Validate the job before the feature list.
Choose one customer type and one workflow. Ask people how they do the work today, where it breaks down, who pays for it, and what makes a change difficult. Listen for actual behavior and constraints rather than asking whether they like your idea.
Write down the current baseline and your proposed improvement. A product can be technically impressive and still fail to change an important customer outcome.
Define good enough before evaluating.
Select examples that resemble ordinary use, including difficult and incomplete inputs. Define what a useful result looks like and which failures are unacceptable. Compare your approach with a realistic alternative.
- What must be correct, and what can be approximate?
- When should the system refuse, ask for help, or defer?
- Who reviews the result before it affects a customer?
- How will you detect changes when a model or prompt changes?
High-consequence uses require appropriate domain expertise, controls, and review. A benchmark result alone is not proof of fitness for every deployment.
Map the data and the dependencies.
Identify what enters the system, which providers process it, who can access it, and what is retained. Confirm the rights and permissions needed for intended uses. Document dependencies on a model vendor, a data source, or a customer integration.
Ask what happens if data access is delayed, costs change, or a provider becomes unavailable. Early awareness can change the architecture and the business plan.
Calculate the cost of a successful outcome.
Count more than API calls. Include retries, longer inputs, evaluation, support, infrastructure, integration, and human correction. Study what happens to costs under real usage patterns.
Then connect that cost to customer value and pricing. If a useful outcome requires extensive manual work, say so and test whether customers will pay enough for the complete service.
Make the pilot a decision tool.
Agree with the customer on the use case, the people involved, the time boundary, and the success criteria. Include both product performance and adoption. Decide in advance what happens if the pilot succeeds or fails.
Customer and buyer / Workflow / Baseline / Proposed outcome / Evaluation method / Human review / Data boundaries / Commercial next step.
Review your next questions.
YOUR WORKING CHECKLIST
Make the unknowns visible.
Check what you have addressed. Unchecked items are useful next questions, not reasons to stop.
0 of 6 areas reviewed
For your own planning. This is not an investment score or an eligibility assessment. Selections are not submitted or saved.