UK SMEs should set data boundaries before scaling AI

The UK AI challenge has moved from access to management

For many UK SMEs, the question is no longer whether artificial intelligence will enter the business. It already has: through standalone tools, AI features embedded in office software and customer platforms, outsourced providers, or employees’ own ways of completing work. The more consequential question is whether management can see those uses, set boundaries for the information involved and retain accountability when an output affects a customer, employee or commercial decision.

This is timely because uptake has accelerated faster than organisational integration. In June 2026, around 35% of UK businesses with 10 or more employees reported using at least one AI technology, up from around 12% in late 2023. Yet adopters used an average of only roughly 1.6 technologies, compared with 1.4 previously; 10% described their use as extensive. Fifteen per cent said that more than half their workforce used AI daily (Office for National Statistics, 2026) ↗. Rapid diffusion, in other words, should not be confused with completed business transformation.

That distinction matters for productivity. Drafting, summarising and search tools can be added at the edge of an existing process with little organisational change. Sustained gains usually require harder work: redesigning hand-offs, connecting systems where appropriate, deciding how outputs will be checked and reallocating the time that has been saved. Those changes require management capacity that many owner-managed and micro businesses have in short supply.

The gap between employee and business reporting makes visibility especially important. The ONS found that 55% of employees reported using AI for work or education, compared with 35% of businesses reporting AI use. These are different surveys of different populations, so the figures cannot establish unauthorised workplace use. They do, however, support a more limited and useful inference: individual access and management awareness may not develop at the same pace (Office for National Statistics, 2026) ↗. A firm can therefore acquire AI-related exposure before it has consciously chosen a use case or a control model.

Sanctuary’s thesis is that SMEs should set a data boundary before trying to scale AI. This is not a new legal category, nor a replacement for sector-specific obligations. It is an operating rule that answers three practical questions: what information may enter a given AI service, what outputs may be used for, and when a named person must review or decide. The purpose is to make low-risk experimentation easier while preventing convenience from becoming the firm’s default risk policy.

Adoption figures reveal an integration and trust gap

The UK Business Data Survey shows why a headline adoption rate is an incomplete measure of readiness. Among businesses handling digitised data, 41% reported using AI-based technologies. Use was more common in large businesses than sole traders, and relatively high in information and communication. But only 21% of AI-using businesses said their tools were integrated with systems such as customer relationship management, finance or productivity platforms (Department for Science, Innovation and Technology, 2026) ↗.

Integration is not inherently desirable: connecting an untested tool to customer, financial or workforce systems can magnify errors and widen access to valuable information. Still, the gap between use and integration is consistent with AI being deployed around existing processes rather than changing the operating model. This is a reasonable early-stage strategy for SMEs. It limits commitment and gives staff time to learn. But it also means that a business may be measuring adoption rather than value.

The governance figures point to a related issue. Only 17% of AI-using businesses reported AI policy or guidance: 5% had a formal written policy and 12% informal guidance. Among those with policy or guidance, 62% included rules on AI access to business data and files (Department for Science, Innovation and Technology, 2026) ↗. The implication is not that every small firm needs a lengthy policy document. It is that staff need a decision they can apply when handling a client file, meeting transcript, supplier price list or employee information.

The survey’s findings on external model training sharpen the commercial rationale. Seventy-three per cent of businesses handling digitised data said they would be uncomfortable with business-owned data being used to train external AI models, compared with 18% who said they would be comfortable. The survey found no evidence that question order affected this response (Department for Science, Innovation and Technology, 2026) ↗. This result does not show that all suppliers train models on customer inputs, nor that all AI services create the same exposure. It does show that businesses want clear terms on retention, access, training and material changes to a provider’s service.

That is a control question rather than simple resistance to technology. For a local adviser, retailer, manufacturer or professional-services business, trust is an economic asset. A useful AI service that weakens customers’ confidence in how information is handled can destroy more value than it creates. Conversely, clear boundaries can enable staff to use suitable tools confidently for lower-risk work.

Why a policy and a licence do not produce productivity

A policy is a useful foundation, but it cannot settle the operational questions that determine whether AI is useful and safe. It does not itself decide whether a generated customer response needs checking, whether a recruitment output may influence a shortlist, or whether an AI feature newly added to existing software changes the firm’s information exposure. These are questions of workflow design, decision rights and supplier management.

The government’s AI Management Essentials guidance takes this broader view. Its SME-focused self-assessment covers internal processes, risk management and communication, and prompts businesses to consider an AI-system record, accessible policies, allocated responsibilities and third-party provider documentation. Crucially, the guidance presents completion as a starting point for better practice, not certification or proof of compliance (Department for Science, Innovation and Technology, 2026) ↗. That framing is sound: management capability is demonstrated in repeated decisions and working practices, not the existence of a document.

Academic evidence also cautions against treating software access as a productivity strategy. Research on high-technology ventures found that AI-related performance gains were associated with sufficiently intensive adoption and were stronger alongside complementary cloud and database investment and relevant internal R&D strategies (Lee et al., 2022) ↗. This study does not offer a direct formula for every UK SME, particularly those outside high-technology sectors. Its transferable lesson is narrower: durable returns depend on complementary organisational and technical investments, not on a licence alone.

Experimental evidence on generative AI helps explain the mechanism. In a multi-firm experiment, access to an integrated generative-AI tool reduced time spent on email and out-of-hours work, but did not by itself produce detectable changes in the quantity or composition of tasks (Dillon et al., 2025) ↗. For an owner-manager, recovering administrative time can still be valuable. But saved minutes become economic value only when they are intentionally redeployed: towards customer service, sales, quality improvement, training or work that had previously been deferred. Without that choice, AI may reduce friction while leaving the underlying business model unchanged.

Cyber practice provides a further warning against assuming that intention equals capability. Among UK businesses and charities using, adopting or actively considering AI, only around a quarter reported practices or processes to manage AI-related cyber risks, although many planned to introduce them within 12 months (Department for Science, Innovation and Technology and Home Office, 2026) ↗. A future plan is not an answer when staff must decide today whether to upload a file, enable a feature or act on a persuasive but incorrect output.

A proportionate data-boundary model for SMEs

Sanctuary recommends a proportionate data-boundary model. It is an implementation framework, not a regulatory standard, and the appropriate level of control should vary with sector, information and consequence. Using public product specifications to draft web copy is materially different from using AI in recruitment, customer eligibility, payment instructions or advice based on client records. The objective is not to prohibit experimentation; it is to separate routine assistance from uses that demand stronger safeguards.

First, maintain a live register of material AI uses. Include paid and free tools, embedded software features, outsourced services and employee-led practices that have become routine. For each use, record the business purpose, accountable owner, users, tool, information involved, expected output, affected decision and measure of success. This need not be a complex system. A well-maintained spreadsheet can give a small firm visibility that it otherwise lacks.

Second, make information categories operational. A practical starting set may distinguish public information, internal operational material, confidential commercial information, personal data and high-consequence records. The important test is whether a member of staff can recognise what cannot enter an unapproved external service and knows who can approve a legitimate exception. A category that staff cannot apply at the point of work is not a control.

Third, match human oversight to the consequence of error. A social-media draft should not receive the same treatment as an invoice-payment instruction, technical specification, customer decision or recruitment recommendation. Where an output could materially affect people, finance, safety or contractual commitments, a named person should have the authority, time and expertise to verify it, reject it and correct the process. Nominal human review is weak when it becomes a rushed approval of content the reviewer cannot meaningfully assess.

Fourth, redesign one bounded workflow at a time. Choose a process with an observable baseline, such as standard customer responses, meeting summaries, internal knowledge retrieval or first-pass marketing content. Assess cycle time, rework, error rates, customer response, staff experience and the use of released capacity. This treats AI as a business experiment rather than a usage target. It also gives firms evidence to stop, amend or scale a use case before it becomes embedded.

Finally, treat suppliers and training as part of the control environment. Establish which AI features are already present in business software, what information controls are available, what provider documentation exists and how material feature changes will be communicated. Training should focus on verification, escalation, information boundaries, customer communication and workflow redesign—not prompting alone. Only 11% of businesses with 10 or more employees reported that more than half their workforce had received AI-related training (Office for National Statistics, 2026) ↗. Better prompts may improve an answer; better judgement reduces the chance that a fluent but unreliable answer becomes a business decision.

The policy priority is implementation capability

One year after launch, the government reported that 38 of the 50 actions in the AI Opportunities Action Plan had been completed (Department for Science, Innovation and Technology, 2026) ↗. That is a relevant indicator of programme delivery, but it is not evidence that ordinary firms are converting AI availability into productive, reliable and trusted services. The local economic return will depend on implementation inside businesses.

This suggests a shift in emphasis for business-support programmes, sector bodies, accountants, lenders and supply-chain leaders. Subsidising licences or promoting generic adoption may be useful, but it is insufficient. Greater value may come from helping firms identify workflow bottlenecks, establish information boundaries, assess suppliers and define credible measures of success. These are practical management tasks, particularly for high-street and local-service businesses that have limited capacity to evaluate technology yet rely heavily on customer trust and responsive service.

There is a valid counterargument. Excessive process can turn a modest productivity tool into administrative drag, especially in microbusinesses. The answer is not to imitate large-company compliance structures. It is to use simple defaults and escalate the documentation, approval and review required as the sensitivity of information and consequence of error increase.

For an individual SME, the first step is modest but demanding: identify the AI already in use, set a data boundary, select one workflow with measurable value and make human decision rights explicit. This will not guarantee a return on investment. It does give the business a disciplined way to discover where returns are real—and to avoid allowing unexamined convenience to determine its exposure to commercial, customer and operational risk.