Eight per cent of organisations have reached established returns on their AI investment, according to KPMG’s own Global AI Pulse survey for the first quarter of 2026. The firm’s UK People Consulting team published its explanation for that number today, and the diagnosis is convincing: businesses are bolting AI onto structures, operating models and job designs drawn up when only humans did the work. The cure they propose is to stop managing roles and start managing tasks. It is the correct answer. It is also, for the overwhelming majority of British firms, an answer arriving several stages too early.
What KPMG is actually arguing
The core claim is that adoption is not the constraint. Most organisations have AI running somewhere. What they lack is any redesign of the work itself, so the technology accelerates the execution of processes that were never re-examined. KPMG calls the outcome an efficiency trap: incrementally faster at the same things, structurally no different.
The alternative on offer is a shift in the unit of analysis. Jobs are not what AI changes; tasks are. So the argument runs that organisations need visibility of work at task level, mapped against the capabilities each one needs, the roles they sit in, the headcount available and the outcomes they are meant to produce, before they can sensibly decide what to automate, what to augment and what to stop doing.
Laurene Batkin, a director in the firm’s People Consulting practice, frames the target state as “connecting workflows and redesigning how AI and human operate together”, turning fragmented activity into something that behaves “more like a smart city”. Megan Butler, a senior manager working on AI and the workforce, is blunter about the prerequisite: today’s data is adequate for today’s decisions, and “we need better and different data to support the more complex decisions”.
Strategic Insight: The diagnosis and the prescription are separable. You can accept that AI transforms tasks rather than jobs without accepting that the only route there is an enterprise-wide task taxonomy. The first is an insight about work. The second is a programme with a budget.
The numbers, and what they cost
KPMG illustrates the approach with client engagements rather than survey data, which makes the scale of the intervention unusually visible.
| Data point | Value | What it implies |
|---|---|---|
| Organisations with established AI ROI | 8% (KPMG Global AI Pulse, Q1 2026) | The value gap is real and near-universal, not a laggard problem |
| One engagement’s task deconstruction | 54 activity clusters covering 2,000 tasks, drawn from over 200 roles | Roughly ten discrete tasks per role, catalogued and clustered |
| A global FMCG workforce intelligence build | 280 roles, catalogued as more than 4,000 tasks | Built on a skills repository that had already taken two years |
| Skills needing to change in one client’s technology function | Over 55% | The scale of capability shift the analysis is designed to expose |
Read those middle two rows as a resourcing estimate rather than a case study. A 2,000-task inventory is not a workshop output. The FMCG example is explicit that the underlying skills repository was developed over two years before AI agents were pointed at it to link roles and skills. KPMG is honest about this, describing the data work as invisible but essential. What follows from that honesty is a question the article does not ask: what does the firm that cannot commission a two-year data foundation do on Monday?
The training gap this lands in
Yesterday, Scottish Engineering published the first AI questions in a member survey series that has run since 1992. Seventy-two per cent of responding manufacturers have staff using AI tools regularly. Sixty-eight per cent provide no formal training of any kind. Five per cent have a finished AI strategy, and 39% have no plans to write one. A further 22% could not say what their usage level was.
We covered those figures in full yesterday, and they are worth holding against the KPMG argument, because the two describe the same country at opposite ends of a telescope. One is a portrait of enterprises with workforce intelligence functions, deciding whether their task taxonomy is granular enough. The other is a portrait of firms where AI is already in daily use, nobody has been trained, and a fifth of employers cannot report their own adoption rate.
Reality Check: An organisation that cannot say how many of its staff use AI is in no position to catalogue 2,000 tasks across its role structure. The task inventory is downstream of basic visibility, and basic visibility is what is missing.
This is not a reason to dismiss the redesign argument. It is a reason to sequence it. KPMG’s five lessons were drawn from organisations already spending seriously on transformation, and they are accurate for that population. Applied to the median UK employer they describe a destination without a first step.
Why the sequencing matters more than the model
There is a version of this that goes badly. A board reads a consultancy point of view, concludes that work redesign is the answer, and commissions a mapping exercise while its actual AI usage runs ungoverned in the background. The output is a taxonomy of how work was supposed to be done, produced at some expense, alongside a reality in which staff have already redesigned their own tasks privately and told nobody.
That is not hypothetical. It is what a 72% usage rate against a 68% no-training rate describes. The informal redesign has happened. The question is whether the organisation finds out about it before or after it builds a model of the work.
Hidden Cost: Every month a task inventory takes to build is a month the underlying work keeps changing. On a two-year timeline, a meaningful share of what you catalogued at the start is wrong by the time you finish, and nobody owns the reconciliation.
Who this actually lands on
| Group | What changes | What they need | How you know it worked |
|---|---|---|---|
| HR and people functions | Ownership shifts from headcount planning to work design, without a corresponding budget shift | Task-level data literacy, and a mandate that survives contact with operations | Workforce decisions cite task evidence rather than role titles |
| Operations and process owners | Their processes become the unit of analysis, and inefficiencies become visible to others | Protection from the exercise being read as a performance audit | Process owners volunteer candidates for redesign rather than defending current state |
| Line managers | Asked to identify automatable tasks in their own teams | A credible answer to what happens to the person whose tasks were automated | Managers surface automation opportunities without being chased |
| Staff using AI informally | Their private workarounds become organisational data | Amnesty on undisclosed use, and training that follows disclosure | Reported usage rises to match actual usage |
The fourth row is the one that decides the others. Task deconstruction depends on people telling you what they actually do, and an employee who suspects the exercise is a redundancy list will describe their role generously. KPMG’s framing keeps the focus on work rather than people, which is analytically right and does nothing to change how it reads from the receiving end. The mitigation is not communications. It is deciding, and saying, what happens to released capacity before the mapping starts.
Success Factor: Publish the redeployment rule before the task inventory, not after. Once staff have seen a task list with their name attached, no reassurance recovers the data quality you have already lost.
A version scaled to what most firms have
The point-of-view piece is written for enterprises. The same logic compresses.
Implementation Framework: Task redesign at one-process scale
Phase 1: Find out what is happening (2 to 4 weeks)
- Survey actual AI use, with amnesty for undisclosed tools
- Pick one process that is genuinely painful and bounded, not the biggest one
- Record where effort concentrates today, from the people doing it
Phase 2: Redesign one process (one quarter)
- Split it into tasks and mark each automate, augment, or stop
- Decide where human judgement is doing real work, and protect it explicitly
- Train the affected people before the change, not after
Phase 3: Decide something (ongoing)
- Reallocate the released capacity to a named priority
- Reuse the task vocabulary on the next process rather than starting again
- Only build a taxonomy once you have three processes wanting one
The difference between this and the enterprise version is that the data foundation accumulates as a by-product of decisions rather than preceding them. That is slower per process and considerably more likely to survive a change of sponsor.
If you are starting from nothing: establish your actual usage rate, write a one-page acceptable-use position, and train the people already using the tools. Nothing else works before this.
If you have pilots running: pick the pilot with the clearest owner and redesign the surrounding process rather than adding another pilot. Pilot proliferation is the recognised failure mode, not a stage on the way to maturity.
If you already have workforce data: the constraint is decisions, not insight. KPMG’s fourth lesson is the sharpest one in the piece, and the one most likely to be skipped.
Resource Reality: One bounded process, honestly mapped, takes a quarter of a competent internal owner’s time over three months. That is affordable for a mid-sized firm. A 2,000-task enterprise inventory is not, and pretending otherwise is how these programmes end up shelved at the halfway point.
Four problems the point of view does not address
The measurement decays faster than it is built
Task inventories describe work at a moment. Generative tools change what a task involves on a quarterly cadence, and agentic deployments change it faster. A foundation built over two years is partly historical fiction on delivery, and the maintenance cost of keeping it current is rarely budgeted alongside the build.
Mitigation: treat the inventory as a rolling sample rather than a census. Re-baseline one function per quarter and accept staleness elsewhere, rather than aiming for a complete picture you cannot refresh.
Retraining is doing more load-bearing work than the evidence supports
Every task-redesign argument ends implicitly at reskilling. The evidence base for that is thinner than the confidence placed in it. A meta-analysis of 56 randomised American trials, reviewed by David Roodman with Anthropic’s Maxim Massenkoff, found that a training place lifts the odds of holding a job by two to three percentage points, worth about $1,000 in annual earnings, at a cost near $13,000. Real, defensible, and not a mechanism for absorbing displacement at scale.
Mitigation: plan redeployment into adjacent work you already have, rather than retraining into roles you hope will exist. The programmes with outsized returns in that review were the ones tied to a specific employer with a job at the end.
Shadow adoption has already made some of these decisions
Task-level redesign assumes the organisation gets to choose. Where 22% of employers cannot state their own usage level, staff have already decided which of their tasks are automated. The mapping exercise then documents an intended state rather than the real one.
Mitigation: run the usage census first and treat undisclosed automation as a finding rather than a breach. What people chose to hand over is the highest-quality signal available about which tasks were worth handing over.
The capability to do this is mostly rented
KPMG’s examples are consulting engagements, and the analysis capability sits with the consultancy. That is a legitimate service, and it also means the client’s internal ability to repeat the exercise next year is an open question. Our earlier analysis of Big Four AI hiring traced how quickly this capability is concentrating on the supply side.
Mitigation: contract for the method as a deliverable alongside the output. If the engagement does not leave behind a task vocabulary your own people can extend, you have bought a snapshot at the price of a capability.
What to take from this
KPMG has identified the right constraint. The value gap in AI is not a model problem, a licensing problem or an adoption problem, and treating it as any of those produces exactly the marginal gains most organisations are reporting. Work designed around human-only capacity does not become AI-enabled by having AI added to it. That is worth internalising regardless of your size.
What does not transfer is the scale of the response. Three things separate the firms that will get value from this from those that will get a document:
- Sequence before scope. Usage visibility, then training, then process redesign, then taxonomy. Skipping to the last one is the common failure and the expensive one.
- Decisions are the deliverable. KPMG’s own fourth lesson is that insight changes nothing. Every mapping exercise should have a named decision waiting for it before it starts.
- Own the method. Whether the capability lives inside the organisation afterwards determines whether this was a transformation or a report.
Strategic Insight: The gap between the 8% seeing established returns and everyone else is not going to be closed by better models. It will be closed, slowly, by organisations that know what their people actually do and are willing to change it. Most UK firms currently cannot answer the first half of that.
Next steps
- Establish your actual AI usage rate this month, including tools nobody has declared
- Write down what happens to capacity released by automation, before any mapping begins
- Pick one bounded, painful process as the redesign candidate for this quarter
- Train the staff already using AI, ahead of any strategy document
- Name the decision each piece of workforce analysis is meant to inform
- Review in 90 days whether reported usage has moved towards actual usage
Source and attribution
This analysis draws on “AI is changing work”, a point of view published by KPMG in the UK on 18 August 2026, with contributions from Laurene Batkin (Director, People Consulting), Megan Butler (Senior Manager, AI Workforce) and Ali Ahsan (Assistant Manager). The 8% ROI figure is from KPMG’s Global AI Pulse Survey for Q1 2026; the enterprise coordination material cites KPMG’s Transforming the Enterprise 2026.
Supporting data: Scottish Engineering member survey figures published by Paul Sheerin in The Herald, 17 August 2026; “Reviewing the evidence on worker retraining programs” by David Roodman and Maxim Massenkoff, Anthropic, August 2026. KPMG separately reported in its Q2 2026 Pulse that nearly half of businesses had paused or scaled back AI projects on cost grounds.
Analysis and UK business framing by Resultsense. Read more of our workforce and adoption coverage or get in touch if you are deciding where AI redesign should start in your organisation.