We've reached that point in the economic cycle of AI where we've done the excitement phase. We've done the experimentation phase. Now we're in the implementation phase.
What we're experiencing now is the bottom of the J-curve. The J-curve is a well-known phenomenon where something that has potential and large upside initially shows up as increased cost and effort.
This is what businesses are seeing right now. Organisations have invested in AI tools, they're spending the tokens, and yet they're looking around wondering when the promised efficiency and opportunity will materialise.
The clearest version of this came from a recent conversation with someone whose job is getting a few thousand lawyers to use AI. He told me colleagues keep pulling him aside to confess the same thing: they went to the training, and they have still never opened the tool.
Then, on Monday, the Office for National Statistics published something that explained what's happening here.
Adoption nearly tripled. Depth barely moved.
The ONS release is called Artificial intelligence in UK businesses: 2023 to 2026, and 38,637 businesses responded to it and there are two numbers worth your attention...
The share of UK businesses with ten or more employees using at least one AI technology went from around 12% in September 2023 to around 35% in June 2026. Over the same three years, the average number of AI technologies used per adopting business crept from around 1.4 to around 1.6.
So the number of companies doing AI nearly tripled, and how much AI any of them actually do barely shifted. The ONS wrote the conclusion in themselves: "This would imply relatively limited transformative impacts to date for most AI-adopting firms." (Source: ONS, 20 July 2026)
There is a second figure further down that I have not been able to put down. Across all UK businesses with ten or more employees, only 11% report that more than half of their workforce has received any AI-related training.
A third of businesses have brought AI in. About one in ten has trained most of the people expected to use it.
It's not just AI depth that's not moving
In February, economists from the Bank of England, the Atlanta Fed, the Bundesbank, Stanford and four universities published a survey of around 6,000 senior executives across the US, UK, Germany and Australia. 89% reported no impact from AI on their firm's labour productivity over the previous three years, measured as sales per employee. (Source: NBER Working Paper)
Then the researchers asked those same executives how much they personally use AI in a working week. The most common answer was up to an hour. The average was about ninety minutes. More than a quarter said they do not use it at all.
Three years in, most organisations still cannot detect the difference AI has made to them, and the people running those organisations are using it for about as long as it takes to eat lunch.
My question to you: when did you last check how much time your leadership team actually spends on these tools?
If leadership teams are pushing these tools, then it only feels appropriate that they should lead by example as to what the future of personal productivity and role redesign looks like.
There are two stories. But neither are true
One says AI never lived up to the hype, the productivity gains were always overstated, and this is the dot-com bubble in a better jacket.
The other says the gains are real and enormous, the laggards just have not adopted fast enough, and the fix is more tools, more licences, more agents.
Both narratives assume capability arrives with the software. It does not. The models are extraordinary and they get better every quarter. The constraint is sitting on the other side of the screen, and it has not moved in three years.
The growing capability gap
AI models are getting better, but we, as humans, are barely getting any better at using them. And, before someone offers up their 'game-changing prompt library', being good with AI is not just about prompting skills, it's the more important stuff like being able to spot opportunities, being able to think in systems, building judgement, cultivating agency. The uniquely human skills we all need to continually sharpen.
I have spent the last month talking to the people running AI adoption inside law firms, banks, airlines, charities and professional services firms. The pattern is consistent enough to be a bit uncomfortable.
Someone at a large UK bank described what he was looking at as a massive delta between the value available and what is actually being done. A business that had rolled licences out across its teams told me, without much ceremony, that they had never really trained those users who now had access.
My favourite one, if favourite is the word, came from a customer success director. Two or three people in his team had built genuinely good automations. He only found out because someone mentioned it in passing. Everyone else on that team was solving the same problems from scratch, in parallel.
None of that is a story about bad technology.
Training is not the same thing as capability
I should declare an interest, because I run AcademyAI, which builds AI capability in organisations. But here is what I keep seeing. A training session is an event, and capability is a state, and organisations keep buying the first one hoping it produces the second. You can run a hundred-person webinar on prompting and change absolutely nothing.
The ONS data hints at why the instinct persists. Firms that admit they have an expertise gap are more than twice as likely to be training existing staff, 62% of them against 26% of firms reporting no barriers. So the instinct is sound. The delivery is where it comes apart.
What I see working is narrower and much less impressive in all honesty. Sitting next to someone at the point of need, which is what the lawyer I mentioned earlier told me was the only thing that had moved the dial for him. Working out what good looks like for one specific role rather than for everyone. Taking the thing one person built and getting it in front of the other two hundred.
Someone at a large charity told me they have 160 members of staff and not one shared job title. Not one. That is why the single generic course fails, and also why the answer is not 160 bespoke ones. It is working out what level of capability each person actually needs, understanding what they need to unlock productivity in their own unique roles and contexts and building the skills day-by-day through always on personalised micro-learning.
The part that should bother you
The models improve roughly every quarter. Organisational capability, left alone, does not improve at all. The ONS just showed us three years of exactly that in a single chart.
So the gap is not sitting still waiting for you. Every quarter you do not deliberately build capability, the distance between what your tools can do and what your people can do with them gets wider.
Everyone has access. Almost nobody is seeing the return.
What would it actually take to close that gap where you work this year? Would love to hear what has and has not worked for you.
Thanks, as always, Andy
PS. Both sources are worth reading in full if this is your world: ONS, Artificial intelligence in UK businesses: 2023 to 2026 and NBER Working Paper 34836, Firm Data on AI.
