How to Choose an AI Consultant

Plenty of businesses do not need one, and any consultant unwilling to say that is telling you something about their sales targets. The honest test has three parts. If your business is small — say under 20 staff — the answer is usually no: the opportunity is one or two off-the-shelf tools used properly, and the consulting fee would exceed the value on offer. If you have someone internally who is genuinely curious, has time, and understands your processes deeply, they will often out-perform an external adviser on the first project because process knowledge beats technology knowledge nearly every time. Where a consultant earns their fee is a specific situation: enough scale that the opportunity is real, enough complexity that the sequencing is not obvious, systems that need to be integrated rather than replaced, and nobody internally with the bandwidth to diagnose properly. If that is not you, save the money. If you are unsure, a free consultation should resolve it — and if the consultant cannot tell you honestly which of the three you are, they cannot diagnose a business either.

Watch the direction of the conversation in the first meeting. A good consultant spends most of it asking about your business — how work actually flows, where things break, what you have already tried and why it failed, what your data looks like, who would have to change their day for this to work. A salesperson spends most of it telling you about capabilities, case studies and a methodology with a name. The specific questions worth asking them: what would make you tell me not to do this project? Who exactly will be doing the work? What does your engagement produce that I keep if I never speak to you again? What have you deployed that failed, and what happened? Where does my data go and what does your vendor agreement actually say? A consultant with real production experience answers the failure question comfortably, because they have plenty of material. One who has only ever produced strategies will find that question difficult, and their discomfort is the answer.

Ask about failure, and ask about operations. Anyone can describe a success at a level of abstraction that hides whether they were there. Ask instead: what broke in the first fortnight after go-live? What did you have to redesign once real users touched it? How did you handle the model provider changing behaviour underneath you? What does your handover look like and who maintains it after? People who have actually shipped AI into production have vivid, specific, slightly annoyed answers to all of those, because those events are memorable. People who have produced recommendations and left will answer in generalities, or redirect to methodology. It is not a trick question and it is not adversarial — it is simply the fastest way to establish which of two quite different professions you are talking to, and both professions are legitimate. You just need to know which one you are buying.

Match the supplier to the size of your problem, not to the size of your ambition. Large firms are built for genuinely large organisations where the coordination problem across thousands of staff and dozens of business units is itself the hard part — their cost base exists to solve that, and for a 40-person business you are paying for a capability you cannot use. Independent specialists and small firms fit the mid-market well: enough capability to build, small enough that the person who diagnosed it is the person who builds it, and priced accordingly. Freelancers can be excellent value for a narrowly defined build where you already know what you want, and are risky where diagnosis is needed, because a single person with a hammer has an understandable interest in your problem being a nail. The failure mode we see most among mid-sized Australian businesses is buying enterprise consulting for a mid-market problem, then being disappointed that it did not fit — which was predictable from the outset.

Small, fixed-price, with a deliverable you keep and a genuine option to stop. That structure does three things at once: it caps your downside while you are still assessing whether these people are any good, it forces the consultant to scope precisely rather than sell a relationship, and it gives you an artefact — a written assessment, a map, a ranked shortlist — that has standalone value even if you never engage them again. What to avoid on a first engagement: open-ended time-and-materials before you know how they work, a long-term retainer signed before anything has been delivered, and any structure where stopping after the first phase leaves you with nothing. The test is simple. Ask what you keep if you stop after the first engagement. If the answer is "our ongoing relationship", you are being sold a subscription, not a diagnosis. If the answer is a document that lets you brief a different firm entirely, that is a consultant confident in their work.

Six, and any one of them should slow you down. First: agreement with your idea before diagnosis — if they endorse the AI project you walked in describing without examining your data or processes, they are selling. Second: nobody mentioned your data, which means either they have never done this or they intend to discover the problem on your budget. Third: no named answer to who is doing the work. Fourth: a proposal that ends at a recommendation while implying you are buying a working system — those are different products and the ambiguity is usually deliberate. Fifth: evasiveness on data handling, model training and where your information goes; "enterprise-grade security" is a slogan, not an answer. Sixth: no scenario in which they would tell you not to proceed. A consultant with no version of "do not do this" has no diagnostic function at all, which means you are buying implementation labour and should price it accordingly.