The UK’s AI Adoption Gap Is Now Inside the Organisation

The new adoption question is not simply whether a business uses AI

## Evidence

UK business AI adoption has accelerated. The Office for National Statistics (ONS) reports that the share of businesses with 10 or more employees using at least one AI technology rose from around 12% in late 2023 to around 35% in June 2026. Yet adoption remains relatively shallow: among adopters, the average number of AI technologies used increased only from around 1.4 to 1.6 over the same period (Office for National Statistics, 2026) ↗.

The more revealing finding is the difference between worker and business reporting. In May to June 2026, 55% of employees reported using AI for work or education, compared with around 35% of businesses with 10 or more employees reporting use of at least one AI technology (Office for National Statistics, 2026) ↗. These are **not like-for-like measures**: the employee figure includes use for education and may include informal or ad hoc use, while the business measure covers specified technologies and excludes most firms with fewer than 10 employees. It would therefore be wrong to treat the difference as a precise measure of unauthorised use.

## Interpretation

Even with that caveat, the direction of travel is analytically important. AI is often arriving through individual tasks before it appears in the organisation’s operating model: drafting, summarising, research, customer communication, coding, document preparation and design work. The resulting gap is not merely a technology-procurement issue. It is a management issue about where work is done, which outputs require human verification, what information may be entered into a tool, and how the time released is reinvested.

This challenges a common assumption: that an employer must first select an enterprise-wide platform and then train the workforce. In many smaller organisations, the sequence is already reversed. Employees are discovering tools, while leadership is still deciding what counts as acceptable, valuable and safe use.

## Sanctuary recommendation

Treat this as an **integration gap**, not an employee-compliance problem. Start with a short, non-punitive discovery exercise: which tasks are already AI-assisted; which tools are in use; what information is being handled; where do colleagues see value; and where could an inaccurate output create harm? The purpose is to establish a realistic baseline before writing policy or buying further technology.

This complements Sanctuary’s earlier argument that productivity depends on redesigning work rather than simply purchasing AI tools AI adoption, SME management capability and workflow redesign.

Why adoption often stalls after the first experiment

## Evidence

The ONS finds that the most commonly reported barriers to AI adoption in 2023 were difficulty identifying suitable activities or use cases (39%), cost (21%) and lack of AI expertise or skills (16%). It also finds that firms with stronger management practices were more likely to adopt advanced technologies and to follow through on intended AI adoption (Office for National Statistics, 2025) ↗. The association between technology adoption and 19% higher turnover per worker in that analysis does not prove that AI caused higher productivity; more capable firms may be better placed both to adopt technology and to perform well. But it does strengthen the case that organisational capability matters.

The 2026 Skills for AI employer evidence points in the same direction. Its employer guide, drawing on 23 workshops, 10 case studies and 536 survey responses, identifies weak leadership support, unclear governance, poor data readiness and lack of sustained impact measurement as constraints on effective AI capability-building (Department for Work and Pensions and Skills England, 2026) ↗. Most organisations surveyed were still at awareness or exploration stages, with relatively few reporting integration (7%), strategy (4%) or scaling (1%) (Department for Work and Pensions and Skills England, 2026) ↗.

## Interpretation

The bottleneck is therefore unlikely to be a shortage of prompts. It is the absence of a repeatable way to decide: **which task should change, who owns the decision, what checks remain human, what data is permissible, and how will value be assessed?**

A generic AI course can raise confidence, but confidence alone may produce inconsistent quality. Conversely, a restrictive policy can reduce visible risk while leaving staff to make unshared decisions about tools and workarounds. Neither response creates a dependable operating capability.

## Sanctuary recommendation

For each proposed use case, use a one-page workflow brief before deployment:

1. **Task and baseline:** define the current process, volume, delay, error or rework burden.

2. **User and beneficiary:** identify who performs the task and who gains from improvement.

3. **AI contribution:** specify whether AI drafts, classifies, searches, summarises or analyses; avoid the vague instruction to ‘use AI’.

4. **Human control:** name the accountable reviewer and define the checks needed before an output is acted upon or published.

5. **Information boundary:** state what data must never be entered and what approved environment is required.

6. **Success test:** select one or two measures, such as cycle time, first-time quality, conversion, response quality or staff capacity released.

This turns adoption from an abstract digital ambition into a decision about a real workflow.

Training should build judgement, not just tool familiarity

## Evidence

Businesses reporting a lack of AI expertise as a barrier were more likely to train or retrain existing staff: around 62% did so, compared with around 26% of businesses reporting no adoption barriers. Yet only 11% of businesses with 10 or more employees reported that more than half of their workforce had received AI-related training (Office for National Statistics, 2026) ↗.

The Skills for AI research reports that organisations face particular difficulty in developing technical AI skills (67% of survey respondents) and responsible and ethical AI skills (32%). It also stresses that judgement, problem-solving and collaboration remain critical even where respondents report fewer difficulties in delivering non-technical skills (Department for Work and Pensions and Skills England, 2026) ↗. Its guidance recommends learning linked to real tasks, repeated practice, feedback, safe and responsible use, and support for different roles and experience levels (Department for Work and Pensions and Skills England, 2026) ↗.

## Interpretation

The useful distinction is between **tool literacy** and **work judgement**. Tool literacy helps someone produce an output. Work judgement helps them decide whether the task was appropriate for AI, whether the output is reliable enough, what needs checking, when to escalate, and when not to use AI at all.

That distinction matters especially for small businesses and community organisations, where one person may combine customer service, operations, finance, marketing and safeguarding responsibilities. A generic course detached from those decisions is unlikely to be sufficient.

## Sanctuary recommendation

Replace one-off awareness sessions with short, role-based learning cycles. A practical cycle should include:

- a live task that participants genuinely perform;

- an agreed example of acceptable and unacceptable input data;

- comparison of an AI-assisted output with the existing process;

- peer review against quality criteria;

- explicit discussion of failures, uncertainty and escalation; and

- a follow-up session after four to six weeks to review whether the workflow has changed.

For managers, training should additionally cover work allocation, quality assurance, staff consultation, capability development and how to protect entry-level learning opportunities when routine tasks are redesigned. This is a human-capital intervention as much as a technology intervention Human Resources Support.

A proportionate governance model for SMEs: visible enough to be useful, light enough to be used

## Evidence

The government-backed Skills for AI guide identifies unclear rules, responsibilities and goals as factors that reduce confidence and limit adoption. It describes a progression from awareness and experimentation to upskilling, integration, strategy and scaling, rather than treating adoption as a single procurement event (Department for Work and Pensions and Skills England, 2026) ↗.

ONS data also indicate that large language models were the most widely used AI technology among businesses with 10 or more employees in June 2026, reported by 18% of firms. This makes governance of everyday text, data and decision-support tasks more urgent than a narrow focus on advanced automation alone (Office for National Statistics, 2026) ↗.

## Interpretation

Governance is often misread as a barrier added after innovation. In practice, proportionate governance can be an enabler: it makes permissible experimentation visible, gives staff a route to ask questions, and lets leaders distinguish low-risk assistance from uses requiring specialist review.

The appropriate control level depends on the task, information and consequences. Drafting a public event description is not equivalent to generating advice for a vulnerable client, processing employee information, or making a decision with financial, safety or legal consequences. This article is not legal advice; organisations should seek specialist advice where personal data, regulated activity, employment decisions, intellectual property or contractual commitments are involved.

## Sanctuary recommendation

Adopt a simple three-lane model:

- **Green — supported use:** low-consequence drafting, brainstorming or formatting using approved tools and non-sensitive information; routine human review remains required.

- **Amber — controlled pilot:** internal analysis, customer-facing content or operational documentation where accuracy, confidentiality or reputational impact require an owner, defined checks and recorded learning.

- **Red — specialist approval:** activities involving sensitive personal data, high-impact decisions, safeguarding, regulated advice, safety-critical activity or binding commitments. Do not proceed without appropriate expert review and an approved process.

The model should sit alongside, not replace, existing data protection, information security, HR, procurement and safeguarding responsibilities. The test is practical usability: staff should be able to identify their lane in under a minute.

What leaders should measure in the next 90 days

## Evidence

Headline adoption is a weak proxy for value. ONS reports rapid growth in the share of businesses using at least one AI technology but only a modest increase in the number of technologies used by adopting firms (Office for National Statistics, 2026) ↗. The Skills for AI guidance similarly identifies missing impact measurement and unsustained training as obstacles to progress (Department for Work and Pensions and Skills England, 2026) ↗.

## Interpretation

The strategic risk is not that every employee uses a different tool. It is that an organisation confuses activity with improvement: more prompts, subscriptions or training completions without better service, stronger quality, greater capacity or new local economic value.

## Sanctuary recommendation

Over a 90-day period, select no more than three workflows and monitor a small balanced scorecard:

- **Performance:** turnaround time, rework, error rates or service responsiveness.

- **People:** confidence, training participation, task quality and whether junior staff still receive developmental work.

- **Risk:** exceptions, escalations, data-handling incidents and outputs rejected in review.

- **Value:** capacity released, customer outcomes, new services tested or revenue opportunities created.

Review the findings with the people doing the work, not only the senior sponsor. Retain, redesign or stop each pilot on the basis of evidence. This creates the management discipline required to translate informal use into resilient capability and sustainable growth. For organisations needing an integrated operating-model review, this is a natural fit for Management Consultancy.