Companies often ask where artificial intelligence should be introduced into the business.

Much less often do they ask the question that should come first:

Which decisions, precisely, would benefit from being supported by a model - and which would not?

An AI initiative never improves a process in the abstract. It improves a decision: which receivable to pursue first, which inventory level to adjust, which maintenance intervention to schedule, which transaction to approve.

If the wrong decision is selected, no amount of technical refinement will compensate for that mistake.

The first task is therefore to identify the right decisions - before the data, before the models, before the vendors.

Three families of decisions

Business decisions differ in many ways, but frequency is one of the most useful distinctions.

Strategic decisions

Acquiring a company, restructuring a business, closing a site, entering a new market or changing the business model.

These decisions are rare, unique and highly dependent on context. Each is taken once, in circumstances that are unlikely to recur in exactly the same form.

They are therefore poor candidates for automation. AI lacks the repetition and historical volume on which its predictive strength depends.

But a decision that should not be automated may still be supported.

A purpose-built model can inform a strategic decision by generating scenarios, testing the assumptions behind an investment case or challenging a recommendation from angles that may have been overlooked.

AI does not make the decision. It improves the preparation.

Tactical decisions

Staffing levels, commercial resource allocation, inventory planning, supplier monitoring, and trade-offs between customers or products.

These decisions translate strategy into day-to-day management. They recur, rely on data and carry real economic value.

They are strong candidates for AI, usually with human oversight.

Operational decisions

Validating an order, routing a service request, approving a transaction or prioritizing maintenance.

Companies take these decisions every day, at every occurrence of a process.

Individually, each may carry limited weight. In aggregate, they matter considerably.

Because they are frequent, numerous and data-rich, they are among the most fertile areas for AI.

At the far end of this spectrum are micro-decisions: one decision per customer, tailored to each interaction - which price to offer, which product to recommend, which follow-up action to trigger.

Their frequency, volume and cumulative value make them the most natural candidates of all.

This requires a change in perspective.

The greatest value from AI is not necessarily found at the top of the organization, in the rare and prestigious decision. It is often found in the large volume of modest, repeated decisions that the organization makes almost without noticing.

Five questions for assessing a decision

Frequency alone is not enough.

Before treating a decision as a serious AI candidate, five questions require explicit answers.

Does it occur often enough?

Repetition provides two things at once: enough past examples for a model to learn from, and enough volume to justify the investment.

A rare decision usually provides neither.

Do we have the relevant data?

Not data in general - every company has data - but the specific signals on which this decision depends, with sufficient depth and consistency to be usable.

This is the criterion that disqualifies the largest number of apparently obvious candidates.

Can the alternatives and desired outcome be clearly defined?

A decision suitable for AI is not necessarily a simple one. In fact, the opposite is often true.

A trivial decision that can be resolved with a two-line business rule does not justify the investment.

Value appears in non-trivial decisions: those that draw on many signals, involve several imperfect outcomes, and exceed what a human can process quickly and consistently.

But the complexity must remain structured.

The inputs must be identifiable. The alternatives must be definable. The desired outcome must be explicit. The acceptable errors must be bounded.

Complex, but frameable.

What cannot be formulated cannot be modelled. What is too simple does not need to be modelled.

Is the impact measurable?

The effect of the decision must be linked to a concrete indicator: margin, cash, cost, service, lead time, quality, risk, conversion or asset utilization.

Without measurement, the company will never be able to demonstrate the model’s value or decide whether it should be maintained.

Can the organization act on the result?

This is the most frequently overlooked criterion - and the most decisive.

Predicting that a customer is likely to leave has value only if the company can still retain that customer.

Anticipating a failure has value only if an intervention can be scheduled.

Identifying a supply disruption risk has value only if the company can change its sourcing or inventory response.

Value does not reside in the score. It resides in the decision that the score allows the organization to change.

An accurate prediction on which no one acts has no value.

A decision that passes all five tests is a serious candidate.

A decision that fails one of them deserves to be paused. In many cases, a business rule, conventional automation or stronger operational discipline will deliver the required result faster and with less risk.

The remaining question once the decision has been selected

A positive answer to all five questions establishes that a decision is suitable for AI.

It does not yet determine how AI should be introduced into that decision.

A sixth question is required:

Is the cost of an error compatible with the level of human control being considered?

One incorrect transaction among thousands may be corrected without material damage.

An unjustified credit refusal or an unnecessary major maintenance intervention cannot.

Depending on the consequences of error, the same decision - even if it passes the first five tests - may require a different operating model.

The real choice is therefore not simply “AI or no AI”.

It is a choice between three modes.

Automate

The decision is taken without human intervention because it is frequent, well structured and the cost of an isolated error remains limited.

Assist

AI makes a recommendation and a responsible manager validates it.

This is the appropriate model for many tactical decisions, where human judgement retains the final say.

Inform

AI supports the decision through scenarios, analysis or structured testing of assumptions, without taking part in the decision itself.

This is the appropriate mode for strategic decisions, where judgement should be strengthened, not replaced.

Confusing these three modes is a source of failure in itself.

Automating what should have remained assisted leads to a loss of control.

Merely informing a decision that could have been automated leaves value unrealized.

The question is therefore not whether the company “uses AI”.

It is which decisions deserve to be supported by AI, and with what degree of autonomy.

This sorting exercise appears simple. The five criteria can be read in five minutes.

Applying them honestly to forty or fifty use cases proposed by the business, comparing those cases with one another, and then testing the strongest candidates against real data is far more demanding.

Few management teams complete that exercise successfully on their own.

Not because they lack capability, but because organizations are rarely impartial judges of their own ideas.

Teams defend the use cases they proposed. Sound prioritization requires a perspective with nothing to protect.

One question is often enough to expose the gap:

How many AI initiatives are currently under way in your company - and for how many of them could you name the precise decision they are intended to improve, together with the business owner accountable for the economic outcome?

If the answer is unclear, the problem is not technological.

It is a prioritization problem.

That is precisely the purpose of a prioritization diagnostic: to sort before investing, distinguish what should be automated from what should be assisted or merely informed, and provide management with a sequence of decisions it can explicitly approve.

The first step requires nothing more than a serious conversation.

This perspective on decision-making draws on frameworks covered in the MIT xPRO AI Strategy and Leadership Program, extended through Nobilys’ operational practice.

Christian Cirino — Nobilys Group