The number of firms calling themselves AI consultancies doubled in 2025. Most of them are rebranded software integrators, digital agencies, or solo practitioners who added "AI" to their LinkedIn headline after attending a conference. Choosing the wrong one costs you six months, a failed pilot, and a board that is now sceptical of every AI proposal you bring.
Start with the problem, not the vendor
Before you speak to any consultancy, write one sentence describing the business outcome you want. Not the technology. Not the use case. The outcome. "We want to cut document review time by 60%" is an outcome. "We want to implement AI" is not a brief, it is an aspiration with no way to measure success.
This matters because the right consultancy for document processing is different from the right one for customer-facing agents, which is different again from the right one for data analytics infrastructure. A firm that is genuinely strong in one area will tell you when a project falls outside it. A firm that claims to be excellent at everything is telling you something important about how they operate.
Five criteria that actually matter
First: they should be able to show you production deployments, not proofs of concept. Ask specifically: "Can you show me a system you built that is running in a client's production environment today, processing real data, with real users?" A PoC in a sandbox environment proves nothing about delivery capability.
Second: they should be model-agnostic. If a consultancy recommends a specific AI model before understanding your data, your infrastructure, and your compliance requirements, they are selling you something. The right model for your use case depends on factors that take weeks to assess properly. Any firm that skips that step is shortcutting your project.
Third: their pricing should be fixed or clearly scoped. Time and materials billing for AI projects is almost always a sign that the firm does not have a clear delivery methodology. Good consultancies have built enough similar systems to know what it costs. Fixed fees mean they have skin in the game on scope and timeline.
Fourth: they should ask hard questions about your data before promising anything. AI systems are only as good as the data they run on. A consultancy that jumps straight to architecture without auditing your data quality, labelling, and access controls is either inexperienced or optimistic about problems they have not found yet.
Fifth: they should be honest about what AI cannot do in your specific context. The best signal that a firm knows what they are talking about is when they push back on part of your brief. That is not a sales problem, it is a quality signal.
Red flags to walk away from
Vague timelines. "This will take a few months" is not a project plan. Any firm worth engaging can give you a week-by-week delivery schedule before you sign.
Vendor affiliations they do not disclose. Some consultancies earn referral revenue from the tools they recommend. Ask directly: "Do you receive any commercial benefit from recommending specific vendors or platforms?" The answer tells you more than anything in their proposal.
Offshore delivery hidden behind a UK front. There is nothing wrong with distributed teams, but if the firm presents UK-based consultants in the sales process and then hands the project to a team you have never spoken to, your governance and communication expectations will not be met. Ask who will actually do the work.
Proposals that skip security and compliance. If an AI consultancy does not ask about your data classification, your ICO obligations, or your internal access controls in the first conversation, they have not worked with regulated clients before.
Questions to ask in the discovery call
Beyond the standard "tell me about your experience" questions, these four tend to reveal the most: What is the most common reason your projects get delayed, and how do you handle it? Can you walk me through a project that did not go to plan and what you did? What would make you walk away from a project mid-engagement? Who owns the IP for anything you build for us?
The answers to these are less important than whether the firm gives them without hesitation. A consultancy that has built and shipped real AI systems has dealt with all of these situations. One that has not will either deflect or give you a rehearsed answer that does not quite answer the question.
What good looks like
A good enterprise AI consultancy starts by auditing what you already have before recommending anything new. They give you a fixed price for a clearly scoped deliverable, not a retainer with a vague mandate. They hand you something that works in production, with documentation your team can maintain. And they tell you what to do after the engagement ends, not why you need to extend it.
At Kelriva, every engagement starts with an AI Readiness Assessment before we touch any infrastructure. We do not recommend specific models until we understand your data. Our pricing is fixed. And we are registered with the ICO because the clients we work with have compliance obligations that matter. If that matches what you are looking for, the conversation starts at kelriva.ai.