In the earlier articles, I considered a hypothetical 20-person equipment servicing business introducing AI-assisted service reports. Imagine an office colleague using a readily available AI tool to turn an engineer’s notes into something suitable for a customer. The draft is ready in minutes. Unfortunately, the notes also include the customer’s contact details, a site address and information about a technical fault. Nobody has checked whether the tool is approved or how it handles the information.
There is no reason to assume the employee intended to do anything wrong. The aim was to complete a job more efficiently. But that does not remove the business’s responsibility to understand where information goes, whether the output is reliable and who approves the result. The precise exposure would depend on the product, its settings and the data involved.
AI use is moving ahead of company guidance
Deloitte surveyed 25,000 UK workers in May and June 2026. It reported that 63% had knowingly used generative AI for work. Among those using it, 31% said they used AI without their employer’s knowledge. These are self-reported findings; they do not establish the position in any particular company. They do, however, suggest that an absence of formal AI projects should not be mistaken for an absence of AI use. [1]
The government’s UK Business Data Survey 2026 found that only 17% of AI-using businesses reported a formal or informal AI policy or guidance. Just 5% had a formal written policy. This was a different survey, covering businesses that handled digitised data, and its figures should not be directly compared with Deloitte’s workforce results. Nevertheless, the lack of organisational guidance is striking. [2]
For a founder or managing director, the first question may therefore be less about the next licence purchase and more about what staff are already using, for which tasks and with what information.
Risk does not begin and end with confidential data
Confidentiality is an obvious concern. Customer records, employee details, pricing arrangements and commercially sensitive documents may be subject to contractual, security or data protection restrictions. Whether material is retained, reviewed, used for model improvement or processed overseas depends on the provider’s arrangements and the settings in use. A business needs to establish those facts rather than assume that every AI service handles information in the same way.
There are other risks. AI may produce a polished report containing a technical error, omit a qualification from a quotation or provide an inaccurate summary of an agreement. If an AI agent can access business systems and act on behalf of an employee, the consequences extend beyond generating text: permissions might permit messages to be sent, records changed or transactions initiated.
The Information Commissioner’s Office makes clear that AI use involving personal data needs appropriate security and data minimisation, with measures proportionate to the risk. Its detailed AI guidance is being reviewed following the Data (Use and Access) Act, so current requirements should be checked for each relevant use case. [3]
Who is allowed to use the tool, with which information, for what purpose, and who checks or approves the result? If those questions have no clear answers, the process is not ready for unsupervised AI use.
Different work requires different safeguards
I would resist a single rule for every use of AI. Drafting a short internal note from public information is very different from analysing employee performance, preparing a customer price or allowing software to update live records. A small business can make sensible distinctions without creating a large compliance department.
For lower-risk tasks using public or approved non-sensitive information, a company-approved tool and ordinary accuracy checks may be enough. Work involving confidential records should use explicitly authorised arrangements, appropriate access controls and a named reviewer. Decisions with significant financial, legal, employment or safety consequences require stronger assessment and human accountability; some proposed uses may simply be unsuitable. Where personal data processing is likely to be high risk, a data protection impact assessment may be required.
These are practical categories rather than universal legal classifications. The required safeguards depend on the information, the process, the consequences of an error and the applicable contractual or regulatory obligations.
Five things a smaller business can put in place
For a business employing perhaps ten to fifty people, I would start with five straightforward measures:
- Establish what is already happening. Ask employees which tools help, which tasks they are using them for and where they are uncertain. The aim is to understand the activity, not to discourage honest answers.
- Agree approved tools and information boundaries. Define permitted use, restricted information, account settings and situations that need prior approval. Check supplier terms, retention, access, subcontractors and incident arrangements.
- Name the person responsible. A manager should own AI usage decisions and maintain a simple register of approved applications. The person checking an important output should understand the underlying work, not just the tool.
- Provide short, practical guidance. Staff need examples of safe inputs, expected checks, common errors and how to report an accidental disclosure or unreliable result. A short session and usable guidance may achieve more than a lengthy policy.
- Review the arrangements as use grows. Test representative outputs, restrict system permissions, monitor incidents and revisit approvals when tools, integrations or business processes change.
The National Cyber Security Centre recommends considering security throughout AI system design, deployment and ongoing operation. The UK government’s voluntary AI Cyber Security Code of Practice also addresses supplier assurance, access controls, testing and meaningful human oversight. These are useful reference points, particularly when a company moves beyond a standalone assistant into integrated workflows. [4] [5]
Governance should make responsible use easier
It is tempting to respond to unfamiliar technology with a blanket prohibition. That may sometimes be appropriate for specific uses, but it is unlikely to resolve the wider problem if staff are already finding the tools useful. Equally, buying an enterprise licence does not settle questions about accuracy, confidentiality or who is authorised to make a decision.
For a small company, I would aim for a short written policy, a small list of approved tools, named responsibility and a clear way to raise concerns. More formal controls can follow where the sensitivity or complexity of a use case justifies them. The objective is not paperwork for its own sake; it is to let people improve their work without unknowingly creating risks for the business or its customers.
My view is that good AI governance should be seen as part of successful adoption rather than a separate obstacle to it. An organisation that understands both the capability and its limits is in a better position to make useful changes with confidence.
The final article will consider what happens after a successful AI pilot: how businesses can embed AI into everyday operations, establish clear ownership, sustain measurable benefits and identify further opportunities for improvement.
Sources and methodological notes
[1] Deloitte UK, GenAI Workforce Survey (16 September 2026). Source — Survey of 25,000 UK workers, May–June 2026; self-reported use and employer awareness.
[2] DSIT, UK Business Data Survey 2026 (18 June 2026). Source — AI policy figures relate to 1,870 AI-using businesses; survey fieldwork October 2025–January 2026.
[3] ICO, Guidance on AI and data protection: security and data minimisation. Source — Guidance under review following the Data (Use and Access) Act; check current obligations.
[4] NCSC, Guidelines for secure AI system development (27 November 2023). Source — Aimed mainly at providers; relevant to secure deployment and operation of integrated AI systems.
[5] DSIT, Code of Practice for the Cyber Security of AI (31 January 2025). Source — Voluntary code, not a general statutory duty; covers human responsibility and system security.
Illustrative scenario: the service-report example is hypothetical and does not describe a client incident or an observed data breach.
