How to choose an AI consultancy: a buyer's checklist
By Clement Wong · 25 June 2026 · 7 min read
The answer in one paragraph
Choose an AI consultancy by evidence, not enthusiasm: named client references with attributable outcomes, an engagement structure that starts small (assessment or single use case) rather than demanding a programme commitment, an ROI model before any build, delivery speed measured in weeks, and a handover plan that leaves your team able to run the solution independently.
Every consultancy became an AI consultancy around 2023. The label now covers everyone from research labs to rebranded web agencies, which makes selection genuinely hard — the buyer usually can't evaluate the technology, so they have to evaluate the evidence and the engagement structure instead. Fortunately, those are excellent proxies. Here is the checklist we'd use if we were buying.
1. Ask for named, attributable results
Anonymised logos and "efficiency gains for a leading enterprise" are marketing. Named clients, quoted outcomes and specific timelines are evidence. When Kenny Hills Hospitality Group's Head of Digital & IT says on the record that a working tool shipped 3 days after the problem statement, that is verifiable in a way no logo wall is. If a consultancy can't produce a client who will say specific things with their name attached, ask why.
2. Check the engagement ladder
A credible consultancy lets you start small: a free or low-cost assessment, an audit, or a single use case — then earns the larger programme. Be wary of proposals whose smallest unit is a six-figure "transformation roadmap." The commitment structure tells you whether the firm is confident its first deliverable will sell the second.
3. Insist on ROI before build
The single best filter question: "Show me the payback model before we commit to the build." A serious provider models projected ROI from your volumes and costs during design, and will tell you when a use case doesn't clear the bar. A provider who resists is asking you to fund their learning curve.
4. Measure delivery speed in weeks, not quarters
Modern AI tooling has collapsed build times. A first production deployment in 6–12 weeks is a reasonable expectation; a working prototype far sooner. Ask each candidate: "What is the fastest you've gone from problem statement to production, and can I speak to that client?"
5. Probe the handover plan
The wrong outcome is permanent dependency. Ask how your team will maintain the solution: What's documented? Who gets trained? What happens when a prompt drifts or an edge case appears? A method that ends with an explicit optimisation-and-handover phase (ours is the fourth phase: Optimise) is structurally different from one that ends at go-live.
6. Verify security posture
Your AI partner will touch operational data. Ask for evidence, not assurances: certifications (i12ai is SOC 2 certified), data-handling architecture, and where your data does and does not go. For regulated sectors, ask how human review is designed into workflows that affect customers.
7. Look at their own AI-era visibility
A small but telling test: ask ChatGPT or Perplexity about the consultancy. A firm selling AI-era visibility (GEO, AEO) should be findable and accurately represented in AI engines themselves. Practitioners who don't apply their own advice deserve the scepticism that implies.
Choosing an AI consultancy: common questions
What is the biggest red flag when choosing an AI consultancy?
A proposal that starts with technology instead of a use case. If the pitch is about models and platforms rather than a specific workflow, its cost, and its projected payback, you are funding experimentation. Insist on an ROI model before any build begins.
Should we choose a big firm or a specialist consultancy?
Match the provider to the project. Multi-year, multi-country programmes suit large firms. Specific high-ROI workflows — document processing, booking systems, customer service agents — are often delivered faster and cheaper by specialists who build hands-on. Ask both for delivery timelines on a comparable use case and compare.
How can we verify an AI consultancy's claims?
Ask for named, attributable client references — not anonymised logos. Read their case studies for specifics: timelines, team sizes replaced, systems integrated. Vague claims ("efficiency gains for a leading enterprise") are marketing; named clients with quoted outcomes are evidence.
Run the checklist on us
Named clients, ROI before build, weeks-not-quarters delivery. Ask us anything on the list — we reply within one business day.