Measuring an AI project from the board: baseline and one metric per problem
An AI project is measured from the board with two things settled before it starts. The first is the baseline: how the process stands today, with its value, its source and the way it is measured. The second is one metric per problem, with an owner. After that, you compare. Without that measurement, the board is given impressions.
Updated: 5 November 2026 · Deibson Silva · Forward Deployed AI Engineer and Founder
Why is measuring an AI project the board's job?
Because the decision to go on, expand or stop is theirs. And a decision like that is only taken well with a before and an after that everybody accepts.
The technical team measures the system: whether it responds, how long it takes, what each query costs. That is needed, but it tells a board little. Senior management cares about the process: whether the work comes out better than before and whether people use it.
Under Regulation (EU) 2024/1689, the company using an AI system under its own authority is the deployer. This article speaks to whoever runs that company, and stays with the method: what to ask for, when and from whom.
The starting rule is short. Without a baseline there is no metric, only a wish.
How do you set a baseline before you start?
It is done before anything is switched on, and it takes less time than it seems.
- 1Choose a specific process. «We want AI in the back office» cannot be measured. «Sorting the documents our clients send in» can.
- 2Name the owner of the process. They will sign the measurement and they will decide afterwards.
- 3Measure how it stands today: how long it takes, how many cases come back because of an error, how many sit waiting for a piece of data.
- 4Note the source of each number: which system or which record it comes from.
- 5Note the instrument: how it will be measured in exactly the same way later on.
- 6Take the measurement with the team in the room and real cases in front of you. A number the team does not recognise is no use.
- 7The process owner signs it. From that moment it is the reference for the whole project.
What cannot be measured today is declared as such, with its measurement plan. It is not filled in with an estimate.
Which metrics are useful to a board, and which are not?
One per problem. If a problem needs five metrics to be explained, it is five problems.
| Not useful | Useful | |
|---|---|---|
| Quality | «That it works well» | Cases resolved with no correction by a person, out of the total |
| Use | People with access to the product | Cases in the process that really go through the system |
| Time | An estimate of hours released | The measured time of the task, against the baseline |
| Risk | «There have been no incidents» | Cases the system passed to a person, and for what reason |
| Cost | The price of the product, on its own | The cost per case resolved, with the model actually in use |
Which questions does the board ask at each review?
Four, always the same. They fit on a one-page table.
Against which baseline?
Each number is compared with the one signed before starting: value, source and date. A figure on its own does not say whether anything has changed.
What was measured, and where does it come from?
Today's value, with the same source and the same instrument as the baseline. No number without a source.
What does it cost to keep going?
The hours a month somebody spends operating and reviewing the system. They are part of the result.
What is the next decision?
Carry on as it is, give the system more autonomy, take it to another process or stop it. Each review ends with a decision by the owner.
Which numbers should not reach a board?
Projections. A tracking table compares what there was with what there is. A column with what there will be turns the report into a sales plan.
Usage figures for the product without the process beside them. A thousand queries in a month do not say whether the work comes out better.
And averages that hide the failures. A system that gets it right nearly always and gets it wrong in the case that hurts most needs somebody to look at that case, by name.
The simplest test is one of use. If on Monday the team goes back to doing the task by hand, the project has not worked, however good the demonstration was.
What is measured before the system goes live?
Two things, and both can be asked for from the board without going into the technique.
The first is the ceiling. The prototype is measured first with the most capable model there is, to find out how far the case goes. Only then is a smaller, cheaper one tried, and you move down if the tests support it.
The second is a suite of tests with real cases. The method starts from 20 to 50 tasks drawn from genuine failures in the process. If the system does not pass them, it does not go out.
With those two measurements, the conversation about cost stops being a matter of opinion. You know what is lost and what it costs with each model, in that specific process.
Where does a senior leader learn to run an AI project this way?
At the SciTechBiz Institute, the Advanced Executive Programme that deals with AI at board level is Applied Intelligence · AI and Cyber Behaviour. It is designed for CTOs, CIOs and heads of AI who have to decide on real deployments and answer for them to a board.
The division of roles is the same as in this article. The senior leader decides what is done and asks for the measurements. The builder, the Forward Deployed Engineer, sets the system up inside the company, measures it and hands it over.
The programmes, in detail
Honest limits
- A good measurement shows earlier that a project is going badly. Whether it goes well depends on the work.
- This article gives no market figures and no averages from other companies. The only comparison that counts is your process against itself.
- The terms and the price of each programme are on its page at the Academy.
- The SciTechBiz Institute is not a university: its certifications and Advanced Executive Programmes are awarded by the Institute itself and are not recognised by any education authority.
- Education attests a competence. It is not legal advice on any particular system.
What people ask most
01What is the baseline of an AI project?
The measurement of how the process stands before you start: today's value, which source it comes from and which instrument measures it. The process owner signs it, and everything measured afterwards is compared with it.
02How many metrics does an AI project need?
One per problem, with an owner. If it takes many to explain a single problem, it is better to split it into several, each with its owner, its metric and its baseline.
03Who should own the metric?
The person who answers for the process at the company, rather than whoever builds the system. They sign the baseline and they take the decision at each review.
04What do I do if something cannot be measured today?
Say so, and write a measurement plan: which data is missing, who will collect it and from when. Filling the gap with an estimate leaves the project with no reference.
05How often is it reviewed?
In the method, the usual pattern is a short review while the system is being deployed and a monthly one with the board once it is in use. Each one ends with a decision.
Sources
- Regulation (EU) 2024/1689, the AI Act (EUR-Lex): Article 3(4) · accessed 23 October 2026

Founder of Legathum and iPersonality and Forward Deployed AI Engineer. He leads the SciTechBiz Community from Madrid and examines the FDE Certification defence.
See it with your own case
A 15-minute call with Deibson Silva to see the platforms with what you do and answer your questions.
iPersonality Global Institute, S.L. · NIF B27580042 · Madrid · Privacy policy
Prices in euros, exclusive of VAT. In Spain, Spanish VAT is added. In the rest of the EU, VAT is calculated at payment according to your country and VAT number. No Spanish VAT is charged to UK customers.