← All field notes

AI MVP Roadmap: Prove the Workflow Before Building the Platform

A minimum viable product should test whether AI helps with one useful task. Keep the first version small enough to review its answers, costs and mistakes.

AIMVPProduct Strategy

An AI product idea can begin with an ambitious vision: an assistant that handles every request or understands an entire document collection. Building all of that at once makes it difficult to discover which assumptions are sound.

A minimum viable product, or MVP, is a small working version that tests the central idea. For AI, the first question is whether the model produces useful work consistently enough to justify the review, development and running costs.

Choose a task someone can assess

Name the person using the result, the information available and the decision or piece of work being improved. Keep the first workflow small enough for a person to review each result. Describe where the task ends so the pilot does not quietly become a much larger product.

Also compare the idea with a simpler rule, search tool or existing workflow. AI should earn its place through the work it helps people complete.

Agree on good answers before testing

An evaluation set is a collection of inputs with a clear description of what an acceptable result looks like. Include incomplete requests, conflicting information and questions the system cannot answer. Use suitable material that can be tested without exposing private information.

If the model searches a document collection before answering, that process is called retrieval. Check both parts: did it find the relevant source, and did its answer accurately reflect that source? Changing the model will not necessarily fix missing or outdated documents.

Build only what the pilot needs

A simple internal screen or assisted workflow may be enough to learn whether the core task works. Account management, billing and a polished dashboard can wait unless they are essential to the question being tested.

Make the source behind each answer easy to inspect. Keep the first outputs as drafts or recommendations where mistakes would otherwise affect customers. Decide how the workflow handles uncertainty: ask for missing information, return no answer or pass the task to a person.

Count review time as part of the cost

Record useful results, corrections, completion time and running cost. Include the time people spend checking answers and dealing with failures. Read the serious errors individually; an average score can hide an unacceptable pattern.

Limit the cost and duration of a run. Clear stopping points make the pilot manageable and help distinguish a useful capability from one that needs frequent intervention.

Finish with a decision

Ask for a working demonstration, the test collection, a breakdown of errors and costs, and a recommendation about the next step. The result may support expansion, a narrower scope, an internal tool or a decision to stop. Learning that before committing to a full platform is a useful outcome.