What a Leader Must Decide Before Buying AI

Buying Before Deciding

The pressure is constant. Every week, a new AI tool promises to change everything: customer service, writing, analysis, prospecting. The temptation to buy one is strong—to avoid falling behind, to show that the company is moving forward. And that is often how the spending that brings no return begins.

Buying an AI tool without having settled what it is meant to do is buying an answer to a question that was never asked. The tool arrives, a few people try it, the enthusiasm fades, and six months later the subscription is still running while no one can say what it delivered. The fault is not in the tool. Today's tools are good. The fault is in the order: the purchase came before the decision.

The market confirms this gap. Adoption is now widespread, but returns have not followed. Having AI and getting something out of it are two different things, and the second is not achieved by buying better or buying sooner.

Yet the same survey shows that diffusion does not produce results: only one organization in three has moved past the pilot stage, and few report more than 5% impact on operating results. In other words, most companies have bought something without yet being able to say what that purchase delivers. It is the symptom of a purchase that came before the decision, not the other way around.

Keeping this distinction in mind governs everything that follows. An AI tool produces nothing by its mere presence. It produces a result when it is connected to a precise problem, fed by reliable data, and judged against a measurable criterion. Until these three conditions are decided, no tool, however good, has any grip on the company. Three decisions therefore precede any purchase: what problem, what data, what measure of success. They belong to the leader, and to no one else.

What Problem, Precisely

The first decision is the most misunderstood, because it seems obvious. A leader believes they know what AI is for: improving customer service, saving time, selling better. But these are not problems. They are directions. A problem, in the sense that AI can address it, is narrower: not "customer service," but "the first-response time for routine requests"; not "saving time," but "the time spent writing a report." A problem that is named, measurable, and bounded.

Precision is not a refinement of method. It is what makes the tool evaluable. Deployments that pay off all start from a task identified with exactness. Klarna's customer service assistant was set on a precise task—handling repetitive inbound requests—and now does the work of about 700 agents. At Qonto, the Moshi agent absorbs 60% of routine requests from its 600,000 micro-business and SMB clients. These results do not come from the power of the model, but from the fact that the target task was named before it was tooled.

The contrast is instructive. "Doing AI in R&D" would have led nowhere; "producing the first draft of a clinical report" produced a result that can be measured. The rule holds for a thirty-person company as much as for a large group: the tighter the problem, the more obvious the tool that serves it becomes, and the more visible the result becomes. A vague problem calls for a vague tool, and a vague tool is never judged.

Naming the problem is also the best defense against a vendor. A leader who walks into a demo with a precise task in mind is not dazzled by what the tool can do in general; they check whether it can do that specific thing, on their own data. The problem comes first: the tool is chosen after, and chooses itself.

What Data, Who Answers

The second decision is the one companies put off and later regret. An AI project is only as good as the data that feeds it. A perfectly named problem starved of actionable data remains an intention. Before any purchase, two questions must therefore be settled: what data is needed, and who answers for it—who guarantees that it exists, that it is reliable, that it is available, and that it can be used.

This is where the cost the sales demo never shows is hidden. The tool installs in an hour; the data that makes it useful takes weeks to prepare. Gathering the history, cleaning it, deciding what does or does not leave the company, making sure the source is up to date: this work precedes the result and appears on no subscription invoice. Neglecting it is buying an engine without fuel. Actionable data is a leader's decision, not a technical detail left for later.

The lesson transfers without any change of scale. An SMB does not need 200 cases or a large group's data foundation. For its one first case, it needs to answer honestly the question: does the data that would feed this tool exist, is it clean, who owns it? If the answer is unclear, the purchase is premature. Deciding on data sometimes means discovering that the chosen problem is not yet ready—and that discovery, made before the purchase, costs nothing; made after, it costs the entire subscription.

What Measure of Success

The third decision is the one almost always skipped, because it requires a commitment. How will anyone know the tool worked? The success criterion must be quantified and set before starting: a timeframe, a rate, a number of hours, paired with a date at which it will be reviewed. Without it, judgment will be based on impression, and impression, faced with a new and appealing tool, always leans toward indulgence.

Measurement is not fussy oversight; it is what distinguishes an investment from an expense. The return on generative AI exists and is documented: an average of $3.7 for every $1 invested, and up to $10.3 at the most advanced organizations. The gap is not about technology, which is identical for everyone; it comes down to the discipline of those who set, before paying, what they expected and by when.

A good measure is concrete and singular. In a trial involving 20,000 UK public-sector employees, an assistant saved an average of 26 minutes per person per day, nearly two weeks a year: a simple, verifiable figure, decided in advance. At Flowserve, the target was the rate of quotes delivered on time; it rose from about 85% to 95%. In both cases, the question "did it work?" had a quantified answer because it had been asked before the first dollar was spent.

A number set in advance, or judgment will fall back on impression—and impression forgives everything.

With these three decisions made, the purchase becomes simple. The search is for the tool that serves this problem, on this data, toward this result, by this date. The tool becomes again what it should always have been: a means, chosen last, once the end has been decided. The sales demo, for its part, stops being a temptation and becomes a simple test: can it do this, or not?

Where AI Falls Short

Setting a measure requires knowing what the tool can do—and what it cannot. This is the clear-eyed side of the third decision. An AI tool excels at certain tasks and degrades on others, often without warning, and a leader who ignores this will set criteria out of reach or, worse, take false results at face value.

The facts are documented. A large firm tested generative AI with 411 participants: first drafts were markedly faster—up to 80% of time saved on certain documents—but average accuracy remained below that of a human, 3 out of 5 versus 3.5. The tool speeds up the draft; it does not replace final judgment. A study of 758 consultants confirms this: within the zone where AI performs well, the gains are real—speed up about 25%, quality up 40%; outside that zone, performance drops by 19%.

Ignoring this frontier costs money, and fast. A firm delivered a government report billed at 290,000 Australian dollars containing AI-fabricated citations, and had to issue a partial refund. The fault was not in the model: it was in the absence of measurement and review before delivery. A serious success criterion always includes human review, precisely because the tool, on the wrong ground, produces a false answer with the same confidence as a correct one. Knowing where AI helps and where it hurts is part of the third decision, not a later step.

Build or Buy

Once the three decisions are made, a question arises, almost always too early: should a market tool be bought, or a custom one built? Asked before the problem, the data, and the measure are settled, it has no answer. Building requires knowing exactly what to build; and that is precisely what the three decisions establish. A premature build-or-buy question is a disguised way of tackling the purchase from the other end: an acquisition mode is chosen before what is being acquired has been chosen.

The major players that build do so afterward, never before. EY launched its internal platform after a $1.4 billion investment, with its own model; McKinsey built an assistant backed by more than 100,000 documents, used by 72% of the firm. These choices follow a portfolio of problems already framed and a data foundation already under control. They are not a starting point; they are an outcome. For an SMB or a mid-market company, imitating that starting point would mean funding the fuel before deciding on the destination.

Most often, buying an existing tool is the right answer—provided it is the last decision, not the first. The accountants at KPMG Nord migrated 650 clients to an AI-equipped market software, generally in two to three months, with 20% to 30% of time saved on data entry depending on the file (editor estimate). The tool existed, the problem was named, the data was there: the purchase followed, and it paid off. The lesson is not "buy" versus "build," but "after" versus "before." Build or buy is an execution trade-off that comes once the three decisions are closed—never in their place.

Three Decisions, One Page

The three decisions require neither budget nor technical expertise. They require being written down, in black and white, before the next sales demo. This table serves as a decision grid: on the left, what the leader settles; on the right, how they will verify that the decision holds.

The strength of this page lies in its order. The problem comes first, because it governs the rest: no data without knowing what for, no measure without knowing what to measure. The data comes next, because it says whether the problem is workable today or only desirable. The measure comes last, because it turns an intention into a commitment. And the tool does not appear in the table at all: it shows up only once the three lines are filled in.

A motivated company quickly generates ten AI ideas. The page keeps only one at a time, the one whose three lines are filled in honestly. Writing down what will not be done this year matters as much as the line that is kept: scattered effort drains more AI budgets than the technology does. Completed French projects confirm this—twenty cases carried through to completion at SMBs and mid-market companies, with sometimes considerable gains, all starting from a single, framed problem, not a large-scale purchase.

Before the Next Demo

Installing this discipline requires neither hiring nor spending. It requires a page and an order held to. Here is the method a leader at an SMB or a mid-market company can apply before their next sales demo—and then hold up against any persuasive vendor.

1. Write the problem. A single task, in one sentence: what gain is targeted, expressed in hours, a timeframe, or a rate. If the sentence stays vague, the problem is not ready.
2. Answer for the data. Which is needed, who guarantees its reliability, is it available and clean today. Vagueness here is enough to postpone the purchase.
3. Set the measure. The number decided in advance that will say whether it worked, the date it will be reviewed, and human review before any delivery.
4. Choose the tool last. Compare offers against the page alone: can it handle this problem, on this data, toward this result? The rest is decoration.

This page turns an impulse purchase into a company decision. It does not replace the leader's judgment: it makes that judgment enforceable. Facing an AI vendor, it provides the only thing that matters—not an opinion about the tool, but a grid for judging what it will actually deliver. Technology will evolve quickly; these three decisions, by contrast, will stay stable, because they rest on the company's problem and not on the tool of the moment.

The message fits in one sentence. AI is not what gets bought: what gets bought is the resolution of a named problem, on owned data, toward a measured result. Settling these three points before paying does not guarantee success, but ignoring them guarantees spending with no return. It is the only order that protects the money—and the only one the leader cannot delegate to anyone.