Write the decision before the solution
Describe who has the problem, how work happens today and what needs to change. A founder may need to validate a product interaction; an operations team may need to reduce manual document handling. Those briefs can require different teams even if both mention AI.
Bring a small set of representative inputs and examples of acceptable outputs. State the decisions you have already made and the questions the provider should resolve. If the technology choice is still open, ask the agency to compare a model-based approach with a simpler existing workflow.
- Name the person accountable for the result.
- Separate necessary capabilities from interesting possibilities.
- State data-access constraints and the consequences of mistakes.
Read past the client logo
Open the project source and identify the agency’s role. Research, interface design, model development, system integration and creative production are distinct contributions. A provider that designed an AI product may be relevant for UX while a different team owns the models.
Check the stage and timing. A proof of concept can demonstrate exploration without demonstrating release operations. An older prediction or conversational-system case should prompt a discussion about current technical experience. Where a client is anonymous, ask what detail or a private reference the provider can share appropriately.
Give candidates the same questions
Send the same brief to a small number of plausible teams. Ask each to explain its proposed scope, assumptions, delivery roles, evaluation plan and ongoing responsibilities. Compare what is included before comparing headline fees. A smaller first engagement may be useful if a material feasibility question remains.
Treat fit as a discussion supported by evidence. Ask who would do the work and how that team contributed to the example you read. Clarify the contracting entity when brands share a parent group.
- What will we receive at each decision point?
- What could stop the project or change the estimate?
- How will we know whether the result is ready for the intended users?
- Who supports the system or workflow after handover?
Agree ownership before kickoff
Assign responsibility for source code, design files, accounts, data, model-provider arrangements and the ability to change vendors. For creative work, ask for a clear agreement on asset permissions and permitted use. For systems, ask how credentials, monitoring and incidents will be handed over.
This is Point of AI’s editorial procurement framework. NIST’s voluntary AI Risk Management Framework is a useful primary reference for incorporating risk considerations into design, use and evaluation. It is not a certification conferred by a directory listing.
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 ↗
Voluntary framework for AI risk considerations through design, development, use and evaluation. Source reviewed 8 November 2026.