A guide for leadership, IT departments and project managers

AI POC: how to go from demo to real-world use

An AI POC (proof of concept) is a short test that checks, on your data and a specific task, whether AI does the job well enough to keep going. Many stay demos, because no one wrote down beforehand what would count as success. A POC is for deciding, not for showing off.

Book 20 minutes with us See the POC sheet

Updated October 11, 2026

01Definitions

POC, pilot, MVP: what's the difference?

The POC proves a task is feasible with AI, the pilot proves a team actually uses it, the MVP is the first version delivered to everyone. Three stages, three questions.

POC, pilot and MVP compared for an AI project in a company
Criterion	POC	Pilot	MVP
Question asked	can AI do this task?	does the team really use it?	can we deliver it to everyone?
Who tests	a few people who do the task	a whole team, on its day-to-day work	all target users
Data	a sample of real files	everyday data	production data
Outcome	continue, fix or stop	roll out or not	continuous improvement

02The causes

Why AI POCs stay demos

An AI POC stays a demo when it proves the tool is impressive, not that it changes the work. Four causes keep coming up, and none of them has to do with the model.

No written criterion up front

People judge after the fact, on a gut feeling. A successful demo convinces everyone and decides nothing.

Demo data

Hand-picked, clean examples with no difficult cases. On real files, the quality isn't the same.

A test outside your tools

The test runs in a separate interface. Using it would mean copying and pasting: no one does.

No one to make the call

No owner, no decision date. The POC quietly fizzles out, or drags on forever.

03The method

Scoping an AI POC: the one-page sheet

Scoping an AI POC means writing down on one page, before the first test, the task, the success criterion, the data, the tool, the end date and who decides.

POC sheet: fill in before the first test

The task: just one, frequent, with a verifiable result: for example, drafting replies to quote requests.
The success criterion: the threshold to reach, compared with the current way of working: time, quality, volume or turnaround.
The data: a sample of real files, difficult cases included, and what is allowed to go into the tool.
The tool and the workflow: the tool the team will actually have, where it already works, and who reviews each answer.
The testers and the end: people who do the task every week, and an end date set in advance.
The decision: who decides, on what date, and what happens if the threshold is reached, nearly reached or missed.

We can fill in this sheet with you, on your own tasks.

Book 20 minutes

04In the field

Test on your data, in your tools

An AI POC is only worth something if it's run on real files, in the tools where the team already works. Otherwise, it measures a demo.

Keep the difficult cases: an incomplete file, an ambiguous request, a scanned document.
Set the data rule before the test: which information can go into the tool, and with which account.
Test with the tool the team will actually have, for example Claude Team or Copilot, not a separate trial version.

If the AI has to chain actions in your tools, you're already testing an AI agent: plan from the POC stage who approves each action.

05The ROI

Measuring the ROI of an AI POC: four metrics

The ROI of an AI POC is measured on the task tested, with metrics recorded before and during the test, never with a percentage gain announced in advance.

Metrics for measuring the ROI of an AI POC
Metric	What we record	The question
Time spent	minutes per file, before and with AI	is it shorter, review included?
Quality	share of answers usable after review	how many go out without rewriting?
Volume processed	files processed per week	does the team handle more?
Turnaround	time between request and reply	does the customer wait less?

Also count the full cost: subscriptions, testers' time, review. A gain that disappears during review isn't a gain.

06The decision

The written decision: continue, fix or stop

An AI POC ends with a written decision, made on the planned date. Stopping a POC that misses its threshold is a result, not a failure.

Illustrative example, fictional team

End of POC: replies to quote requests

Criterion set: 8 replies out of 10 ready to send after a quick review.
Result: 6 out of 10: incomplete requests make the rest fail.
Decision: we fix: a check of the attachments before the AI step.
Next checkpoint: new test in three weeks, same criterion.

Our three-step method follows the same logic: each step ends with a written deliverable, and each follow-up checkpoint with a written decision.

07Next steps

After the POC: scaling up and an AI roadmap

A successful POC doesn't scale on its own: you need a pilot with a real team, then an AI roadmap that ranks the next projects by value and by risk.

Pilot: a whole team, on its day-to-day work, with the same criterion.
Industrialization: accounts, data access, usage rules, someone to ask questions.
Training: the team learns the workflow, not just the tool.
Roadmap: every new project goes through the same sheet, so you can compare and decide.

Without an official tool after a successful POC, everyone carries on alone with their personal account: that's how shadow AI is born.

Kailli, an applied AI transformation firm, helps companies that already use AI scope their POCs and make decisions. It sells no tools.

Written by Rémi Carbonnefounder and CTO of Kailli

Frequently asked questions about AI POCs.

What is a POC in AI?

An AI POC (proof of concept) is a limited test that checks whether an AI tool can perform a specific task, on real data, at a quality level set in advance. It doesn't prove the team will adopt it: that's the pilot's job.

How long should an AI POC last?

Long enough to process a representative sample of real files, no longer. A daily task can be judged faster than a monthly one. The key is to set the end and decision date before you start.

How much does an AI POC cost?

There's no universal price: the cost depends on the task, the data to prepare, the tools and the team's time. Count subscriptions, testers' time and review, not just the service fee. At Kailli, everything is priced in writing after a 20-minute call.

How do you measure the ROI of AI?

On a specific task, by comparing before and after: time spent, quality, volume processed, turnaround. Add the full cost: subscriptions, review, training. Be wary of universal percentage gains: the only number that counts is the one measured at your company.

When should you stop an AI POC?

When the criterion set at the start isn't met by the planned date and no simple fix is in sight. Stopping early frees up budget for a better candidate. A written stop decision, with its reasons, helps the next POC.

Twenty minutes. And you'll know whether we can help.

No sales pitch. You tell us where you stand with AI, we tell you how far you can go.

Book 20 minutes Prepare for the call