Short answer: The minimum useful output is a ranked use case, evidence about data and workflow, explicit risks, a measurable pilot scope, cost drivers, and a go/no-go recommendation.
Discovery should leave a decision record
The purpose of discovery is to reduce uncertainty before implementation money is committed. It should connect a business problem to evidence, constraints, and a bounded pilot. A slide deck listing possible chatbots is not enough.
Inputs the client should expect to provide
- the current workflow and its owner;
- representative examples, including failures and exceptions;
- current volume, time, cost, quality, and delay;
- systems, accounts, access constraints, and integrations;
- data classifications, contracts, policies, and relevant regulatory requirements;
- the people who review work today and the people affected by a change.
Required discovery outputs
A ranked use-case register
Each candidate should state the user, task, expected value, frequency, available evidence, risks, dependencies, and reason for its priority. Ranking should compare AI with simpler process, search, template, or deterministic automation options.
A workflow and data assessment
The assessment should show where information originates, who may use it, its quality, what is missing, and whether it can lawfully and contractually be processed by proposed providers. Unknowns must remain visible rather than becoming assumptions.
A solution outline
Record the proposed model or service category, retrieval or integration needs, access boundary, human-review point, logging, evaluation approach, security controls, and shutdown path. This is an outline, not premature production architecture.
A measurable pilot plan
The plan needs a fixed scope, evaluation set, baseline, acceptance thresholds, review owner, schedule, cost drivers, exclusions, and go/pause/stop decision. It should say what the pilot will not do.
Commercial and ownership clarity
The engagement should state who owns the discovery materials, prompts, evaluation sets, integration code, generated outputs, and customer data. It should identify third-party subscriptions, usage charges, likely ongoing work, and which estimates remain uncertain.
Vendor red flags
- A solution is chosen before the workflow and evidence are reviewed.
- Benefits are described without a measurable baseline.
- The proposal assumes broad production or document access for convenience.
- Human review is described as a temporary inconvenience rather than a control.
- Risks are reduced to a generic disclaimer.
- The client receives no usable evaluation set, decision record, or exit path.
- A successful demonstration is treated as production approval.
The minimum useful handover
At the end, the client should be able to explain what problem is being tested, why AI is appropriate, what information it may use, how results will be judged, what it will cost to learn, who is accountable, and what decision follows. If those answers are absent, discovery is not complete.