A firm offers you an AI consulting engagement: three days of workshops, a forty-page report, a five-figure invoice. You sign. Six months later the report is in a drawer and your teams work exactly as they did before. The problem is not AI, and it is not always the consultant. It is the scope of the engagement, often copied from a large-group template when your small or mid-sized company needs one precise decision, quickly. Here is what an AI consultant really does, what they do not do, and the situations where their work saves you months instead of costing you time.
Key Takeaways
- An AI consultant prepares decisions: which use cases to pursue, with which data, in what order and on what budget.
- They generally do not build the solution. That step belongs to an integrator, an editor or your technical team.
- Their value shows in the documents delivered, not the number of slides: process map, scoping note, specification, rollout plan.
- They are useful when several departments ask for AI with no clear priority, or when nobody internally can arbitrate between business and technical needs.
- According to RAND (2024), more than 80 percent of AI projects fail, twice the failure rate of IT projects without AI. Scoping addresses the first of the five root causes the authors identified.
What an AI consultant actually does
The job title covers very different backgrounds: former data scientist, organisation consultant, ex-digital project manager, pre-sales engineer at a software editor. The quality of an engagement depends less on the CV than on the activities actually carried out in the first weeks. There are five.
Understanding the real work, not the org chart
A serious engagement starts with interviews of the people who do the work: the assistant re-keying orders, the salesperson following up by hand, the accountant checking the same lines every month. The consultant looks at where time is lost and which decisions are taken without data. At this stage a company often discovers that its problem is not the one it thought: the quote that takes three hours comes from a badly configured tool, not from a lack of artificial intelligence. An honest consultant says so, even when it shrinks the engagement they could have sold.
Prioritising use cases
Fifteen ideas do not make a roadmap. The consultant reduces the list to three or four workstreams ranked by expected gain, data availability, risk and the workload the team can absorb. This is the most useful arbitration of the engagement, and the hardest to hold alone as a business owner: every department defends its project with good arguments. The consultant's role is to bring an outside criterion, not to decide in your place.
Turning the need into a specification
A need stated in one sentence produces five incomparable quotes. The specification states what the solution must do, with which data, under which acceptance criteria, what is out of scope and how success will be measured. Without it, each supplier prices a different project, and you compare prices that do not cover the same work.
Setting data and compliance rules
Where the data lives, who can access it, how long it is kept, what happens when an employee pastes a client contract into a consumer tool, what obligations the European AI regulation creates for your use. These questions are not settled after go-live. A consultant raises them beforehand, with your legal counsel or data protection officer where one exists.
Preparing teams and measurement
A tool adopted by three people out of thirty produces no return. The consultant identifies key users, organises their onboarding, defines the indicator that will show whether the project delivers, and the date at which that number gets reviewed. This is the part technical suppliers handle worst, because it happens after delivery.
The need is real. According to Bpifrance Le Lab (2025), 31 percent of French micro, small and medium businesses say they use generative AI, but only 8 percent on a regular basis. And more than two out of three owners who do not use it say they struggle to identify use cases for their business. The gap between trial and embedded use is precisely where a consultant works.
What an AI consultant does not do
Misunderstandings are expensive, in both directions.
- They do not build the solution. Unless explicitly sold as a hybrid profile, they write specifications, not code. Building belongs to an integrator, an editor or your own team.
- They do not guarantee results. Nobody can promise that a model will hold its performance on your data before seeing that data. A firm commitment on gains, signed before any diagnosis, is a warning sign.
- They do not decide in your place. They inform the arbitration, they do not choose which department gets served first.
- They do not replace training. A scoping note makes nobody autonomous. Training your teams is a separate workstream, with its own deliverables and budget.
- They do not provide support. Once the solution is live, monitoring, fixes and changes belong to a maintenance contract, not a consulting engagement.
Deliverables that separate scoping from talk
A consultant who speaks well is not a consultant who produces. The difference shows in the documents handed over and in the questions they accept to receive. This table works as a reading grid before you sign.
| Expected deliverable | Question to ask | Warning sign |
|---|---|---|
| Process map | Which processes did you spend time on with operational staff? | A list of tools to buy |
| Prioritisation note | Why this use case before the others? | A ranking with no criterion and no numbers |
| Specification | Who signs off the acceptance criteria? | A document with no measure of success |
| Rollout plan | Who does what after you leave? | An oral debrief, no written document |
| Data register | Which data, held by whom, kept for how long? | "We will see later" |
If two rows of that table are still empty at the end of the engagement, the scoping did not happen.
When to call in an AI consultant
Four situations justify an engagement, and they are fairly easy to recognise.
Several departments ask for AI at once
Sales, operations, HR, finance: each has an idea and an emergency. The owner then arbitrates on volume of demand, not on value. An outside view puts projects on the same scale, with comparable criteria.
You do not know whether the problem is an AI problem
This is the most common case. A slow process may come from unclear organisation, an unsuitable tool or a genuine automation opportunity. Paying for a diagnosis that concludes there is no AI project to fund is a good investment: it avoids six months of effort and a wasted budget.
The budget is ready but the scope is vague
Committing a supplier to a poorly defined need is the fastest way to lose money. Scoping turns an intention into a comparable order, with a written scope and acceptance criteria.
The internal team lacks a technical referee
An IT manager alone against an editor selling an AI solution does not always have the arguments. The consultant brings an independent reading of the sales pitch: feasibility, hidden costs, reversibility, dependency on a single vendor.
In all four cases, the standard engagement falls under AI audit and advisory: a few weeks, interviews, a scoping note and a roadmap you can execute with the supplier of your choice.
When you can do without one
Conversely, three situations make consulting unnecessary.
- The need is already scoped and the tool chosen. What remains is integration, training and monitoring: an integrator or an editor is enough, and consulting adds a middleman with no new decision to produce.
- The project fits in a single department and a single tool. Training focused on that tool costs less than a general diagnosis and produces a visible effect faster.
- Your team already has a senior data or IT profile available to arbitrate. They will do the work internally, with the advantage of knowing the company and its constraints.
When in doubt, start with one day of scoping shared with your teams: it costs little and immediately tells you whether a larger engagement is justified.
How to pay without paying for nothing
Two models coexist: the day rate and the fixed price per deliverable. The first is comfortable for the supplier, the second protects you. A consultant billed by the day has no economic interest in finishing early. This is not a question of honesty, it is a question of incentive. Ask for a price per deliverable, with an indicative number of days to gauge the effort.
Three clauses deserve your attention before signing: the delivery date of each deliverable, the option to stop after the first phase, and ownership of the documents produced. The questions to ask resemble those for choosing an AI agency, and the warning signs of an unreliable supplier apply to a freelancer too.
Avoid the twelve-month accompaniment signed before the first deliverable. A consulting engagement is judged on what it produces in three weeks, not on the quality of the sales proposal.
FAQ
What is the difference between an AI consultant and an AI agency?
A consultant sells a decision, an agency sells a delivery. The first scopes, prioritises and specifies; the second builds, deploys and maintains. Some firms do both, which creates a quiet conflict of interest: whoever advises has an interest in recommending a project they will then deliver. Ask explicitly who builds, and at what price.
Can you hire a consultant without a defined project?
Yes, and it is often one of the best moments. The engagement then consists of finding use cases and deciding whether there is material. The deliverable can be a documented negative conclusion: some processes do not need AI, and a company that learns this early saves several months.
Can a consultant work with our current IT supplier?
It is often the most effective setup. The consultant brings the need and the acceptance criteria, your supplier brings knowledge of the system. Agree on a joint checkpoint from the start and a written split of responsibilities, so that neither side waits for the other.
How long does a scoping engagement last?
It depends on the number of processes examined, not on company size. An engagement limited to one process counts in weeks; a map covering several departments takes longer, especially when data is scattered. Ask for a calendar with deliverable dates, not a duration in working days.
How do you check that a consultant is competent?
Ask for two reachable references and one recent piece of work, anonymised if needed: a scoping note, a specification. A serious consultant agrees to show a document, not just logos. Also check that they can say no: an interview where all your ideas are excellent announces an engagement that will confirm your intuitions without testing them.
Conclusion
An AI consultant does not make your company intelligent. They turn a vague intention into decisions: which process, which data, which budget, in what order, and who does what next. According to RAND (2024), more than 80 percent of AI projects fail, and a survey cited by the authors found that only 14 percent of organisations reported being fully ready to adopt AI. The causes they document are organisational before they are technical: a misunderstood problem, a model optimised for the wrong metric, insufficient data, technology placed ahead of the problem to solve. That is exactly what scoping corrects, provided it is bounded and delivered.
Before signing, ask three questions: which document will I receive, on what date, and what happens if the conclusion is that I do not need AI. If the answers are vague, the engagement will be too. If you want to review your situation before choosing a supplier, contact NexeAI.
