AI in business: three uses that pay off, three that waste time

The topic is saturated with promises, and most lists of use cases look alike. What is missing is not another list: it is a sorting rule you can apply to your own situation, including to a use that does not exist yet.

Here it is, in three conditions. A profitable use has a verifiable output, sufficient volume, and cheap failure. All three are required. A use meeting only two produces a convincing demo and an abandonment six months later.

The sorting rule, condition by condition

Verifiable output. Can someone tell, in seconds and without particular expertise, whether the result is right? An amount extracted from an invoice is checked at a glance against the document. A six-month sales forecast can only be checked after six months — and by then nobody remembers to compare.

Sufficient volume. Setting up has a fixed cost: integration, testing, training, supervision. A task performed three times a month will never repay it, however perfectly automated. The question is not “is it feasible” but “how many times a week”.

Cheap failure. What happens when the output is wrong — because sometimes it will be? If the error is visible and recoverable, the use holds. If it reaches a customer, a filing or an irreversible decision, it needs supervision that cancels the gain.

Three uses that pay off

Document extraction

Reading supplier invoices, delivery notes, statements, contracts, and pulling structured fields out of them. All three conditions hold: the result is checked against the source document, the volume is daily, and an error is corrected before validation.

It is the least spectacular use and the most reliably profitable. It replaces data entry, not judgement.

First-line answers

Answering questions whose answer already exists in your documentation: opening hours, return procedure, order status, contract terms. The condition to respect is that the system answers from your documents and cites its source, rather than producing a plausible reply.

Second condition, non-negotiable: a handover to a human, visible and immediate. A system that traps the user in a loop costs more in frustration than it saves in time.

Reviewed drafting

First versions of product descriptions, standard replies, meeting notes, working translations. The gain is real because correcting a text is faster than starting from a blank page.

It disappears entirely if the review is skipped. That is the only condition, and it is more fragile than it looks: after a few weeks the temptation to publish directly sets in, precisely because the output is often fine.

Three uses that disappoint

Unsupervised decisions

Granting credit, rejecting an application, refusing an order, applying a penalty. Beyond the regulatory risk — several legal frameworks govern automated decisions with a significant effect on a person — it is the cheap-failure criterion that disqualifies it: a wrong decision is hard to detect, since you never see what would have happened otherwise.

Publishing without review

Filling a site or a news feed automatically. The volume is there, the cost too, but nobody checks the output. A factual error published under your name costs more than all the time saved, and it stays indexed.

Forecasting on a short history

Predicting demand, stockouts or customer churn from two years of irregular data. A result will be produced, presented cleanly, and unverifiable for a long time. It is the use that consumes the most budget for the least observable return.

Your data: what to check before connecting anything

Using an online service means sending data to a third party. Four questions to ask the provider, and to get in writing:

• Where is the data processed, and under which legal regime? The point is decisive if you handle personal data of European residents.
• Is it retained, and for how long? A vague answer is an answer.
• Is it used to train the provider's models? That must be in the contract, not in a help page liable to change.
• What happens if the service closes or changes its terms? A business process built on a single service is a borrowed process.

For the most sensitive data, models can run on your own infrastructure. That is more expensive and often less capable — but the comparison only means something once the confidentiality constraint is stated, not before.

What it actually costs

Three items, only one of which appears in offers.

Usage-based billing is the one everyone looks at. It is predictable once the volume is known, and it is rarely the problem.

Integration — wiring the system into your tools, handling edge cases, dealing with failures — is most of the initial budget. A demo is built in an afternoon; a production use takes the same work as any other feature.

Supervision never disappears. Someone has to keep checking a sample, otherwise a silent drift sets in. That time belongs in the calculation from the start, or the announced gain is a gross figure.

A use that requires as much checking as it saves in work is not a bad use: it is a use that is not yet in the right place.

Where to start

One case, chosen because it satisfies the three conditions — not because it is impressive. Measure beforehand: how long the task takes today, how many errors it produces. Without that baseline you will never know whether the result is good, and the discussion will be settled on impressions.

Refuse a general rollout as step one. It turns a reversible trial into a commitment, before anyone has seen how the system behaves on your edge cases — which are, as always, half the work.

Our AI & Data and IT Advisory & Digital Transformation pages cover implementation and case selection respectively. The sorting rule above applies without us.