On this page
What Healey said about AI
Earlier today (7 September 2026), John Healey gave part of his growth speech over to artificial intelligence. He called it a general-purpose technology, meaning one that reaches across the whole economy instead of a single industry. He said every firm, community, hospital and school could benefit.
He was careful about the risks too. He named pressure on resources, uncertain effects on the labour market and risks to national and business security. Then he put both halves in one sentence:
As Chancellor, I will not let this opportunity of AI pass Britain by, but nor will I allow this technology to proceed with no oversight.
Take that as the test he has set himself. Speed and oversight together, with neither traded away for the other.
One number in the speech deserves a second look. Healey said the capability of frontier models is doubling every four months on some measures. He didn’t say which measures, or whose. Until someone puts a source to that figure, I would not build a decision on it.
Most of what followed had been announced already. Healey pointed to the AI Security Institute, the AI Economics Institute and an AI minister in Cabinet. Then he said: ‘I did this at Defence, I announced this for Sovereign AI last week.’ HM Treasury and the Department for Science, Innovation and Technology set up the AI Economics Institute in June 2026. The Sovereign AI competition opened in late August.
That still tells you something useful. A Chancellor put AI in the innovation section of a growth speech. That signals where policy attention and money will sit for the rest of this Parliament.
One item pointed forward. Healey set an aim of having new sandboxing powers in place, ready to deploy across the economy next year.
What is open now, and who it is for
Two routes came out of the speech. Here’s how to tell quickly whether either one is yours.
The £100 million Sovereign AI competition
This buys AI capability from UK companies to address four challenges set by government. They are compute efficiency, AI in defence mission environments, safe AI agent adoption, and NHS productivity.
The guidance is clear about who can bid. The lead applicant must be a UK-registered company, and the project must be delivered in the UK. Universities and research organisations can take part only as subcontractors, and no more than half the work can be subcontracted. There is no minimum turnover, trading history or net asset requirement.
Contracts run from £250,000 to £10 million, with most expected between £1 million and £3 million. All project activity and payment claims must be complete by 31 March 2030.
If you build AI, the timetable is tight. Companies join an Approved Supplier List through an expression of interest, assessed pass or fail and returned within two weeks. Approved suppliers then submit a full proposal into a competition batch.
Batch one closes on 16 October 2026, batch two on 31 December and batch three on 26 February 2027. The guidance asks for the expression of interest at least a fortnight before the batch you are aiming at. For batch one that means 2 October.
If you adopt AI instead of building it, you can set this one aside and move on.
Sandboxing
A sandbox is a controlled arrangement for testing something new under agreed conditions, usually with a regulator watching.
Healey described inventors testing frontier technologies, from pavement robots to drones to lifesaving medical treatments. He set the aim of having powers ready to deploy across the economy next year. Powers that are ready next year are powers that do not exist now.
What exists today sits with individual regulators, and each arrangement is tied to a sector and a technology. Medicines, financial services and legal services each have their own routes. If you work outside a regulated sector, keep an eye on the legislation and carry on with your own work.
The number worth your attention
The Office for National Statistics published its combined view of AI in UK businesses on 20 July 2026. It draws on the Business Insights and Conditions Survey, wave 159, with fieldwork from 5 to 28 June 2026.
Among UK businesses with 10 or more employees, 35% reported using at least one AI technology. That is up from around 12% in late 2023. Adoption runs at 28% among businesses with 0 to 9 employees and 49% among those with 250 or more.
Now set that against a second ONS finding. Over half of employees, 55%, report using AI for work or education. That sits well above the 35% of businesses reporting they use it.
That gap is your head start. Your people have already begun. Somebody in your organisation has worked out how to draft a funding report faster. Someone else is turning a pile of interview notes into something readable. The capability is already in your building, and you didn’t have to buy it.
The catch is that you probably can’t see it. Only 11% of businesses with 10 or more employees have given AI training to more than half their workforce. Use is running ahead of training, and ahead of whatever the organisation believes it has approved.
Depth fills in the rest of the picture. The average adopting business uses around 1.6 AI technologies, up from around 1.4 in late 2023. Only 10% of adopters with 10 or more employees describe their use as extensive. Adoption across the UK is wide and shallow.
You may have seen 16% quoted as the UK adoption figure. That comes from the Department for Science, Innovation and Technology’s AI Adoption Research, and it measures something narrower. The study covered 3,500 UK private sector businesses with at least five employees, with fieldwork between 12 February and 2 May 2025. It left out the smallest businesses, along with public administration and defence, education, and human health and social work. The ONS work is a year fresher and counts micro businesses.
The DSIT study is still worth reading on depth. Among businesses using AI, three quarters said staff were getting more done, while over three quarters said revenue had not changed. Those answers are self-reported, so they tell you what businesses believe about AI rather than what AI caused.
Both surveys carry limits worth naming. The ONS survey excludes agriculture, oil and gas extraction, energy, public administration and defence, public provision of education and health, and finance and insurance. Neither figure covers the whole UK economy.
What this looks like inside an organisation
The patterns below come from Insightful AI discovery work and governance reviews. They describe what we find and what we recommend. We haven’t measured what happens to staff behaviour after a control goes in, so I have kept all of this qualitative.
Staff reach for free or personal AI tools before leadership has any view of what is in use. A list of paid licences looks like an inventory. It misses consumer accounts, and the AI features already sitting inside software you pay for.
That informal use is rarely reckless. People reach for whatever is accessible because the official route is slow or badly matched to the work in front of them. Treat that as information about your processes.
Training tends to follow access. Where it exists, it often covers general awareness without naming approved tools, data limits, how to check an output, or when to escalate a concern.
A policy is not evidence of control. Your board will want to show the policy is being followed. That needs a way to find undeclared tools, record use cases, handle exceptions and show that staff understand the rules.
Where to start
Nothing in the speech changes what good adoption looks like.
Insightful AI’s Adoption Roadmap runs in six stages: assess and align, build AI fluency, establish governance, pilot and prove, scale and integrate, sustain and improve.
Start at the first stage, because that is where the statistics point. Find out what your people are already using, including personal accounts and features embedded in software you already buy. Ask them directly, and make it safe to answer honestly. You’ll learn more in an afternoon of open conversation than from any licence report.
Then separate the four things that headline adoption figures blur together. Whether a tool is available. Whether people use it. Whether those people have been trained. Whether anyone governs it.
Work out which of the four you can evidence today.
For a structured route through that, our AI strategy support starts with your workflows and the data behind them.
If you build AI, open the Sovereign AI competition guidance and read it properly. Batch one closes on 16 October, and the expression of interest needs to be in by 2 October. If you adopt AI, start by asking your team what they are already using. The answer is usually more than you expect, and that is a better place to begin than a blank page.