Home › Blog › Designing an AI PoC

Designing an AI proof of concept to finish in four weeks

PoCs consume six months not because the scope is large but because there is no end condition. Write the condition for finishing first and four weeks is frequently enough.

4-week design Weekly goals Stop conditions

The short answer

The purpose of a PoC is to produce the minimum evidence an adoption decision needs. Not to build a polished system, and not to handle every exception.

So before starting, write two sentences: "if this result appears, we adopt" and "if this result appears, we stop". A PoC without those two does not end.

Before starting — settle three things

One target process — Validating several processes at once will not finish in four weeks. Pick one that repeats often and is reversible.

Success and stop criteria — Write them as numbers — "30% reduction in minutes per item, error rate under 5%". Qualitative criteria produce an argument in week four.

The current method’s numbers — Monthly volume, minutes per item, current error rate. Without these the PoC result cannot be interpreted.

Weekly goals

Week	Goal	Output
1	Secure data and access, set up accounts	Sample dataset, accounts, baseline numbers
2	Basic flow working — happy path only	Results and review records on normal inputs
3	Exceptions and wider scope	Exception list and how each was handled
4	Measure, write up, decide	Comparison table and adopt/stop recommendation

How to actually finish in four weeks

Do not build integrations. File download and manual upload are enough at PoC stage. Integration work belongs after the decision, and it is the most common reason half the PoC window disappears.

Do not build a UI either. You can judge results from a table. If a UI feels necessary, that is a pilot-stage requirement, not a PoC one.

And record review time without fail. Omit that number and every post-adoption savings estimate comes out optimistic.

Signals to stop

The happy path still fails at end of week 2 — The tool or the data does not fit. Three or four more weeks rarely changes that.

Review takes longer than doing it manually — If that holds through week 3, stopping is right. If fixing the output format and prompts does not help, it is the tool.

Data terms are unconfirmed — Even if it works technically, you cannot adopt what does not meet the terms. If confirmation drags, pause the PoC and finish the terms first.

Scope keeps growing — Repeated "it would be nice if it also did this" is a sign the end condition blurred. Return to the original criteria.

Frequently asked questions

Is four weeks too short?

With one target process and no integration or UI work, it is enough. Feeling short usually means the scope is large, and reducing scope produces better results than extending time.

How is a PoC different from a pilot?

A PoC establishes whether it is possible; a pilot establishes whether it holds up in real operation. A PoC can run on sample data and manual steps; a pilot has to attach to the real flow.

Do we need outside help?

Internal people who know the process matter more. Outside help saves time on tool setup and technical judgement, but the success criteria and business rules have to come from inside.

What if the result is ambiguous?

Do not lower the criteria after the fact. Stop, and record what was missing. That record is what makes the next PoC’s scope accurate.

Baseline numbers

Build the comparison baseline

The AI ROI calculator turns your current volumes, times and costs into a table — the baseline the PoC compares against in week four.

Calculate ROI