How to choose a profitable AI project among several ideas

A plant rarely has a single possible AI project. It has ten: reconstructing the inspection history, predicting failures, reading supplier reports, monitoring energy, preparing audits. The real question is not "can we do AI", but "which of these projects should we fund first".

Picking the wrong project costs more than launching nothing: the wrong project consumes a budget, ties up teams and closes without proving anything, leaving the conviction that "AI does not work here".

Choosing a profitable AI project is therefore not a matter of intuition. It is a sorting exercise: laying a list of candidates side by side, scoring them on the same criteria, and funding the one that pays back earliest and most reliably.

This article is about that upstream step, the choice itself. Once the project is picked, its return still has to be costed in the business case of an AI project, and then executed without scattering, as described in a first AI project in 90 days. Here, we only decide which one deserves the budget.

What makes an AI project profitable rather than a money pit?

The profitability of an AI project comes down to one simple idea: the size of the pain multiplied by the share you will actually capture.

The size of the pain is its frequency multiplied by what it costs each time. A painful but rare task weighs little; a small task repeated every day weighs a lot. It is the product of the two that counts, never one without the other.

The share you will capture is what separates theory from real money. It depends on three things: technical feasibility on your documents, the availability of the data that feeds the tool, and the organisation's ability to turn the time saved into time reallocated. A gain that stays on paper is not a gain.

A profitable project, then, is a large pain that the technology can handle, on data that is present, whose gain will actually be harvested. Remove a single one of these elements and the project slides towards the money pit, however well it dazzles in a meeting.

What are the six criteria that do the sorting?

Six criteria are enough to separate a list of candidates. They are not about the solution, but about the problem-data pair, which is to say about you.

1. Data availability

An AI feeds only on what already exists. If the reports are scattered across PDFs, poorly tagged or impossible to find, the first project is not AI, it is the data. A project that assumes the data has to be built before it can even start is not a profitable AI project, it is a data project in disguise. This question often decides on its own, and it is developed in choosing an industrial AI solution.

2. Frequency of the pain

How often does the task recur: every day, every week, once a year? The higher the frequency, the more a small unit gain compounds into an annual gain that shows. An annual task, however heavy, takes years to pay a project back.

3. Value per occurrence

What does the task cost each time: time spent, people involved, the consequence of an error or a delay? Frequency and value multiply together. A frequent but trivial pain is worth no more than a heavy but rare one.

4. Technical feasibility

Can the tool actually do it, on your real documents and not on a clean demo? Reading a well-formatted report is easy; finding a measurement on a crooked, hand-signed scan is far less so. Feasibility is checked on your own material, never on the brochure.

5. Reversibility

What happens if the AI gets it wrong? An error caught before any consequence is an incident; an error that carries a safety decision is an accident. A profitable first project sits on tasks where you are allowed to be wrong while you learn. Anything that carries liability is kept for later.

6. Time to first result

How long before you see something verifiable? A project that proves itself within a quarter keeps trust alive and unlocks the next one. A project that promises a big result in eighteen months runs out of steam before it lands, and its gain, however real, arrives too late to count.

How do you score and rank a shortlist?

These six criteria only help if you apply them to every candidate the same way. The method comes in three moves: list the projects under consideration as their champions frame them, score each candidate from 1 to 3 on each of the six criteria (1 for weak, 2 for medium, 3 for favourable), then add up and see what rises.

The score is not a measurement, it is a shared judgement. Its virtue is not the precision of the number, but forcing an explicit conversation: why does this project score "1" on data? The discussion is often worth more than the total. The values below illustrate the method; yours will depend on the real state of your site.

Table scrolls horizontally on small screens.

The ranking reads at a glance: the document tasks, where the data already exists, rise to the top, while the most spectacular subject, failure prediction, falls to the bottom for lack of data and feasibility.

Recap: what to check, and the red flag

Table scrolls horizontally on small screens.

Which traps turn a project into a money pit?

Four projects come up at every scoping workshop, seductive and consistently disappointing. Recognising them stops you sinking a budget into them.

The vanity pilot. You pick the subject that will impress the board, not the one that pays. The demo shines, everyone applauds, then nobody opens the tool again because it solves no real pain. A pilot is judged on what it pays back, not on the effect it produces in a meeting.

The project with no data. The pain is real, the value is there, but the data that would feed the AI does not yet exist. You then fund a collection exercise disguised as an AI project, whose cost and timeline no longer bear any relation to the initial estimate. Without data present, the best use case waits.

The boil-the-ocean scope. Wanting to handle the whole asset base, every document, every site at once. The project never finishes, the gain dilutes, and the proof never arrives. A narrow scope that lands proves more than a total scope that bogs down.

Phantom savings. The gain was real on paper, but nobody harvested it. The freed hours were neither reallocated nor removed: they dissolved into the day-to-day, and no saving appears anywhere. A project is only profitable if someone decides, in advance, what the time saved will become.

Once you have chosen the project, what next?

Sorting is only the first step. The chosen candidate still has to be costed, to check that the estimated gain exceeds the entry cost, then executed on a bounded scope. And to place each use case within the overall picture of what AI changes in the plant, industrial AI sets the scene.

To put down the first half of the calculation, the current cost of a task, Assets 4.0 offers an ROI calculator, useful for comparing several candidates on the same basis before keeping one.