KPMG has just published a report called Make AI Scale, and buried under the survey graphics is a single, sharp idea worth every leader's attention: most organisations are managing AI as a change programme when they should be managing it like the building of a new production line.
The distinction matters because it explains the thing everyone is quietly worried about. Almost everyone is now using AI — nine in ten UK organisations are already working with agentic tools — yet only around one in twenty say they've reached established, repeatable return on it. Enormous activity. Very little of it yet running the business.
Three kinds of change, not one
The report's most useful move is to split "using AI" into three very different levels of change, defined by how much of an end-to-end process the AI actually takes on. They are not stages of the same project — they behave like different animals.
Faster drafts, fewer manual steps. Feels safe — and it's where most teams quietly declare victory.
Roles blur, the old metrics and controls stop fitting, and the value starts to get real — along with the friction.
You're building a new production line. AI becomes part of the operating model — think flow, throughput, controls and escalation, not a tool.
Most stalls happen because a business is running a manufacturing-level change with a redecoration-level mindset. The pilot worked, the demo impressed, and then the moment the AI reaches deep enough into a process to change who does what — everything the organisation is built on starts to resist.
Why this matters more if you're not a FTSE 100
KPMG's proof points are the businesses it works with: global insurers, large financial institutions, enterprises with data-science benches and transformation offices. Read quickly, you could conclude this is an enterprise problem.
It's the opposite. The mid-market firm, the broker, the regional operator — you face the identical stall, at the identical point, with none of the bench. You don't have a spare head of data science to pull onto the hardest bottleneck, or a governance function that can redesign itself. That's not a reason to wait. It's the reason to be deliberate: pick fewer battles, put your genuinely best person on the one that matters, and build it so it holds up. The thesis is the same at every size. The margin for wasted effort is not.
Before you scale anything, get an honest read on where the return is — and where it isn't. About three minutes, no sign-up.
Take the free AI analysis →Three things that unstick it
The report names three forces that stall AI as it moves into real work. They translate cleanly to a smaller business.
1. Fear — usually a clarity problem, not a technology one
Adoption is the most-cited barrier in the UK, and it rarely means people hate the tool. It means they're asking "what does this do to my job?" and not hearing a straight answer. The fix KPMG suggests is a time-saved contract: be explicit that time AI frees up gets reinvested in better work — deeper analysis, better customer outcomes, upskilling — not quietly banked as headcount reduction. Say what success looks like for the people doing the work, not just for the programme. It's a simple, honest commitment, and it's worth putting in writing.
2. Focus — put your best person on the hardest bottleneck
Above roughly half-automation, "improving how we work" quietly becomes "replacing how we work" — and you end up running the old way and the new way at once. That takes your best operators, the ones who know where the exceptions and the hidden controls live. In a smaller firm that's a real cost: the person you'd second to the AI work is the person holding today together. Doing it anyway, on one thing that matters, is the choice that separates the firms that scale from the firms that stay in pilot.
3. Friction — design it in, don't apologise for it
Governance, approvals and risk checks feel like the brakes. KPMG reframes them as the organisational immune system doing its job — and, importantly, most stalls dressed up as "we can't move faster without more risk" are really a governance design problem. The answer isn't to weaken controls; it's human-in-the-lead, not human-in-the-loop: put people at the decision points that need judgement, let AI do the heavy lifting in between, and monitor continuously rather than inspecting at the end. For a regulated smaller firm, that's the difference between AI you can defend and AI you have to keep switched off.
A short test before you scale
The report's whole argument reduces to a few questions worth asking before you push any AI initiative wider:
- Which of the three levels is this really — redecoration, renovation, or rebuild? Are we resourcing it accordingly?
- Where does the time saved actually go, and have we told the team that honestly?
- Who is the one person accountable for how this runs, including when to pause it?
- Are our controls designed into the flow, or bolted on at the end where problems surface too late?
If you can answer those, you're ahead of the 95%. If you can't, that's the work to do first — and it's a lot cheaper to do now than after you've scaled the wrong thing.
Managing a rebuild, not a rollout?
We help leaders work out which level of change they're really taking on — then put a senior owner and real operating discipline behind the one that pays.
Book a call Start a conversationSources: KPMG, Make AI Scale — From experimentation to transformation (2026), and the KPMG UK AI Pulse survey referenced throughout it. Figures are self-reported by surveyed leaders and are directional rather than audited; the three-level framework and the "time-saved contract" and "human-in-the-lead" ideas are KPMG's, discussed here with our own commentary for UK SME and mid-market readers.