Everyone advising on AI has something to sell, including us. The Five-Test Screen checks incentive, evidence, alternatives, scope-out and reversibility, so a board can tell advice from a pitch.
Every organisation currently receiving advice about AI is receiving it from someone with something to sell. Vendors recommend their products. Training companies recommend training. Implementers recommend implementation. Most of this is delivered sincerely, and most of it contains real expertise. That is precisely what makes it hard to judge.
We are a consulting firm, so this article applies to us as well. If you cannot tell an advisory recommendation from a commercial one, you are at the mercy of whoever presents most fluently. The remedy is a short screen that any executive director, finance lead or board member can apply without technical knowledge.
Good advice can be checked. A sales pitch usually asks to be believed.
We suggest five tests. None of them requires you to understand the technology. Each asks about the recommendation's structure, not its content.
Test one: the incentive. What does the adviser gain if you accept? A recommendation to buy the adviser's own product or service is not automatically wrong, but the incentive should be stated by them, not discovered by you.
Test two: the evidence. What is the recommendation based on, and can you inspect it? Look for named sources, described methods and results that someone outside the adviser's organisation could reproduce. Claims that rest on "our experience" alone deserve a follow-up question about which experience, and how it was measured.
Test three: the alternatives. Did the adviser present options they would not profit from? A recommendation compared only with doing nothing is a sales argument. A recommendation compared with a cheaper, slower or in-house route is a decision aid.
Test four: the scope-out. Will the adviser tell you what you do not need? In our own work, we say so plainly when an organisation does not need us. Ask any adviser what they would leave out, and how they would know.
Test five: the reversibility. How costly is it to stop, and what would you keep if you did? Good advice leaves you with assets you own, such as documented methods, trained staff and written policy, and does not tie you to a supplier.
If an adviser cannot describe a scenario in which you should not buy from them, you are not being advised. You are being sold to.
Here is an illustrative pair of recommendations, not drawn from any real engagement.
An organisation is told, "You need an enterprise AI platform across all departments." The source is a platform vendor. The evidence offered is a case study from a much larger organisation. The alternative considered is to do nothing. There is no statement of who should not adopt. Exit involves migrating everything. On the screen, it fails four of five tests, even though the platform may well be excellent.
The same organisation is told, "Before buying anything, map where staff already use AI, decide which data may never enter a prompt, and train each department on its own tasks. Then decide whether a platform is justified." The adviser states their incentive, cites the method, offers an in-house path and would leave you with a written policy and trained staff either way. That advice may still be wrong. But it can be checked, and that is the difference that matters.
Three types of claim deserve extra caution, whoever makes them.
Buyers can shape the quality of advice they receive. Ask the adviser to disclose commercial relationships with any vendor they mention. Ask for the assumptions behind any figure. Request a short written statement of what the recommendation does not cover. And where the decision is material, invite a second opinion from someone with no stake in the outcome.
None of this is adversarial. Serious advisers welcome it, because it separates them from the noise. Firms that resist being tested have told you something useful.
A board does not need to judge AI. It needs to judge the advice about AI. That is a governance skill boards already have. The Five-Test Screen turns it into five questions that fit on the front of a paper, and can be applied to any proposal that reaches the table, from any source, including ours.
It would be inconsistent to publish these tests and exempt our own work. So here is how we answer them. Our incentive is that we sell assessment, policy, training and support for AI adoption, and we say so. We build no AI systems, so we have no platform to recommend for commercial reasons. Our evidence rests on named public studies and on methods we document so that others can inspect them. We state the alternatives, including doing the work in-house. We tell organisations when they do not need us. And we aim to leave behind assets they own: written policy, trained staff and a documented method.
You should still apply the screen to us. A firm that welcomes being tested is doing what it asks of others, and a firm that does not is worth a second look.
If you receive an AI recommendation this month, run it through a simple routine. Ask the adviser to write, on one page, the incentive, the evidence, the alternatives considered, the cases where the recommendation does not apply, and the exit cost. Score each answer as clear, partial or absent. Anything with more than one absent answer goes back for revision before it goes to the decision-maker.
The exercise takes perhaps half an hour and produces a document that a board can read in minutes. It also changes behaviour on the supply side. Advisers who know the page will be requested prepare differently, and the quality of the conversation rises.
Judge AI advice by its structure, not its fluency: state the incentive, show the evidence, compare real alternatives, name who should not buy, and keep exit cheap. Advice that passes those tests can be trusted more than advice that merely sounds sure.
If this is the question on your desk, a thirty-minute call tells you whether the service fits, or that you do not need us yet.