Skip to content

AI prototyping and Prove Sprint

Some AI ideas hinge on one question: can the model actually do this, on your material, well enough to be useful? We build the smallest working prototype that answers it and test it on your real examples. The packaged version is the Prove Sprint. It is for uncertain AI capabilities, not a required step for every project.

About the Prove Sprint3 to 4 weeks, scope agreed per project

When you need this

  • You have a promising idea, but nobody knows if the model can do the hard part.
  • Your documents, images or data are unusual, and generic demos do not prove much.
  • You need evidence before asking for a larger budget.
  • A previous attempt stalled and you want to know whether to try again or stop.

How we do it

  1. 1. Name the uncertainty

    Agree the single question the prototype must answer, and what result would count as success, partial success or failure.

  2. 2. Build the test set

    Collect 30 to 100 representative examples from your real material, including the difficult cases.

  3. 3. Prototype

    Build the smallest working version of the capability, tried with more than one model or approach where it matters.

  4. 4. Report honestly

    Results on the test set, the failure cases, the likely cost to run, and a recommendation to proceed, revise or stop.

What you get

  • Working proof of concept
  • Representative test set and results
  • Documented failure cases
  • Feasibility conclusion and running-cost estimate
  • Recommendation to proceed, revise or stop
Example
Illustrative

A logistics company wants to read handwritten delivery notes. The prototype is tested on 80 real scanned notes. Printed fields read reliably; handwritten quantities do not. The recommendation is to proceed with printed fields now and route handwritten quantities to a person for review.

Start with a Prove Sprint

3 to 4 weeks, scope agreed per project

One uncertain AI capability built on your systems and tested against your own examples. It ends with the results, a running-cost forecast and a scoped plan for production, or a clear recommendation to stop.

About the Prove Sprint

Questions

What is a Prove Sprint?

The packaged version of this work: usually 3 to 4 weeks, with the scope agreed per project, ending in test results and a go or no-go decision.

What if the prototype fails?

Then you have learned that early, with evidence of why. A clear no is a useful result.

Often combined with