Define the user and the task
Write a one-sentence statement: this product helps a specific user complete a specific task. Describe how the task works today, where time is lost, and what a better outcome looks like. If the scope includes several unrelated workflows, choose one for the first release.
Collect representative examples
Prepare sample inputs and expected outputs. Include ordinary cases, ambiguous cases and failures. Remove sensitive data unless confidentiality and handling terms are agreed. For a knowledge product, include the documents needed to answer each example question.
Decide whether AI is needed
Compare an AI approach with a simpler rule, search feature or standard workflow. AI can help with variable language and unstructured information, but deterministic operations are often easier to verify. A hybrid approach can use AI only for the part that benefits from it.
Choose the product pattern
Use retrieval when the answer depends on proprietary knowledge. Use an agent when the task needs decisions and tool calls across systems. Use a standard model integration for a focused generation or classification step. These patterns can be combined, but each adds evaluation and operating requirements.
Define acceptance criteria
Specify what a successful task looks like and how it will be measured. Consider accuracy, task completion, latency and cost per successful task. Include behavior when the model cannot answer, the source is missing, or an external service is unavailable.
Map permissions and data handling
List who can access each type of record, what data may leave your infrastructure and which actions require approval. Tenant isolation, source permissions and log redaction need implementation beyond the prompt. Agree provider and contractual requirements before using production data.
Separate proof from production
A feasibility proof can reduce uncertainty about a difficult AI task. A production release also needs authentication, persistence, observability, failure recovery and a usable interface. Decide which questions a proof must answer and which deliverables belong to the subsequent build.
Prepare a brief for your development partner
Summarize the user, workflow, sample data, essential integrations, constraints and acceptance criteria. Share an existing prototype if available. Ask for a proposal that identifies assumptions and dependencies as well as features, then review how the product will be maintained after launch.