what evidence they see, how much time review requires and what they do when uncertain. “Human in the loop” is not a control if the person cannot detect an error, override the output or reach the original source. Include user training, a route for reporting problems and the owner who may pause the pilot.
Risk and compliance depend on the use, people and sector. The European Commission's AI Act overview distinguishes prohibited, high-risk, transparency and minimal-risk uses, and identifies employment, access to essential services and some safety-related applications among areas requiring particular attention. The brief should flag the intended use for qualified legal, data-protection, security or sector review rather than declaring compliance from a generic checklist.
- Reviewer, authority, source access and time allocation.
- User instruction for correct use, limitations and escalation.
- Named owner for incidents and the pause decision.
- Required legal, privacy, security or sector review flagged early.
6. Finish with deliverables, ownership and a go-or-stop review
List what the pilot will actually deliver: configured prototype, approved test set, results log, user feedback, known limitations, integration findings and a recommendation. Add responsibilities for data preparation, access, testing, user availability, technical changes and final approval. Tool subscriptions, implementation work, training and ongoing operation should not be hidden inside an undefined “pilot”.
At the review, compare results with the pre-agreed baseline and critical limits. Decide whether to stop, adjust and retest, run a narrower controlled trial or prepare a separate production plan. Production needs its own security, monitoring, support, ownership and change process. The pilot brief is not a legal assessment, procurement contract or guarantee of savings; it is a disciplined way to learn whether one proposed use deserves the next investment.
- Deliverables and exclusions visible to every owner.
- Responsibilities for access, testing, feedback and approval.
- Known costs and dependencies separated from hoped-for benefits.
- A scheduled review that can approve, narrow, repeat or stop.

