How to choose an AI implementation partner (and spot a strategy-deck consultancy)

Choose the partner whose senior person will personally build the thing, who insists on a baseline before they start, and who can describe their own exit. Brand, deck quality and team size are secondary. The four questions that matter are who does the work, how fast, what it costs, and whether it still runs after they've gone.

A disclosure before the table, because it would be worthless without one. I'm one of the options below, the senior solo practitioner, and I've tried to write the comparison honestly enough that it sometimes points away from me.

Why does the choice of partner matter more than the choice of tool?

Because the tools mostly work now, and the projects mostly still don't.

Gartner's September 2026 survey found that "only 22% of organizations have successfully scaled AI across multiple business units", and that roughly one in ten couldn't say what their function had spent on AI in 2025 (Gartner). McKinsey's 2026 survey put the share attributing any profit impact to AI at 37%, "about the same share as last year" (McKinsey). And the MIT study everyone quoted last summer, the one that found 95% of organisations getting zero return, had a finding that got less attention: pilots built with an external partner "were 2x as likely to reach full deployment as those built internally", on a small sample the authors were careful to caveat (MIT NANDA).

The reasons projects die aren't new. Gartner's January 2026 update listed them as "poor data quality, inadequate risk controls, escalating costs or unclear business value" (Gartner). Every one of those is a decision somebody made, or didn't, in the first fortnight, and the partner you choose is the person making those decisions with you. Pick the wrong type and you can end up with a beautifully argued reason to do nothing, or a proof of concept that never finds its way to a number.

What are the five types of provider, and what do you actually get from each?

Most buyers compare firms. It's more useful to compare shapes, because the shape predicts the outcome better than the logo does.

Why do so many AI engagements end as a slide deck?

Because the deck is the product the pyramid is built to make.

The economics are simple. A large firm has to keep a lot of junior people billable, and a strategy phase is the safest work to bill them on: interviews, workshops, a maturity assessment, a roadmap. Implementation is where you can be wrong in public. So the incentive is to sell the phase that can't fail, and to quote the phase that can as a separate engagement once the roadmap has been "aligned".

This model is under real pressure. Reporting on PwC's consulting overhaul noted that "the work that graduates and junior staff would previously have spent days compiling can be skipped" now that AI can do it (City AM). KPMG announced in the spring that it would cut more than 500 UK roles (Consultancy.uk). And in August, Consultancy.uk reported that all four of the largest firms had "potentially alienated customers" by publishing thought leadership "alleged to be majority-generated by AI" (Consultancy.uk). If the deck itself can be generated, what exactly is the deck-shaped engagement selling?

None of this means strategy is worthless. A business with forty candidate processes and no idea which to start with needs somebody to think, and I do that thinking on every engagement. But an SME with a manual process that costs it a morning a week doesn't need a forty-page roadmap. It needs the process automated, measured, and handed over. If the proposal in front of you ends with a roadmap, ask what the second proposal will cost before you sign the first.

What should you ask on the first call?

Seven questions. The answers matter less than whether they're answered without hesitation.

1. Who, by name, will build this, and will I speak to them every week? A good answer names a person and puts them on the call. A bad answer is "we'll assemble the right team".
2. What will you measure before you start? You're listening for a baseline: the volume, the time per item, the cost of getting it wrong. "We'll define success metrics in discovery" is fine. "We'll see what the data tells us" is not. Step two of the pilot playbook is the standard here.
3. What's the smallest thing you'd build first, and how long would it take? If the answer is measured in quarters, the scope is wrong.
4. What do you need from us? Data access, a process owner, a decision-maker who'll turn up. Someone who says "nothing, we'll handle it" hasn't done this before.
5. What happens at the end? Where does the workflow live, whose accounts run it, who has the keys, who fixes it in March? The right answer is "in your systems, under your logins, with your people trained to run it". Anything else is a dependency with a monthly invoice.
6. Which model or platform, and what happens when we want to change it? You want the answer in what vendor-agnostic AI means: the model in one place, the logic and data yours. Be wary of a partner whose answer is their own platform.
7. What would make you tell us not to do this? A partner who can't think of anything will sell you anything.

What are the red flags?

Most of them are visible in the proposal before you've spent a pound.

• Phase one is a roadmap and phase two is a separate quote. The commercial model has told you what it's optimised for.
• The people in the pitch aren't the people on the project. Ask directly. Watch for the pause.
• Everything runs in their tenancy. Their cloud account, their subscription, their API keys. The day you stop paying is the day it stops working.
• A partner tier with a platform vendor. A firm that is a certified partner of one CRM or one cloud is going to recommend that CRM or that cloud. That's not corruption, it's incentives, and you should price it in.
• They never say no. Every idea you raise is a great idea. Nobody with twenty years of scars agrees with everything.
• A proposal you couldn't hand to a rival to build from. If it's too vague for that, it's too vague to hold them to.

What, exactly, should the money buy?

What the money should buy is a working thing, measured against the number you agreed before it was built, that pays for itself within a year on its own figures. That last clause is the ratio that matters. A £15,000 engagement that removes a £50,000-a-year process is cheap. A £5,000 "AI workshop" that produces a document is expensive, whatever the day rate was. Size the first engagement so the outcome funds the second, and you'll never have to argue with a board about an AI budget again.

How should an accounting or professional services firm choose?

The same way, with two extra tests, because the sector has its own shape.

Accountancy is further into this than most. An ICAEW survey of mid-tier firms this spring found "86% have a technology strategy that includes AI adoption", but that AI "is being incrementally embedded into firms with most firms only claiming moderate use of the technology to date", and only 17% felt confident about its effect on their workforce (ICAEW). The barriers the Institute lists are "skills shortages, concerns about AI risks and the regulatory environment" (ICAEW). Strategy documents exist in abundance. Working automations are rarer.

The first extra test is data handling. Client confidentiality isn't optional in a regulated practice, so ask where the data goes, which models see it and under what terms, and whether the partner can explain the difference between a consumer AI tool's terms and an enterprise one's. A partner who waves this away is a partner who hasn't worked with a regulated firm.

The second is where the value lives. A professional services firm runs on a practice management system, a CRM, an inbox and a great many spreadsheets, and almost all the manual work happens in the gaps between them. The most instructive engagement I've done in the sector started with a B2B services firm whose vendor had just implemented an AI-enabled CRM, complete with lead scoring and automated outreach. Six account managers were then spending fifteen hours a week between them rekeying client metrics into it by hand. The vendor's implementation team had built for the platform's headline features, not for how the work actually flowed. A two-week pipeline from the systems that already held the data, using nothing the firm didn't already own, removed the rekeying entirely and was worth £50,700 a year. That is the partner-type problem in one story: the platform's implementers optimised for the platform. Ask any prospective partner what they'd have done differently on that project, and listen for whether they'd have started with the workflow or the tool.

The bottom line

The best AI implementation partner for a UK SME is whoever will personally build a working thing, measure it against a number you agreed first, put it in your systems under your control, and leave you able to run it. That describes a good senior practitioner, a good small agency and, occasionally, a good in-house hire. It rarely describes a pyramid, because the pyramid is built to sell the deck and quote the build. Use the eight questions, watch for the red flags in the proposal, and size the first engagement so it pays for the second.

If you'd like a straight answer on which shape fits your situation, including "you don't need any of us yet", a thirty-minute discovery call sorts it out. See AI consulting for how I run a fixed-price project and fractional AI leadership for the ongoing version.