Choose the next stage you need
Discovery should resolve an important uncertainty: the user need, data suitability, integration access or a technical approach. A prototype explores behavior. A proof of concept tests whether a selected approach is plausible. Neither term by itself specifies the work required for a maintained release.
Write down the decision the first stage will support. If it is “can this extract the fields we need?”, use representative documents and agree how results will be reviewed. If it is “can our customers use this safely?”, include interaction research and failure cases. Avoid paying for a polished demo when the unresolved question is elsewhere.
Make assumptions visible in the estimate
Ask for the same scope breakdown from each provider. Identify whose time, accounts and data are required, what has been assumed about access and what will trigger a change request. A fixed fee can still depend on a tightly bounded input set; an hourly engagement still needs a clear decision process.
Agree what counts as acceptance, how defects are handled and which work is explicitly outside the estimate. Ask how the team will report an infeasible assumption before spending the remaining budget.
- Discovery, design and evaluation activities.
- Data preparation and integration access.
- Implementation, release checks and handover.
- Documentation, staff training and maintenance.
Include the operating work
An implementation estimate and a running-cost estimate answer different questions. Ask providers to identify expected usage, model and infrastructure assumptions, third-party subscriptions, monitoring and support. Request a way to revise the estimate when actual usage becomes available.
For a recurring workflow, include the person who reviews exceptions and maintains connectors. For a product, clarify who responds to failures and evaluates changes to models or prompts. For a creative engagement, clarify revision rounds and the cost of producing additional formats. These are scope questions, not universal price benchmarks.
Budget for the decision to continue
Make evaluation an explicit deliverable. Ask to see results for the full task, including difficult inputs, and separate a demonstration from a repeatable check. Decide which failures stop release, which can return to a person and who can approve expansion.
NIST’s voluntary AI framework is a reference for considering risks across the lifecycle. The practical checklist here is editorial guidance: use it to make agency proposals comparable, then confirm commercial and contractual terms with the provider.
Sources and editorial context
Procurement checklists are Point of AI’s editorial guidance. The following primary sources support the specific principles identified here; they do not endorse this directory or its listed providers.
- NIST AI Risk Management Framework ↗
Lifecycle consideration of AI trustworthiness and evaluation; not a price guide. Source reviewed 8 November 2026.