AI Engineering 9 min read

What an AI proof of concept should prove first

A proof of concept should show it works on your records, how often it is right against a test set, what each task costs, where a person checks it and what it connects to. Without all five, you watched a demo.

Founder and Head of AI Engineering

An AI proof of concept should prove five things before you spend more on it: that it works on your own records, how often it is right against a test set written before the build, what it costs to run per finished task, where a person checks its work, and what it has to connect to. If a proof of concept ends without a written answer to all five, what you watched was a demo.

Keep it small. One job, one team, two to four weeks, and a test set of real cases from your business, including the awkward ones. Agree the pass mark before the first output, so "good enough" is a number and nobody argues about it afterwards.

Done means a short write-up: the score against the pass mark, the cost per task at your real volume, the cases it got wrong, where a person stays in the loop, and a recommendation to go, change or stop. Any of those three is a successful proof of concept, because each one stops you spending money on a guess.

What is the difference between a proof of concept, a pilot and production?

The three words get used loosely, and a vendor demo is sometimes sold as a pilot. Here is how we use them, and what each stage should hand you.

StageThe question it answersWho uses itDataWhat you hold at the end
Proof of conceptCan this do the job on our records, well enough and cheaply enough?The builder, plus two or three of your people who grade the outputA copy or sample of real records, and a test set with the correct answersA score against the pass mark, cost per task, a list of failures, and a go, change or stop call
PilotDoes it keep working beside the people who do the job today?A small group in their normal day, with a person checking the outputLive records, read only or with approval before anything changesGraded results on live work, the time each human check takes, and the gaps found
ProductionCan the business rely on it every day?Everyone who does the jobLive, with access no wider than the user's ownA log of every run and its cost, tests rerun on every change, a rollback, and one named owner

Each stage costs more than the one before and is harder to walk back from, which is why the proof of concept has to answer its questions properly before a pilot starts. Our guide to what AI consulting costs in Australia sets out how the stages are usually priced.

What five questions should an AI proof of concept answer?

These are the five we write at the top of every proof of concept, before anything gets built.

  • Does it work on your real data? Run it on your own records, with your field names, your gaps and your duplicates. A model that answers well from a tidy sample can fall over on a CRM where half the companies have no industry set. If the data needs cleaning first, the proof of concept should say so and size the job.
  • How often is it right against a test set? Write a few dozen to a couple of hundred real cases, each with the correct answer, before the build starts. Some teams call this a golden set. Include the edge cases: missing data, ambiguous requests, and things it must refuse. Score every run against the whole set, and have someone other than the builder do the scoring.
  • What does it cost to run per task? Add up the model usage or platform credits across the test set, divide by the number of cases, and multiply by how many times a month the job really happens. That gives you a monthly running cost worked from your own numbers.
  • Where does a person check it? Decide which outputs go straight through, which a person approves first, and which it should never touch. Time the checks, because that staff time is a running cost too.
  • What does it connect to? List every system it reads from and writes to, and whose access it uses. Read only is usually enough for a proof of concept. Each extra system adds integration work, and that is where most of the build cost goes.

The fourth question is the one a demo skips. A demo shows the model's best answer. A proof of concept should show what happens on its worst one, and who catches it before a customer sees it.

This video from our YouTube channel looks at where AI needs a human, and where it does not.

Where AI needs a human, and where it doesn't, from our YouTube channel. Watch it on YouTube.

If the thing being proved is an agent, add a sixth question: what is it allowed to do on its own? Anthropic's engineering guidance on agents notes that their autonomy brings higher costs and the potential for compounding errors, and recommends the simplest solution that does the job. So an agentic AI proof of concept should come with a written action policy: routine steps run and get logged, sensitive ones notify a person, and anything that cannot be undone waits for someone to approve it. Often a workflow with a single AI step does the job with less to go wrong, and the proof of concept is where you find that out.

How do you scope an AI proof of concept in two to four weeks?

The scope should fit on one page. If it does not, the proof of concept is too big. This is the order we work in:

  • Week 1: pick the job and write the test set. One job your team can name, with a volume you can count. Pull real cases, write the correct answer for each, agree the pass mark for each type of case, and get read access to the data.
  • Week 2: build the smallest version that can attempt the test set. The prompts, the lookups and the one connection it needs. Nothing goes in front of a customer.
  • Week 3: run, score, fix, run again. Every run goes against the full test set, scored by someone who did not build it, with the cost of each case logged.
  • Week 4: write it up and decide. The score, the cost per task, the failure list, the human checks, what a pilot would need, and a recommendation.

A simple job on clean data can finish in two weeks. An agent that touches several systems needs closer to four. If a proof of concept runs past four weeks, it has usually turned into a build without anyone deciding to start one.

From a named job to a decision in four weeks
  1. Week 1Real cases written up with correct answers, and a pass mark agreed for each type of case.
  2. Week 2The smallest version that can attempt every case, with one connection and nothing customer facing.
  3. Week 3Every run scored against the full set by someone other than the builder, with cost logged per case.
  4. Week 4The write-up and the go, change or stop call.

The order described in this post. A simple job on clean data can finish in two weeks.

Before week one, get five things agreed in writing: the job, its monthly volume, who owns the test set on your side, the pass mark, and the one person who makes the go, change or stop call.

Weighing up an AI build for your business?

Talk to Matt about AI projects and how we work.

What does "done" look like at the end of an AI proof of concept?

A write-up of one or two pages that someone who was not on the calls could read and act on. It should hold:

  • The score. Results for each type of case, against the pass mark agreed in week one.
  • The failure list. Every case it got wrong, with the cause: the data, the prompt, missing context, or the model's judgement.
  • The running cost. Cost per finished task, and per month at your volume, with the staff time for checks included.
  • The human checkpoints. Which outputs a person approves, and how long each check took.
  • What a pilot would need. Systems, access, data clean-up and people.
  • A recommendation. Go, change or stop, with a cost range for the next stage.

Ask for the build cost and the running cost of the next stage as two separate numbers. Our guide to what a custom AI build costs an Australian business explains what moves each one. For reference, our AI projects start from A$10,000 and ongoing AI partnerships from A$8,500 a month, both ex GST.

Why do AI proofs of concept stall after the demo?

Most stall for reasons that have little to do with the model. We plan for each of these before the build starts:

  • No pass mark agreed up front. Without one the result is an opinion, and a room full of opinions rarely approves a budget.
  • The demo ran on tidy data. The first week on live records then turns up the gaps, and the project becomes a data clean-up nobody planned for.
  • Nobody priced the running cost. The build looked affordable, then the monthly bill at real volume arrived as a surprise.
  • The risk questions came late. What happens when it is wrong, who gets to see which data, and whose access it runs under. We answer those before we build anything.
  • No owner in the business. If nobody is named to make the call, the proof of concept sits there until somebody asks about it.
  • The connection was left for later. Reading a spreadsheet export is easy. Reading and writing your CRM with the right access is the real project.

The same questions help when you compare firms. Our guide to choosing an AI implementation partner in Australia covers what to ask, and the five checks we run on CRM data before an agent goes near it covers the data side.

Go, change or stop: how do you decide?

Use the write-up and this table. The person named in week one makes the call, on the evidence.

CallWhat the result looks likeWhat happens next
GoMeets the pass mark on the case types that matter, costs less per task than doing it by hand, and failures are quick for a person to catchA pilot with a small group, a review queue and a way to roll it back
ChangeClose to the pass mark, with failures clustered on one cause such as missing data or one difficult case typeFix that one cause and rerun the same test set, on a date agreed now
StopWell short of the pass mark, more expensive per task than a person, or the job happens too rarely to pay its wayStop spending. A plain workflow or a data clean-up is often the better first move

Stopping after three weeks is a cheap answer to an expensive question. Keep the test set whichever way the call goes. It is the most reusable thing a proof of concept produces: when a newer model or a better approach comes along, you can rerun it in minutes and know whether the answer has changed. Our post on how we test an AI agent before it goes near a customer shows what a test set and a pass mark look like in practice.

A change sends you back to the same test set
How a proof of concept ends in go, change or stopA test set and pass mark feed the proof of concept, which runs, scores and fixes over two to four weeks. It ends in one of three calls. Change means fixing one cause and rerunning the same test set. Go means a pilot with a small group, a review queue and a rollback. Stop means stopping spending and keeping the test set. Test setreal cases, pass mark agreed Proof of conceptrun, score, fixtwo to four weeks Changefix one cause, rerun the same test set Gopilot: small group, review queue, rollback Stopstop spending, keep the test set

The three calls from the table above. Whichever call is made, the test set is kept for the next run.

Frequently asked questions

What is an AI proof of concept? A short test of one AI use case on your own data, scored against a test set and a pass mark agreed before the build. It ends with a written go, change or stop decision.

How long should an AI proof of concept take? Two to four weeks for one job with one team. If it runs longer, the scope has usually grown past the question it was meant to answer.

What is the difference between a proof of concept and a pilot? A proof of concept tests whether AI can do the job on your data, against a test set, away from live work. A pilot runs it beside your team on live work, with a person checking the output.

Why do AI proofs of concept fail after the demo? Usually because no pass mark was agreed up front, the demo used tidy data, nobody priced the running cost at real volume, the risk questions came late, or nobody was named to make the decision.

What should an agentic AI proof of concept test? The same five questions as any proof of concept, plus what the agent may do on its own. That means a written action policy, a person approving anything that cannot be undone, and test cases it must refuse.

Sources


Neighbourhood builds AI systems on Claude for Australian businesses, starting every build with a test set and a pass mark. How that work runs is on our AI engineering page.

Got a job you want AI to prove itself on?

Talk to Matt about AI engineering. He can explain how we work and whether our AI engineering service fits what you need.

Neighbourhood

Neighbourhood is a HubSpot Diamond Partner in Brisbane. We build AI systems and the revenue operations they run on, for businesses across Australia and New Zealand.