If your company ran an AI pilot in the last year or so, there’s a fair chance it’s now stuck in an odd kind of limbo. The demo worked. The steering group liked it. Then someone asked how it would connect to the CRM, who would own it once the project team moved on, and what it would cost at full volume, and the answers were vague enough that nobody signed off the next phase.
That’s usually when a business starts talking to outside AI firms. It’s a sensible move, but it’s also where money gets wasted, because the market is full of consultancies that are very good at the part you’ve already done.
Why pilots stall in the first place
Most stalled pilots don’t fail on the model. Getting a large language model to summarise contracts or draft replies to support tickets is the easy bit now, and your own developers can probably get a prototype running in a few weeks. The trouble starts when the prototype meets real conditions. Data that looked clean in a sample of a few hundred records turns out to be patchy across the full set. The security team wants to know where prompts and outputs are logged. Legal asks what happens if the tool gives a customer the wrong answer. The finance director wants to see running costs, not just licence costs.
None of these questions is exotic, but answering them takes skills that most internal teams have in pieces rather than in one place. Your data people know the data, your platform team knows the cloud tenancy and your product owner knows the process, but nobody has taken an AI system all the way to production before. That gap is what you’re really hiring for, and it’s why the useful end of the market tends to be specialist AI consultancies that work across strategy and delivery, rather than firms that only advise or only build.
Spotting the slide-deck consultancy
The slide-deck firm is easy to recognise once you know what to look for. The first meeting is about “AI maturity” and a roadmap with three horizons. You’ll hear plenty about their methodology and very little about your systems. When you ask who’ll actually do the work, the answer is a pool of consultants to be confirmed after signature.
A better first conversation feels more like they’re interviewing you. Good consultancies ask to see the pilot, the data behind it and the architecture it runs on. They want to know who owns the process being automated and whether that person has time to be involved. They might tell you, fairly early, that the use case you’ve picked isn’t worth taking further, or that a simpler approach without generative AI would do the job for less. That’s a good sign, even if it’s awkward to hear.
A few questions separate the two quickly. Ask what they’d need from you in the first month, and see whether the answer is specific (access to a data source, a named process owner, a test environment) or generic. Ask about a project that didn’t go to plan and what they changed because of it. Ask how they’d judge whether the thing works, and listen for real measures such as handling time or error rates rather than login counts. Then ask who writes the code, who looks after it after go-live, and what happens to it if you part ways.
Terms worth arguing about
The contract matters as much as the pitch. Push for a short, paid discovery phase with a defined output, whether that’s a working build against production data or a clear recommendation to stop. It costs less than a long engagement that drifts, and shows you how the firm works.
Be careful about who ends up owning what. If the consultancy builds on its own accelerators or proprietary frameworks, find out whether you’ll have the rights and the documentation to run and change the system without them. The same goes for prompts, evaluation sets and any fine-tuning data, which are easy to overlook. Handover should be in the plan from the start, with your people working alongside theirs, not bolted on in the final fortnight.
Watch platform incentives too. Many consultancies are closely tied to one cloud vendor, which is fine if you’re already committed to it. Still, ask directly whether they earn anything from the vendor when you buy more, and whether any of their recommendations would change if you ran a different stack.
The last thing to settle is what “done” means. For most AI systems the build isn’t the end of the cost. Models get updated or retired, and outputs need checking against real cases for months after launch. A consultancy that prices only the build is leaving you to discover the running costs yourself. The one that can tell you plainly how much monitoring and evaluation the system will need, and which person on your side should own it, has most likely shipped something like it before.