Among firms leading on AI deployment, 49% report that their return on investment is meeting or beating expectations. Among laggards, the figure is 15%. That three-to-one gap is the load-bearing number in The Adoption Decade: Closing the Execution Divide and Making AI Work for Britain, published on 18 August by the CBI with Oliver Wyman, and it is what justifies the group’s call for the next ten years to be designated Britain’s adoption decade. The gap is real and well sourced. The problem is that the survey it comes from also says, in plain terms, what causes it, and none of the four things the CBI asks government for would change that variable.
Strategic Insight: The report contains the right answer. It files it under what businesses should do, while the asks directed at government are a convening body, a voluntary code, a literacy standard and a regulatory map. The diagnosis and the policy prescription are pointing in different directions.
What the 49% actually measures
The figure comes from the 2026 CEO survey run by the Oliver Wyman Forum with the New York Stock Exchange. Its methodology section deserves reading before the finding is applied to British business. The survey captured 415 chief executives, of whom 266 ran public companies and the rest private ones. Those public companies alone carry roughly $12tn of market capitalisation, around a tenth of the world’s listed equity. The Forum describes it as “the largest annual survey of large-company chief executives”.
That is not a UK dataset, not an SME dataset, and not a general business dataset. Europe as a whole supplies only 30% of the survey’s deployment leaders, against 47% in Asia-Pacific and 37% in North America. A firm turning over £800,000 in Stoke is not in this sample, has no counterpart in it, and shares few of its constraints.
More importantly, “deployment leader” has a definition, and it is not a flattering one for the skills argument. The Forum defines one as a company running AI at scale across two or more use cases. Leadership is defined by deployment breadth. Not by literacy scores, not by governance maturity, not by whether the firm signed anything.
And in the same passage where the 49% appears, the Forum reports what distinguishes those companies:
Critical Context: Deployment leaders are redesigning workflows at 49%, against 32% for those stuck in pilots. The Forum’s own summary is unambiguous: “It is the single clearest finding on AI in this year’s survey: Deployment drives ROI.”
Workflow redesign. Not training. The variable that separates the two sides of the execution divide, according to the research the CBI cites to establish that the divide exists, is whether the organisation has restructured how work actually happens.
| Data point | Value | What it implies |
|---|---|---|
| Deployment leaders meeting or exceeding ROI expectations | 49%, against 28% of all CEOs surveyed | The CBI’s 15% laggard figure sits below that all-CEO average; both cuts come from the same 415-CEO survey |
| Leaders redesigning workflows vs firms stuck in pilots | 49% vs 32% | The operating model, not capability, is the discriminating variable |
| CEOs saying AI ROI met or exceeded expectations | 27%, against 38% twelve months earlier | The returns picture is deteriorating across the sample, not improving |
| Firms qualifying as AI ROI leaders (>10% cost or revenue gain) | 12%, against 17% the previous year | The top of the distribution is thinning, which complicates a race-to-adopt framing |
That third row is awkward for an urgency argument. The share of chief executives who say it remains too soon to judge AI returns rose to 53% from 41%. Nearly a quarter report no revenue impact at all. The Forum treats this as a recalibration rather than a loss of nerve: reshaping how work gets done at enterprise scale is proving slower and more demanding than the early excitement implied. That is a reasonable conclusion and a poor foundation for a decade-long national mission framed around falling behind.
Do the four asks reach the mechanism?
The CBI wants four things from government, which we reported when the study landed: a time-limited Business AI Adoption Delivery Group with a minister sponsoring it, a voluntary Responsible AI Adoption Code carrying a public pledge, AI skills investment at all levels underpinned by a National AI Literacy Standard, and a “no wrong door” route through regulation backed by a map of which rules apply.
Assessed against the mechanism the evidence identifies, they perform differently, and none of them performs well.
| The ask | What it changes | Does it touch workflow redesign? |
|---|---|---|
| Delivery group | Coordination between Whitehall and large corporates already at the table | Indirectly at best. Convening bodies do not redesign anyone’s processes |
| Voluntary code and pledge | Gives firms a shared reference point on responsible adoption | No. It governs how you deploy, not whether you restructure |
| Skills investment and literacy standard | Raises baseline capability and confidence across the workforce | No. Literate staff inside an unchanged process produce a faster unchanged process |
| Regulatory navigation | Reduces uncertainty about which rules apply | Removes a barrier that ranks below cost and data quality in UK survey evidence |
The literacy standard is the one worth dwelling on, because it is the ask most likely to be adopted and the least likely to shift the number. Consider what happens when you raise AI literacy inside an organisation that has not touched its operating model. Staff become more capable of using tools within processes designed before those tools existed. That is the efficiency trap: incrementally faster at the same things, structurally no different. We made this argument in full against KPMG’s task-level redesign work earlier this week, and it cuts the same way here. Work redesign is the answer. Training is a prerequisite for it, not a substitute.
The evidence base under the skills ask is thinner than the ask assumes
Britain’s instinct in every technology transition is to reach for retraining, and the assumption underneath is that the constraint is funding. A meta-analysis of 56 randomised trials, covered here on 13 August, found that offering someone a training place improves their odds of holding a job by two to three percentage points, worth around $1,000 a year against a cost of roughly $13,000 per place. Programmes come close to paying for themselves. They are a defensible use of public money and a weak instrument for absorbing a structural shift.
The one category that performs several times better is employer-linked and sector-specific, with firms hiring participants directly. Attempts to replicate those programmes have repeatedly failed. So the intervention that works is the one nobody has reliably managed to copy, and a population-wide literacy standard is not it. A standard is classroom provision by another name, delivered at population scale, which is precisely the design the evidence rates lowest.
Reality Check: A literacy standard is cheap, popular, measurable and almost certainly worth doing. It is also, on the best available evidence, worth about two to three points of employment effect. The execution divide is a thirty-four point gap. These are not the same order of magnitude.
Whose divide is it?
The Lloyds Business Barometer, published the same day as the CBI report and drawn from 1,200 UK firms, gives the domestic picture the Oliver Wyman sample cannot. Six in ten firms use AI. Above £10m of turnover that average rises to 79%. Firms trading internationally sit at 63% against 46% for those selling only at home.
Then the number that should govern how anyone reads the CBI’s asks. Asked whether failure to adopt AI leaves a business at a competitive disadvantage, 73% of firms turning over £10m or more agreed. Among firms below £1m, 54% did. We covered the full barometer breakdown on 18 August.
Strategic Reality: The businesses furthest from adoption are also the least convinced they need it. A voluntary code and a public pledge are instruments that require a firm to already believe it has a problem before they do anything at all. They select for the converted.
This is the structural weakness in the whole package. Every one of the four asks is an opt-in instrument. A delivery group convenes the firms who turn up. A voluntary code binds the firms who sign. A literacy standard reaches the firms who send people on courses. A regulatory map helps the firms already trying to navigate regulation. For a business on the wrong side of both the £10m adoption gap and the sub-£1m confidence gap, none of these is a door that opens from the outside.
Lloyds also puts cost as the leading barrier at 18%, with data quality and access to skills tied on 17% each. Skills is a genuine constraint. It is not the top one, and 31% of firms reporting a workforce short of the skills it needs is a real problem that a decade of literacy standards would partly fix whilst leaving the operating model untouched.
The measurement problem nobody in the report raises
There is a further difficulty with the 49% that follows from how it was collected. It is a chief executive reporting whether returns met the expectations that chief executive held. It is self-assessment at the top of the organisation.
Research commissioned by Certinia and reported by Consultancy.uk found 69% of executive leaders calling their AI deployments successful against 53% of the decision-makers and practitioners actually using the tools. Within professional services the split widened by discipline: 55% at consulting firms, 43% across audit, accounting and tax. The caveat belongs in the same breath as the finding, and we flagged it when covering the study: the research was paid for by a vendor whose recommended remedy happens to be the category it sells. The perception gap itself does not depend on that remedy being right.
If executive self-reports on AI success run something like sixteen points ahead of practitioner assessment, then a survey of 415 chief executives measuring whether AI met their own expectations is measuring executive confidence with reasonable precision and organisational outcomes with rather less. That does not make the execution divide fictional. It does mean the number quantifying it should carry an error bar the report never draws.
What Simon Johnson is doing in this story
The Standard’s write-up places Professor Simon Johnson, who chairs the UK’s AI Economics Institute, in the same piece as the CBI’s call. Speaking to Newsnight, he described an intense industrial transformation arriving fast, and then said this: “But I think the companies are exaggerating. They are obviously trying to attract and impress investors. The technology is impressive, but we do have some time.”
He went further, arguing that the UK has time if it acts now to prepare, and framing the objective as a stronger and more resilient economy carrying “more good jobs”, one that incorporates AI instead of, in his phrase, “getting run over by AI”.
That is not a minor difference of emphasis. An adoption decade framed around the risk of being left behind and an economist saying the vendors are overstating the clock are two different theories of the situation, and they imply different policy. If Johnson is right, the correct response is not to accelerate diffusion for its own sake but to be selective about where AI is applied and rigorous about measuring whether it worked. The deteriorating ROI figures in Oliver Wyman’s own survey are consistent with his reading, not the CBI’s.
⚠️ Warning: Urgency framing is not neutral. A firm persuaded it is behind buys tools. A firm persuaded it has time redesigns a process and then buys tools. The second sequence is the one the data associates with returns.
Who this lands on
| Group | What changes | What they need | How you know it worked |
|---|---|---|---|
| Sub-£1m turnover firms | Little. Every instrument on offer is opt-in and they are least likely to opt in | One redesigned process with a measured before and after, not a strategy | A named process runs differently and the difference is quantified |
| £1m to £10m firms | Most exposed. Adoption is spreading, governance is not | Ownership of AI use at director level and visibility of what staff already do | You can state your actual usage rate without guessing |
| Large corporates | Already inside the delivery group and the survey population | Nothing from this package they did not have | Workflow redesign rate, tracked as a metric |
| Government | A convening role and a standard to build | Diffusion measured by outcomes, not pilots launched or pledges signed | Productivity data, not adoption headlines |
The Scottish Engineering member survey, published on 17 August, shows what the middle row looks like in practice. 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 strategy and 39% have no plans to write one. A further 22% cannot say what their usage level is.
Read that against a literacy standard and the mismatch is obvious. These firms do not have a literacy problem. Their staff are already using the tools daily, which means literacy is arriving informally and at speed. What they have is an absence of ownership, no visibility of their own usage, and no redesigned process for any of it to sit in.
Hidden Cost: A fifth of Scottish manufacturers cannot report their own AI adoption rate. Every governance instrument, code and pledge in the CBI package assumes an organisation that knows what it is doing with AI. The prior question is whether it knows what is being done.
What to do instead of waiting for the package
None of this argues for inaction, and none of it argues that the CBI is wrong about the destination. The report’s own advice to business is the correct one, and it is worth quoting because it is stronger than anything asked of government: “treat AI as business transformation, not a technology roll-out”, and to “prioritise a few of the highest-value use cases and move them from pilot to scale, with senior ownership, redesigned workflows, proportionate governance and security built in”. The CBI also states plainly that “Value increasingly depends on implementation, not access to technology.”
That is the whole answer. It requires no delivery group, no code and no standard.
💡 Implementation Framework: Crossing the divide without waiting for policy
Phase 1: Find out what is actually happening (2 to 4 weeks)
- Ask every team which AI tools they use and for what, without penalty for the answer
- Record the usage rate as a number you could defend to an auditor
- Name one director as owner of AI use, whatever their other title
Phase 2: Redesign one process end to end (1 quarter)
- Pick a process with a measurable cycle time or error rate, not the most exciting one
- Map how it currently runs, then how it would run if the AI step were assumed rather than bolted on
- Change the handoffs, approvals and roles the redesign implies, or the redesign has not happened
Phase 3: Measure, then repeat or stop (following quarter)
- Compare the before and after on the metric you chose at the start
- Ask the people doing the work whether it improved, and record the gap against your own assessment
- Scale to a second process only if the first one cleared its bar
Priority actions by where you are now
Firms not yet using AI in any structured way
- Establish the baseline: You almost certainly have shadow usage. Find it before you plan anything, because it tells you where the appetite already is.
- Choose one process, not a strategy: A single redesigned workflow with a measured result is worth more than a document nobody implements.
- Set the bar before you start: Decide what result would justify continuing, in writing, while you are still unbiased.
Firms with AI running but no redesign
- Audit for the efficiency trap: For each deployment, ask whether the surrounding process changed. Where it did not, you have bought speed on an unexamined activity.
- Move ownership up: The Forum’s leaders combine senior ownership with redesign. Delegated pilots are the pattern associated with the 15%.
- Check your own reporting: Ask the practitioners whether the deployment works before the number goes upward. The perception gap is measurable and it flatters management.
Firms already scaling
- Track redesign rate as a metric: Percentage of deployments accompanied by a documented process change. This is the variable the evidence associates with returns.
- Treat governance as scaling infrastructure: Boards at firms posting high AI ROI got about twice as involved in AI strategy as the survey average, 42% against 23%. Governance follows scale in the data rather than obstructing it.
- Engage the code on your terms: A voluntary code will be a reasonable articulation of practice you already have. Signing it is cheap. Expecting it to change your results is not warranted.
Resource Reality: Phases 1 and 2 need a director’s attention for perhaps two days a month and no new software licences. The binding constraint is willingness to change handoffs and approvals, which costs nothing and is harder than any budget line in this article.
Four things that will go wrong
The delivery group becomes the deliverable
Time-limited groups sponsored at ministerial level have a way of producing a workplan, an interim report and a successor body. Progress gets measured in convenings. The CBI’s own report warns against measuring success by “the number of pilots launched”, and the same discipline needs applying to the machinery: a delivery group that has not moved a productivity number in two years has not delivered.
Mitigation: Insist that the group publishes a diffusion metric at the outset and reports against it, with firm-size breakdowns. If nothing can be measured below £10m of turnover, the mission is not reaching the firms it was written for.
The pledge becomes a procurement filter
Voluntary codes with public pledges tend to acquire weight in supply chains. Large buyers start asking whether suppliers have signed. That is not the stated intention and it is a predictable consequence, and it lands hardest on small suppliers who have neither the governance function to satisfy the code nor the leverage to decline.
Mitigation: If a code arrives, read it as a future tender requirement and cost the compliance before your customers ask. Small firms should push for a proportionate tier now, while the drafting is open.
Literacy is measured, redesign is not
A standard creates a metric. Metrics attract reporting. Within two years the country will know how many workers hold an AI literacy credential and will still have no idea how many organisations changed a process. The measurable proxy displaces the thing that mattered, which is a familiar failure and entirely avoidable.
Mitigation: Whatever the standard measures, track your own redesign rate internally alongside it. Do not let a training completion percentage become your board’s AI metric.
The SME simply never appears
The £10m adoption gap and the sub-£1m confidence gap describe the same population from two angles: firms that have not adopted and do not think they need to. Nothing in the package creates a reason for them to engage. In ten years the aggregate adoption figure will have risen because large firms and international traders adopted further, and the distribution will have widened.
Mitigation: For policymakers, target the confidence gap directly with sector-specific, employer-linked provision, which is the only intervention class the retraining evidence rates highly. For SMEs, do not wait for it.
The takeaway
The CBI has produced a good report with a well-evidenced central finding and asked government for four things that do not act on it. The execution divide is an operating model divide. The survey establishing its existence says so explicitly, naming workflow redesign as the discriminating behaviour and calling deployment the driver of returns. A delivery group, a voluntary code, a literacy standard and a regulatory map are all reasonable and none of them redesigns a process.
Three things decide whether a British firm ends up on the right side of this:
- Ownership at the top: Leaders combining senior ownership with redesign are the ones reporting returns. Delegated pilots are the profile associated with the laggard figure.
- Redesign before deployment: The sequence matters. Tools inside unchanged processes produce faster versions of what you already did, and the ROI data across the whole survey is drifting downward as firms discover this.
- Honest measurement: Executive self-assessment runs ahead of practitioner experience by a measurable margin. Ask the people doing the work, and record the gap.
Strategic Insight: The most useful sentence in the entire package is addressed to businesses, not government: value depends on implementation rather than access to technology. Any firm that acts on that one line has done more for its position than the four policy asks combined will do for it.
This week:
- Establish your actual AI usage rate as a defensible number
- Name one director as owner of AI use
- Pick one process with a metric you already track
This quarter:
- Redesign that process assuming the AI step rather than adding it
- Change the handoffs and approvals the redesign implies
- Ask practitioners, not managers, whether it improved
This year:
- Track redesign rate as a board metric alongside any literacy measure
- Cost voluntary-code compliance before a customer requires it
- Decide, in writing, what result would justify the next deployment
Source: Make AI adoption Britain’s national economic priority, says CBI, reported by Holly Williams (The Standard, 18 August 2026). Report details and recommendations verified against the CBI’s own summary and its media release. ROI, workflow-redesign and methodology figures taken from the 2026 Oliver Wyman Forum and NYSE CEO survey. UK adoption and skills figures from the Lloyds Business Barometer release of 18 August 2026. Executive-versus-practitioner perception figures come from vendor-commissioned research paid for by Certinia and reported by Consultancy.uk.
This strategic analysis was written by Resultsense, a UK-focused AI news and analysis publication. We will keep tracking whether the adoption debate shifts from training volume to workflow redesign. Read more analysis at Insights, or get in touch.