UK businesses have adopted AI widely and used it shallowly, and closing that gap is management work that begins before any purchase. The decisions that matter are which process changes, who is responsible for the change, and how anyone outside the team will know it worked.
What the adoption figures show, and what they hide
Adoption has almost tripled since late 2023 while the depth of use has barely moved, and that gap is where the work sits. The Office for National Statistics reports that self-reported use of AI among UK businesses with 10 or more employees rose from around 12% in late 2023 to around 35% by June 2026. Over the same period the average number of AI technologies used per adopting business moved from about 1.4 to about 1.6, and among businesses using at least one AI technology only 10% described their use as extensive.
Figures below are from the ONS bulletin Artificial intelligence in UK businesses: 2023 to 2026.
| Measure, UK businesses with 10+ employees | Late 2023 | June 2026 |
|---|---|---|
| Using at least one AI technology | ~12% | ~35% |
| Average AI technologies per adopting business | ~1.4 | ~1.6 |
| Reporting extensive AI use | not measured | 10% |
| Information and communication sector adoption | 58% | |
| Construction sector adoption | 13% |
Sector spread is more important than the headline number. The ONS lists information and communication at 58% and construction at 13%. A strategy copied from a software company will not fit a builder, a wholesaler or a care home, and the ONS evidence says so directly.
The ONS also found that employees report more AI use than their employers do, with 55% of employees saying they use AI for work or education against 35% of businesses reporting any AI technology at all. Some of that is informal, unapproved use, and some of it is work that no manager has recorded. The ONS notes that larger businesses are the most likely to be unsure how much AI their own staff use.
AI for business is a management discipline
Judgement about where AI fits and how to make it accountable is the scarce skill, and it arrives in the form of people. The University of Edinburgh's AI for Business MSc lists the managerial tasks plainly: allocating resources to buy AI applications, envisioning how AI enters an organisation, translating generic ethics guidelines into practice, and deciding which criteria to use when judging whether an AI deployment succeeded. Only one of those involves a vendor contract. The programme also covers the technical floor a manager needs, including basic programming and using data to support decisions, which is a sensible limit: enough to interrogate a model, and enough to notice when somebody is bluffing about one.
For a firm of forty people, this compresses into three artefacts: An appointed person, a list of approved tools, and one process chosen for improvement with a linked measure. All other requirements are optional until those exist.
Write the acceptable use policy before you buy anything
An acceptable use policy is the cheapest control you will deploy, and it is what lets staff use AI in the open, under rules they can read. Arbor Law's guide to developing an AI acceptable use policy sets out the structure: purpose and context, scope, links to existing IT and data protection policies, a list of pre-approved tools, the process for requesting new ones, evaluation criteria, and rules for inputs and outputs. It also identifies shadow use as the problem the policy exists to solve, which matches what the ONS data implies about unrecorded employee use.
The output rules are the ones staff break most often, so the wording is worth reproducing exactly:
Do make sure there is a human-in-the-loop when using any AI tool, at each material stage of its use and with overall human oversight. Do check and verify the accuracy of all AI outputs before you use them more widely, share or publish them.
That is Arbor Law's wording, and it is the standard to follow. A register is the operational half of the same idea:
# ai-tools-register.yml
- tool: Microsoft 365 Copilot
status: approved
owner: operations
permitted_data: internal, non-personal
prohibited: customer records, contracts, source code, staff data
human_review: required before external publication
review_date: 2027-03-01
- tool: any new request
status: pending
owner: legal
requires: DPIA if personal data is processed, vendor terms review, security sign-off
Arbor Law recommends reviewing the policy at least annually, because the tooling and the law both move faster than that. A dated review line in the register is what makes the review happen.
Choose a governance framework by cost and certification need
Pick the framework that matches your size and your obligation, and leave the rest until something changes. The CIPD's guide to AI standards, produced with the British Standards Institution, maps four risk management frameworks and six international standards, then makes the choice concrete. Small enterprises should start with the NCSC AI Cyber Security Code of Practice as a security baseline plus the NIST AI Risk Management Framework. Medium and large enterprises should consider BS ISO/IEC 42001, the certifiable AI management system standard. Regulated sectors should pair that certification with a governance programme. The cost column is blunt: The NCSC code is free guidance, whereas ISO/IEC 42001 involves audit and implementation costs.
Two other standards earn their place. BS EN ISO/IEC 23894 covers AI risk management without a certification bill, and the CIPD's worked example of buying an AI recruitment system shows how it behaves in practice: the people team kept the shortlisting decision with human recruiters, restricted the tool to recommending, and ran random checks on both rejected and shortlisted candidates to catch biased decisions. That is what oversight looks like when it is written down.
What government has already published
The public sector has assembled a working governance stack in the open, and a private firm can copy the structure without being bound by it. The Government Analysis Function collates approved AI guidance from departments and specialist teams, and points new readers first to the cross-government Data and AI Ethics Framework and then to the Artificial Intelligence Playbook for the UK Government. The framework sets out core principles for designing, developing and deploying analytical or data-driven work, covering transparency, fairness, accountability and wider impacts. The playbook contains ten principles for civil servants using AI and covers what AI can and cannot do, how to select it, and how to mitigate the risks it brings.
The playbook's own framing is worth keeping:
"The AI Playbook will support the public sector in better understanding what AI can and cannot do, and how to mitigate the risks it brings."
A private firm does not need ten principles. It needs two or three, written in its own language, and a route to escalate a concern. Borrowing the framework's headings is faster than inventing them and easier to defend when a client or an insurer asks.
The risks no one budgets for
The failure modes are mundane, and most of them arrive before the first model does. nibusinessinfo's guidance on the risks and limitations of AI in business lists them in the order a small firm meets them: * poor-quality data that is fragmented, inconsistent or out of date; a skills shortage that pushes the work to outside consultants and pushes the cost up; ongoing spending on training, maintenance and updates after the initial software agreement; and existing infrastructure that cannot support the processing demand.
Two more are below those. AI systems need regular updating to stay effective, and restoring one after a failure takes longer than restoring ordinary software, because the behaviour depends on data as well as code. Privacy and transparency then arrive together: systems handling customer data raise data security questions, and a model whose reasoning cannot be explained makes it harder to satisfy a privacy regulator or to answer a customer. The guidance is direct that these are the conditions to fix before adoption, and the fix for most of them is a data management practice plus a written strategy.
Where the skills come from
Training is a line in the budget, and it is the cheapest part of the programme. The Business Academy series from business.gov.uk runs free sessions on using AI for market research, analytics and reporting, marketing measurement, customer service and lead generation, all aimed at firms that have not started. Wharton's AI for Business programme takes the other end of the same market: a self-paced online course, listed at $850, covering big data, machine learning, generative AI and governance, with an optional stack into a digital leadership certificate. Neither one makes a deployment safe on its own, and both cost less than a software subscription that no person was trained to use.
The order that works is the order on this page. Record what counts as AI use and who has the legal right to it. Publish the rules. Pick one framework and one process. Then spend money, and measure the process afterwards.
Sources
- Artificial intelligence in UK businesses: 2023 to 2026
- AI for Business MSc, University of Edinburgh
- How to develop an AI acceptable use policy for my business
- Navigating AI standards: a practical guide for people professionals
- Artificial Intelligence (AI) guidance, best practice and standards
- Risks and limitations of artificial intelligence in business
- Use AI to drive your business' growth
- AI for Business Online Course, Wharton Executive Education





































