Andrew Crossley

    How do you build an AI MVP?

    SHORT ANSWER

    Build an AI MVP by choosing one task where being right 80% of the time is still useful, wrapping a hosted model rather than training your own, and designing the interface around correction — users must be able to see, edit and accept output. Set an evaluation set before launch, measure task success and human-override rate, and only consider fine-tuning once the workflow itself is proven.

    Answered by Andrew Crossley, Fractional Chief Product Officer · Updated 2026-08-01

    Why it matters

    AI products fail at the workflow layer, not the model layer. The model is a commodity; the useful part is where it sits in someone's job.

    Without evaluation, teams ship confidently wrong output and discover the problem through churn instead of through data.

    How it works in practice

    1. 1

      Pick a tolerant task

      Drafting, summarising, classifying, triaging. Avoid tasks where a single wrong answer is unacceptable until you have evaluation and guardrails.

    2. 2

      Start with a hosted model

      Use a gateway to a frontier model. Model choice is a tuning decision later; the workflow is the risk now.

    3. 3

      Build the evaluation set first

      Fifty to a hundred real examples with expected outputs. Run it on every prompt or model change. This is the single highest-leverage practice in AI product work.

    4. 4

      Design for correction

      Show the reasoning or source, allow editing, and capture every correction. Corrections are your training data and your quality metric.

    5. 5

      Instrument the honest metrics

      Task success rate, human-override rate, time saved per task, and cost per successful task — not tokens processed.

    6. 6

      Only then optimise

      Cheaper models, caching, retrieval or fine-tuning, chosen against the evaluation set rather than intuition.

    Common mistakes

    • Building a chat box because it is easy, when the user needed the answer inside their existing workflow.
    • Fine-tuning before there is any evidence the task is valuable.
    • No evaluation set, so quality changes are invisible until users complain.
    • Ignoring inference cost per task until the unit economics are already negative.

    FROM EXPERIENCE

    Where AI features actually earn their cost

    The AI features that hold up in production are the boring ones: triaging inbound requests, drafting a response someone edits, extracting structured data from messy input. Each replaces a repeated manual step with a reviewable draft.

    The ones that fail are open-ended assistants bolted onto a product, where the user has to invent the use case themselves.

    Frequently asked

    Do I need my own model?

    Almost never at MVP stage. Hosted frontier models plus good prompting and retrieval cover the vast majority of use cases.

    How do I stop hallucinations damaging trust?

    Ground answers in your own data, show sources, and design the UI so output is reviewed rather than auto-applied.

    What does an AI MVP cost to run?

    Model spend usually lands between £0.01 and £0.30 per task depending on model and context size. Measure cost per successful task, not per call.

    IN SHORT

    • One tolerant task, hosted model, workflow-first design.
    • Evaluation set before launch; measure task success and override rate.
    • Optimise cost and model choice only after the workflow is proven.

    Build your AI MVP in 5 days

    The self-paced course for founders shipping their first AI product.

    Build your AI MVP in 5 days

    THE FRAMEWORK

    The Crossley Method: idea to first revenue in seven stages

    See the full method
    1. STAGE 1DiscoverWeek 1
    2. STAGE 2ValidateWeek 2
    3. STAGE 3PrototypeWeek 3
    4. STAGE 4Build MVPWeeks 3-4
    5. STAGE 5LaunchWeek 5
    6. STAGE 6First RevenueWeek 6
    7. STAGE 7ScaleOngoing