How to Choose an AI Consulting Firm: 5 Questions to Ask Before You Sign

Hiring AI help is easy. Hiring the right AI help is not.

The market for AI consulting has roughly tripled since 2023, according to Grand View Research. That growth brought talent. It also brought noise. Every software agency added “AI” to their website. Every strategy firm launched an AI practice. And every freelancer with a GitHub repo started calling themselves a consultant.

For a mid-market company trying to make a smart hire, the landscape is genuinely difficult to navigate. Proposals look similar. Everyone claims end-to-end capabilities. The jargon is dense enough to obscure real differences. And by the time you realise you picked the wrong partner, you’re three months and five figures deep.

We hear this constantly from companies that come to us after a failed first engagement. The story is always some version of: “They seemed credible in the pitch. The proposal was polished. But once work started, it became clear they didn’t really understand our business, or AI implementation.”

The fix is straightforward. Ask better questions before you sign. Not more questions. Better ones. Five, specifically, that cut through the pitch and reveal what you’re actually buying.

Five questions that separate real capability from good marketing

Question 1: “Can you show me a system you’ve built that’s still running in production?”

This is the single most revealing question you can ask. Not a demo. Not a proof of concept. Not a pilot that ran for two months. A system that is currently live, processing real data, in a real business. Many firms have impressive case studies that describe projects they started but never finished, or pilots that were deemed “successful” but never deployed. The gap between a working demo and a production system is enormous - it’s where most of the hard engineering lives, and what the NIST AI Risk Management Framework formalises as the difference between development-time controls and deployment-time controls. If a firm can’t point to something running in production, they haven’t crossed that gap.

Question 2: “Who exactly will work on our project, and what have they built before?”

In consulting, the people who sell are rarely the people who deliver. This is true at firms of all sizes, but it’s especially common in the big consultancies. You get a partner in the pitch meeting and a team of recent graduates doing the work. Ask to meet the actual team. Ask for their LinkedIn profiles or portfolios. A firm that hesitates here is a firm that plans to staff your project after you sign, not before. And staffing after signing means they’re assembling a team for your project, not deploying a team that already knows how to work together.

Question 3: “What does your discovery process look like before you propose a solution?”

Any firm that sends a detailed technical proposal after a single meeting is either guessing or selling a pre-built solution. Real AI consulting requires understanding your operations, data landscape, team capabilities, and business objectives before recommending anything. A credible firm will describe a structured discovery phase (interviews, process mapping, data assessment) that happens before the technical work begins. If the answer is “we’ll build you a chatbot” before they’ve seen how your team works, walk away.

Question 4: “How do you define and measure success for this engagement?”

Vague answers here are a major red flag. “Improved efficiency.” “Better customer experience.” “AI-enabled operations.” These phrases sound good in a proposal and mean nothing in practice. Push for specifics. A good firm will tie the engagement to measurable business outcomes: hours saved per week, error rate reduction, cost per transaction, response time improvements. They should be comfortable putting those numbers in the contract, or at minimum, defining them during the discovery phase before committing to a build.

Question 5: “What happens after you leave?”

This question reveals whether the firm is building for your long-term success or for their ongoing revenue. Some consultancies deliberately create dependency: proprietary platforms, undocumented code, systems that only their team can maintain. A credible firm builds systems your internal team can operate, documents everything, trains your staff, and defines clear criteria for when the handoff is complete. They should also be honest about what you’ll need internally to maintain the system, even if that means telling you to hire someone.

Running your own evaluation process

Beyond the five questions, there’s a practical process for evaluating AI consulting firms that most procurement teams overlook. It doesn’t require technical expertise. It requires structure.

Talk to three firms minimum. Not to get competitive pricing - to calibrate. After three conversations, you’ll have a much clearer sense of what’s standard, what’s exceptional, and what’s smoke. The differences in how firms approach the same brief will tell you more than any proposal document.

Ask for client references and actually call them. Not the references the firm provides. Those are cherry-picked. Ask for the name of a recent client, then ask that client what went wrong. Every project has friction. The question is whether the firm handled it well. A firm that refuses to provide references, or only offers anonymised case studies, is hiding something.

Pay attention to what they ask you. A good consultant asks hard questions early. About your data quality. About your internal resources. About what you’ve tried before. About what your actual budget is, not your aspirational one. If the firm is mostly talking about their capabilities and barely asking about your situation, they’re selling, not diagnosing.

• Compare at least three firms on the same brief to calibrate quality
• Request and call client references, then ask specifically what went wrong
• Evaluate the questions they ask you, not just the answers they give
• Insist on meeting the delivery team, not just the sales team
• Look for measurable success criteria in the proposal, not vague outcomes

One more practical note: the EU AI Act creates new obligations for companies deploying AI systems, particularly around transparency, human oversight, and risk assessment. A firm that understands these requirements and factors them into their delivery process is a firm that thinks about production, not just pilots. If regulatory compliance never comes up in the conversation, the firm is either unaware or unconcerned - neither is good.

The audit-first shape we run at gamgi sits on the process page and is the structural answer to most of the five questions above. The diagnostic version of the same conversation, whether you actually need an AI audit, is worth reading before you brief any firm. A structured audit before signing reframes the selection problem against your actual operational bottleneck rather than the most marketable use case.

When the five questions aren’t the right filter

The five questions assume you’re hiring a partner to build a custom production system against an unsolved internal problem. Several situations don’t fit that shape, and forcing the questions through them produces the wrong shortlist.

• You’re hiring for a one-off research deliverable. A model benchmark, a feasibility study, or a market scan against your data isn’t a production engagement. Question 1 (production systems) and Question 5 (handover) don’t apply because nothing is being deployed. The right shortlist is closer to a boutique research firm than an implementation partner.
• The decision is already “buy a SaaS tool” and you need integration help. If the AI capability lives inside an off-the-shelf product you’ve already picked, the firm you need is a systems integrator with that product’s certification, not a custom-build consultancy. Question 3 (discovery process) collapses to vendor-product knowledge.
• Your internal team is the build team and you need senior advisory only. Some buyers have the engineering capacity and need an architect or a part-time AI lead, not a delivery firm. The five questions are written for full-delivery vendors. For advisory engagements, the diagnostic is depth of operator experience, not production-portfolio breadth.
• The budget can’t support a real engagement. Below roughly €30K of true scope, most credible firms decline rather than under-resource the work. Pushing the five questions at a budget the market can’t deliver against produces a shortlist of firms willing to lie about scope. The fix is to halve the problem, not the budget.