CBI Adoption Decade Report: The AI Execution Divide Is a People Problem

In August 2026 the CBI, with knowledge partner Oliver Wyman, published The Adoption Decade: Closing the Execution Divide and Making AI Work for Britain. Its central finding is that the UK's AI problem is no longer invention or strategy. It is delivery. Firms leading on deployment hit their ROI expectations 49% of the time; laggards manage 15%. The gap between them is made of people, not technology.

What the CBI and Oliver Wyman actually published

The report landed on 18 August 2026 and was picked up the next morning across the trade and national press. It is not a technology forecast. It is a delivery argument, aimed squarely at Whitehall, and it makes four asks: raise productivity through adoption, build a competitive and resilient UK AI stack, enable trusted and responsible adoption, and build an AI-ready workforce.

Rain Newton-Smith, the CBI's Chief Executive, put the framing plainly: "The next chapter of AI won't just be defined by what we invent, but by how quickly we can put it to work at scale." Britain, the CBI argues, needs to treat AI adoption as a national economic priority rather than a departmental initiative.

That is a meaningful shift in emphasis from a body that has spent recent years talking about AI mostly in terms of investment, compute and regulatory clarity. The subject has moved from what Britain can build to whether British firms can use what already exists.

The execution divide, in one number

The statistic doing the heavy lifting is this: almost half of firms leading the way on AI (49%) report meeting or surpassing their expected return on investment, compared with just 15% of firms still stuck in pilot mode. The underlying research comes from the Oliver Wyman Forum and the New York Stock Exchange.

Sit with the size of that gap for a moment. It is more than three to one. And the two groups are buying from the same vendors, at broadly the same prices, with access to broadly the same models. Nothing in the technology explains a 34-point spread in whether an investment pays back.

"If two companies buy identical licences from identical vendors and one of them gets three times the return, the variable under test was never the software. It was everything the company did around it."

The report is honest about where value now accumulates: close to deployment, in workflow integration, applied models, trusted data and specialist implementation. Access to a frontier model is table stakes. What you do in the ninety days after the contract is signed is the entire game. We wrote about that gap in detail in moving AI from pilot to production, and the failure patterns in why AI adoption fails in companies.

Our perspective: the tech is ready, the people are not

We have been saying a version of this for two years, usually in a room with eight executives who have already bought the licences and cannot work out why nothing has changed. What is striking about the CBI report is not that it disagrees with us. It is that it says the same thing at national-policy scale, with a peer-reviewed number attached.

Here is our position, stated flatly. AI adoption is a people challenge far more than a technology challenge. The models are good enough. They have been good enough for eighteen months. The constraint is that most organisations have not changed a single management routine, job description, review cycle or definition of "done" to reflect the fact that a capable assistant now sits inside every workflow.

Toni Dos Santos, our Co-Founder and AI Advisor, puts it this way: "Every failed AI programme we are asked to rescue was a competent technology decision followed by no behavioural decision at all. Nobody changed what a good week looks like for a marketing manager. So the marketing manager kept having the old week, with an expensive tab open."

Meera Sanghvi, our other Co-Founder, adds the governance half: "The firms that stall are rarely the ones with the strictest rules. They are the ones whose rules are unwritten. When people cannot tell whether using AI on a client document will get them praised or disciplined, they choose the safe option, which is not to use it."

That is the trust dimension the CBI report keeps returning to, and it is not a soft consideration. Trust is the rate limiter on usage. Ambiguity reads as prohibition. If you have never published a one-page position on what staff may and may not put into a model, you have effectively banned AI without knowing it. Our UK ICO-aligned governance framework is where we usually start that conversation.

What actually separates a leader from a laggard

Across roughly fifty enterprise and mid-market engagements, the firms that land in the 49% share four traits. None of them is budget.

1. A named owner with a diary, not a steering committee

Leaders have one person accountable for adoption, with time formally allocated to it. Laggards have a cross-functional working group that meets fortnightly and produces slides. The difference shows up within six weeks.

2. Workflows, not tools

Training that teaches "prompting" produces a brief spike and a fast decay. Training built on the actual artefacts a team produces — this quote, this QBR, this credit memo, this tender response — produces habits. The unit of adoption is a workflow that somebody is already paid to complete, not a feature.

3. Managers who use it visibly

This is the single strongest predictor we see. If a head of department cannot demonstrate one thing they now do differently, their team reads AI as an initiative to survive rather than a tool to use. We built a playbook for turning sceptical managers into champions precisely because this layer decides the outcome.

4. A number they check weekly

Leaders measure something specific: hours from trigger to approved deliverable, percentage of a team's outputs touched by AI, weekly active use inside a named workflow. Laggards measure licences issued, which measures procurement. Our CFO guide to measuring AI ROI sets out the small set of metrics that survive a finance review.

The skills number nobody wants to look at

The CBI's fourth goal is an AI-ready workforce, and the supporting data explains why it is there. Research from Lloyds Banking Group cited alongside the report found that while 58% of firms believe AI has created jobs inside their organisation, almost a third (31%) say their workforce does not yet have the skills needed. The government's answer is an industry partnership targeting AI skills for 10 million workers by 2030, backed by more than £200 million.

Ten million is the right order of magnitude. But the word "skills" is carrying an enormous amount of weight, and it is worth being precise about what it has to mean, because most of what is delivered under that heading is awareness training.

An awareness session tells people that AI exists, that it sometimes makes things up, and that they should not paste client data into a consumer chatbot. That is useful for about forty minutes. It changes nothing on Monday. Skills, in the sense the CBI needs, means a named person can complete a named piece of their actual job faster, to the same or better standard, and can show you. That is a much more expensive and much more valuable thing to deliver. We laid out the distinction in closing the AI skills gap and in our guide to AI adoption training in the UK.

The UK evidence base backs the concern. British Chambers of Commerce data has 54% of SMEs reporting AI adoption in 2026, but only 11% using it extensively, and over 60% of UK businesses naming the skills gap as their main barrier, ahead of cost. We keep a running set of these figures in our UK SME AI adoption statistics roundup.

Where we push back on the report

Two things, offered in good faith.

First, "adoption" is doing too much work as a single word. The report treats adoption as one thing that firms either have or lack. In practice there are three distinct stages that fail for entirely different reasons: getting people to try it, getting them to use it for real work, and getting the organisation to redesign a process around it because usage is now reliable. A policy that funds stage one and calls it adoption will produce a lot of trained people and very little productivity. Most of the £200 million risk sits here.

Second, the demand side is under-specified. The report is strong on supply — skills programmes, the UK stack, compute, regulatory clarity. It is lighter on why a profitable mid-market firm with a full order book would voluntarily disrupt a working process. In our experience they do not, until either a competitor forces them or a leader personally understands the delta. That is a change management problem, and it is the one thing government cannot procure. We have written about the mechanics in AI change management for enterprises.

Neither of these makes the report wrong. The diagnosis is correct and the "adoption decade" framing is the right one. We would simply argue that the decade will be won or lost in middle management, not in Whitehall, and the report's own 49-versus-15 number is the evidence for that.

What a UK business should do in the next 90 days

If you read the CBI report and want to act on it rather than circulate it, this is the sequence we run.

1. Days 1–14: pick three workflows, not thirty. Choose ones with high volume, a clear quality bar and an existing owner. Tender responses, client reporting, first-draft marketing, support triage, credit memos. Write down the current time-to-deliverable for each. That number is your baseline and you will need it later to defend the spend.
2. Days 15–30: publish one page of governance. What data may go in, what must not, which tools are approved, who to ask. One page, signed by an executive. Ambiguity is what is suppressing your usage, not risk appetite. Align it to UK GDPR and ICO guidance rather than inventing your own scheme.
3. Days 31–60: train on the artefacts, in role groups. Not a company-wide webinar. Two to three hours per business unit, working on live documents, ending with each person having produced one real deliverable they would send. Managers attend the same session as their teams, and go first.
4. Days 61–90: measure and name owners. Re-measure the three baselines. Name one accountable owner per workflow with time in their diary. Report the delta to the board in hours and pounds, not in licences activated.

If the numbers have not moved by day 90, the design was wrong, and that is recoverable. If you never took the baseline, you will be arguing about vibes at the budget review, and that is not.

The line worth keeping

Strip the CBI report to one sentence and it says: Britain does not have an AI capability problem, it has an AI usage problem. That is a much better position to be in, because usage problems are solvable without inventing anything.

They are also harder than they look, because they are made of habit, incentive, permission and managerial courage rather than budget. The technology is ready. The organisations around it are not, and the distance between those two facts is worth roughly three times your return on investment.