AI Proof of Concept in 6 Weeks
A fixed-scope engagement that answers one question: does AI work on your data, for your process, at a quality worth paying for. A working prototype, an evaluation report and a go or no-go decision.
One use case, six weeks, a defensible answer
Most AI budgets die between the slide and the system, in pilots with no success metric and no end date. Our proof of concept is the opposite shape: one use case, your real data, six weeks, and success criteria written down before we start. The scope and the price are fixed in the proposal, so the decision you buy is the decision you get.
The outcome is honest by design. If the prototype clears the agreed bar, you have a validated case and an architecture to take to production. If it does not, you have spent six weeks and a bounded budget to avoid a six-figure mistake, with a report that says exactly why.
- Fixed scope and fixed price, agreed in the proposal before we start.
- Success criteria and a stop rule written down in week one.
- Your data, your security perimeter, EU jurisdiction throughout.
The six weeks, week by week
- Week 1: scope lock: the use case, the data access, the success criteria and the stop rule, all in writing. Nothing is built until the bar is agreed.
- Weeks 2-3: first working version on real data: the pipeline, the model approach, the integration stubs. Ugly is fine; real is mandatory.
- Week 4: evaluation against the criteria: a test set your team helps assemble, measured accuracy, failure cases catalogued.
- Week 5: iteration where the evaluation says it pays, and projected production costs: model usage, infrastructure, review workload.
- Week 6: the decision package: live demo, evaluation report, production architecture and roadmap, or a documented no-go.
Success criteria set at the start
The week-one scope document names the metric, the bar and the test method: for a document pipeline, field-level accuracy on a held-out month of your documents; for an assistant, answer quality over a question set your team writes; for a prediction, lift against your current baseline. It also names the stop rule, because a proof of concept that cannot fail is marketing, not engineering.
This is the discipline we argued for in our writing on why AI pilots fail: no baseline, no bar, no end date are the three ways a pilot becomes a permanent science project. The PoC format makes all three impossible.
Proof: from PoC to production
The Mayster parts assistant in our case studies started exactly this way: a bounded retrieval prototype on the real catalogue, evaluated on real dealer questions, with the fine-tune-or-retrieve decision settled by measurement rather than fashion. It cleared its bar and went to production; the architecture the PoC proved is the architecture that runs today.
What you get
-
A working prototype
Running on your real data, demonstrated live with your team driving it, not a slide of screenshots.
-
An evaluation report
Measured accuracy against the agreed criteria, failure cases documented, running costs projected for production volume.
-
A go or no-go decision
A production architecture and roadmap if the bar is cleared; a reasoned stop with lessons captured if it is not.
Frequently asked questions
What does the proof of concept cost?
The price is fixed and stated in the proposal once the use case is scoped. We do not publish rates, but the scope is bounded by design: six weeks, one use case, a defined team.
What do you need from us?
Access to the relevant data, a business owner for the use case who can spend a few hours a week, and security sign-off for the environment. Week one is designed to settle all three.
Who owns the prototype and the code?
You do. Code, prompts, evaluation sets and documentation are handed over whatever the decision.
What happens after a go decision?
The evaluation report doubles as the production plan: architecture, timeline and the team to build it, either with us or with your own engineers on the handover.
What if the data is not ready?
Week one surfaces that honestly. Sometimes the right outcome is a data-readiness sprint before the PoC; we would rather delay the start than measure a model on data that misrepresents you.
Six weeks to a decision you can defend
Tell us the use case you are weighing. We will scope the PoC, fix the price and write the success criteria with you.