Guide: How-to
How to Choose an AI Consultancy: What to Ask Before Signing Anything
NILG.AI · October 25, 2026
The short answer: most AI consultancies won't say no
When everyone claims to do AI, the question stops being who says it and starts being who can prove it. The difference between a good consultancy and a mediocre one rarely shows up in the proposal. It shows up in the second meeting, when they start talking technology before they understand your business.
This article is a practical guide. Concrete questions, real warning signs, and the reasoning behind each one.
The first warning sign: they arrive with solutions before listening to your problems
A serious consultancy spends most of the first conversation asking questions. They want to understand how the operation works, where the bottlenecks are, what data exists and what shape it's in. They don't show up with a Computer Vision pitch or a GPT chatbot before checking whether any of that actually makes sense in your context.
If they're already showing you product demos in the first meeting, that's a yellow flag. They may be selling a solution in search of a problem.
AI should only be applied where it makes business sense. Any partner that skips this validation is working for their own sales pipeline, not your outcomes.
The questions you should ask before signing
Do they have concrete cases in your sector?
Don't ask for generic references. Ask for a specific example: what problem was solved, how success was measured, and what happened when the project moved out of pilot and into production. That last part is where most consultancies fall short.
Many AI projects get stuck between prototype and real operations. A consultancy that has already navigated that transition knows exactly why things go wrong and how to avoid it. One without that experience will find out on your project.
Who will actually work on it?
The obvious answer is "our team." The question that matters is different: is it the team that closed the contract, or a junior team you've never met? In small firms and specialist boutiques, it's usually the former. In larger structures, that's not always the case.
Ask to meet the people who will be doing the work. Understand their actual experience level, not what's on the company's LinkedIn profile.
How do they measure project success?
If the answer is vague, that's a problem. A good consultancy defines business metrics before building any model. Reduction in process time, fewer errors, higher conversion. Real numbers, not technical metrics your CFO won't be able to interpret.
The ROI of an AI project needs to be readable by people who make business decisions, not just by the people building the model.
What happens when the model starts to degrade?
Machine learning models degrade. Data changes, customer behaviour changes, markets change. A consultancy that delivers and disappears is leaving you with an asset that will deteriorate without you noticing.
A serious partner talks about monitoring, proactive alerts, and continuous improvement. If this conversation never comes up during the selection process, you'll have to have it later, under worse conditions.
Warning signs worth taking seriously
They promise results without seeing your data. Any impact estimate made before a serious analysis of available data is fiction. It might be optimistic fiction or pessimistic fiction, but it's fiction either way.
They talk more about tools than processes. Technology is a means to an end. If the conversation revolves around their stack rather than how they'll solve your problem, the focus is in the wrong place.
They can't explain what they do without jargon. If they can't tell you, in plain language, why a particular approach makes sense for your situation, they'll struggle to bridge the gap between their technical team and your operational managers. And that bridge is critical for internal adoption.
They have no opinions. An experienced consultancy takes clear positions: this approach works, this one doesn't. If they agree with everything? Be careful. Either they lack the experience to push back, or they're telling you what you want to hear.
What separates a technical consultancy from a business partner
Some consultancies reach the end of a project with excellent models that nobody uses. It happens more than you'd think. The model works technically, but it wasn't designed with the organisation's real processes in mind. The team wasn't brought along. The model's output isn't integrated into the workflows people use every day.
A business partner treats adoption as part of the project, not as a problem the client sorts out afterwards. That means working with operational teams from the start, not just with the executive sponsor. It means designing interfaces and outputs that make sense for the people who will use the system daily.
The simplest test: ask how they ensure the project is adopted by the team, not just delivered to the client. The quality of that answer says a lot.
Social proof: what's worth something and what isn't
Reviews on platforms like Clutch or TechBehemoths are useful as a starting point, but read them with attention to detail. Reviews that describe the process, the communication, and concrete results are worth more than high ratings with no context.
Ask for contacts from previous clients with projects similar to yours. A 20-minute conversation with someone who went through the process tells you more than any proposal will.
And check whether the consultancy produces knowledge publicly: articles, case studies, technical content. Whoever shares knowledge consistently is demonstrating competence, not just claiming it. The NILG.AI blog is one example of that commitment to transparency, with articles ranging from healthcare applications to demand forecasting in construction.
A note on size
Bigger is not better in AI consulting. A specialist boutique with a solid portfolio in a specific sector will almost always outperform a large generalist that treats your project as another ticket in the pipeline. What matters is whether they have a structured methodology, experienced teams, and evidence that they take projects from pilot to production with real impact.
Projects like detecting errors in insurance claims or multi-product forecasting in the construction industry are the kind of evidence you're looking for: concrete cases, specific domains, measurable results.
Choosing the right partner doesn't eliminate the risk in an AI project. It does reduce it substantially. And the fastest way to spot the wrong partner is to pay close attention to what they say in the first two meetings, before any formal proposal reaches the table.