There is a stage in most AI projects that nobody puts on a slide. The pilot worked, everyone was pleased, and then the thing had to survive contact with the actual business, with its messy data, its nervous compliance team, and its staff who never asked for any of this. A surprising number of projects quietly die right there, in the gap between the proof of concept that impressed the board and the boring work of making it real. That gap is where the AI consulting market is currently making most of its money, and where a lot of it is being wasted.
Boards across the UK spent the last two years being told they were behind, and responded by writing cheques. What many discovered is that buying AI advice is not like buying other professional services: the field moved faster than the people selling into it, and a fair amount of the guidance on offer was repackaged general consulting, or a thin layer of strategy over somebody else’s model.
Why the pilots stall
The reason so many projects fail in the awkward middle is rarely the technology. The models are good enough for most business uses and getting better. What kills projects is everything around the model: data that is scattered across systems nobody has tidied in a decade, processes that were never written down, and a workforce that has not been told what any of this means for their jobs. None of that is glamorous, and none of it is what the flashier end of the market wants to talk about.
This is the unhelpful truth about deploying AI at scale: the hard part is organisational, not algorithmic. A model that hallucinates in a demo is a curiosity. A model that confidently gives wrong answers to your customer service team because nobody governed its inputs is a liability, and the difference is almost entirely down to the unglamorous groundwork. The firms that understand this spend the first chunk of any engagement on plumbing and governance, precisely the work that does not photograph well.
What the useful ones do differently
If you spend any time comparing how different advisers approach this, a pattern emerges. The ones worth hiring are oddly reluctant to start with the AI. They want to see your data first. They ask awkward questions about ownership and accountability before they will talk about use cases. They are comfortable telling you that your most exciting idea is the one most likely to fail, and that the boring internal process you mentioned in passing is where the real value sits.
Transparity, a UK firm that does this kind of work, makes a version of this point on its artificial intelligence consulting services page, framing the job less as bolting on clever tools and more as getting an organisation into a state where AI can actually be trusted to operate. Whether or not you ever speak to them, the framing is the right one, and it is a useful test to apply to anyone pitching you: do they start with your readiness, or with their product?
The contrast with the opportunist end of the market is stark. They lead with the model, the demo, and the promise of transformation by next quarter, because the easy part demos beautifully and the hard part does not. By the time the hard part shows up, the engagement is signed and the invoices are flowing, and the awkward middle becomes your problem rather than theirs.
See also: How AI and EdTech are Redefining Research Standards in US Universities
How to tell them apart
You do not need to be technical to separate the two. A few signals do most of the work. Ask a prospective adviser what they would refuse to do, and listen for whether they have any limits at all. Ask how they would measure success in twelve months, and be wary of answers that are all adoption metrics and no business outcomes. Ask what happens to the project when they leave, because a consultancy that builds dependency rather than capability has an incentive that does not align with yours.
The unfashionable conclusion
AI is going to matter enormously, which is exactly why the rush to buy advice about it deserves more scepticism, not less. The technology is the least of anyone’s worries. The organisations that come out of this period ahead will not be the ones that bought the cleverest model. They will be the ones that did the dull, patient work of getting their data, processes and people ready, usually with the help of an adviser honest enough to insist on it.
That is not the message most of the market is selling, because it is slow, unglamorous and hard to put a transformation logo on. It is, however, the one worth listening to. The awkward middle is unavoidable. The only real choice is whether you cross it with someone who warned you it was coming, or someone who pretended it was not there.
