AI Governance Framework: Questions to Ask Before Scaling AI

An AI governance framework sets what your AI tools and agents may access, which decisions they may make or influence, where a person must approve and how the results are checked. Once AI moves from pilots into daily workflows, those choices become part of running the business, and each one needs a named owner.

In the government's UK Business Data Survey 2026 (June 2026), 58% of medium-sized businesses (50 to 249 employees) that handle digitised data said they use AI, but only 22% of medium-sized businesses using AI had a formal, written AI policy. The gap matters more once AI acts as well as advises: an agent that can update records or issue refunds holds permissions that need the same discipline as any employee's. In Deloitte's August 2026 survey of 501 US leaders at organisations already piloting or using AI agents, 70% said they did not yet feel able to trust and govern them.

Why AI governance becomes an operational issue

An isolated experiment needs little more than rules on which tools to try. A pilot using real data or connected systems already needs the controls below, and once AI sits inside a workflow, it reads live data, drafts work for customers and, increasingly, acts. The risk moves from what a tool says to what it does.

In IBM's Cost of a Data Breach Report (July 2026), more than 20% of the 602 breached organisations studied reported a breach targeting AI models or applications, most often through the systems around them, such as compromised APIs or plug-ins. OWASP calls the agent version excessive agency: damaging actions caused by unexpected or manipulated output, usually because the system had more permissions or autonomy than it needed. Its Top 10 for Agentic Applications (December 2025) adds tool misuse and identity and privilege abuse.

Governance is the control layer around the capabilities the first article in this series described. For the wider reasons AI projects stall, read why most AI implementation fails.

What the law requires, and what is good practice

The table summarises the position as at September 2026; take advice on how each applies to you.

The UK position. The government's 2023 AI regulation white paper set five principles for existing regulators to interpret and apply, and the King's Speech of 13 May 2026 did not include an AI bill. The ICO's guidance on AI and data protection is under review after the Data (Use and Access) Act 2025. Its published plans put new agentic AI guidance in autumn 2026 and updated automated decision-making guidance in winter 2026, and it plans to consult on a statutory AI and automated decision-making code in 2027.

The EU position. Prohibited practices and the AI Act's literacy duty have applied since 2 February 2025, the rules for general-purpose AI models since 2 August 2025, and most other provisions since 2 August 2026. The Digital Omnibus on AI, adopted as Regulation (EU) 2026/1744 and in force since 27 July 2026, moved the high-risk obligations to 2 December 2027 for uses such as employment and critical infrastructure, and to 2 August 2028 for AI built into regulated products. It also relaxed the literacy duty: providers and deployers must still take measures to support their staff's AI literacy, but no set level is required. The Commission's AI Act Service Desk confirms that AI agents fall under the same rules, including the transparency rules from 2 August 2026 where an agent interacts with people or generates content.

Good practice. ISO/IEC 42001:2023 sets requirements for an AI management system that can be certified. The voluntary NIST AI Risk Management Framework has four functions: govern, map, measure and manage. For agents, the NCSC and international partners published joint guidance in May 2026, Careful adoption of agentic AI services, which advises starting with tightly bounded, low-risk pilots and points to the ETSI EN 304 223 standard for baseline AI security requirements. Adopt their controls, sized to your risk, before deciding on certification.

Five areas leaders should control

Each needs an owner, a rule and a check.

Data: what should AI be allowed to see?

An AI assistant connected to shared drives, email or a CRM usually inherits its user's permissions, so old sharing links and over-broad folders become part of what it can read and repeat. Decide which data each AI use may touch, and keep special category data such as health information out unless the use needs it, you have both a UK GDPR Article 6 lawful basis and an Article 9 condition (plus a Data Protection Act 2018 Schedule 1 condition where required), and your DPIA supports it. Give each data set an owner who fixes errors at source. In the UK Business Data Survey 2026, 62% of businesses with an AI policy said it covered AI access to their data and files.

Access: what may people and agents do?

Apply role-based access and least privilege to AI agents as you do to people: give each agent its own identity, scope its permissions to the task, separate reading from writing, and review access when the task changes. NIST's identity specialists warned in August 2026 against handing agents people's credentials or long-lived API keys, and the NCSC advises giving agents the minimum access for the shortest time, with temporary credentials revoked when a task is complete. OWASP recommends requiring a person to approve high-impact actions before they are taken, and warns of prompt injection, where instructions hidden in an email, web page or document change what the model does. An agent that reads outside content should not also hold the power to make payments or delete records.

Decisions: who owns what AI decides or influences?

Each AI-enabled workflow needs one owner, accountable for its outcomes. Sort its decisions into three kinds: AI decides within set limits, AI recommends and a person decides, or AI only informs. Where a significant decision about someone is made from their personal data with no meaningful human involvement, the safeguards in Article 22C of the UK GDPR apply: information about the decision, a chance to make representations, human intervention and a way to contest it. The ICO's guidance also says that if you use AI in decision-making, you need to be able to show it does not result in discrimination under the Equality Act 2010, which covers Great Britain; Northern Ireland has its own equality legislation.

Oversight: where must a person approve?

As good practice, keep human approval mandatory where a decision has a legal or similarly significant effect on a person, moves money above a set threshold, commits the business to a customer or supplier, changes master data such as bank details, or cannot easily be reversed. The law is narrower: it permits most solely automated significant decisions if the Article 22C safeguards are in place. Write your thresholds into the workflow, and, as the NCSC recommends, decide before an agent goes live who owns it, who approves its access, who monitors it, who reviews incidents and who can stop it. Approval only protects you if the reviewer has the time, the information and the authority to say no.

Monitoring: what should you watch after launch?

AI systems change after launch: vendors update models, data drifts and staff find new uses. Decide before go-live what you will measure and who reviews it: a weekly sample of outputs checked against source, error and override rates, complaints, the actions agents took and the data they touched, and every incident. Review each use at least quarterly and whenever the model, data or task changes, and make sure your incident plan covers an agent that fails, is misused or must be stopped. The NCSC's test: if you cannot understand, monitor or contain an agent's actions, it is not ready for deployment.

The human layer around AI

Governance is carried out by people, so it needs roles, time and training. A senior sponsor, often the COO, owns the framework and the register of AI uses; each workflow has an accountable business owner; reviewers check outputs and approve decisions above the thresholds; data owners keep records accurate; and a small leadership group approves new high-risk uses and reviews incidents.

Everyone who uses AI needs to know which tools are approved, what data must never go into them, how to spot a wrong answer and how to report a problem. Within the EU AI Act's scope, supporting staff AI literacy is also a legal duty. Build review into job design: in our experience, controls break down where nobody has the time or the authority to apply them.

Governance for distributed and offshore teams

A distributed team needs the same framework, applied with the same rigour in every location. Four controls matter most:

• Access. Offshore colleagues work in your systems under your identity and access controls and least privilege, using approved AI tools under the same data rules as UK staff.

• Approval. Apply the same thresholds everywhere: an offshore team can run a process and resolve exceptions within your limits, escalating decisions above them to the named owner in the UK.

• Monitoring and logging. Logs and output checks cover every location in one view.

• Training. Everyone completes the same AI and data training before they get access, and again when the rules change.

Under the Embedded Offshore Team Model, your team works in your systems, to your policies and reporting into your managers, and four things make these controls easier to hold:

• One employer. Your team is employed by us under local employment contracts in each hub, so your data and AI rules reach everyone through one employer, not a mix of agencies and freelancers.
• On-site supervision. Teams work from our offices with local leadership on site, so how people use data and AI tools is supervised in person.
• HR support. Our local HR business partners support performance, conduct and development, so your AI rules are backed by the same people processes as any other policy.
• Consistent onboarding. Every new starter follows the same structured onboarding into your systems, tools and policies, including your rules and training on AI.

Offshore access and UK GDPR. Under the ICO's three-step test, making personal data held in the UK accessible to people outside the UK is a restricted transfer when they work for a separate legal entity, which is the position when we employ your team. How the transfer is covered depends on where the team works.

Since the Data (Use and Access) Act's transfer changes took effect on 5 February 2026, UK law calls this assessment a data protection test: acting reasonably and proportionately, you must decide that protection for people's information will not be materially lower than in the UK after the transfer. The ICO still calls it a transfer risk assessment. Take advice on the right mechanism before anyone is given access; we agree the arrangements with you before a team starts. See also offshore locations for technology teams, offshore service desks for MSPs and offshore workplace culture.

A practical leadership checklist

Ask these questions before a pilot uses real data or connected systems, and again before it scales. Any questions you cannot answer is your next piece of governance work.

1. Do we know every AI tool and agent in use? Keep a register of owners, purposes, data and decisions.
2. Is each one's access the minimum its task needs? Review permissions for people and agents alike.
3. Who owns each AI-enabled workflow? Name one person accountable for its outcomes.
4. Where must a person approve, and at what threshold? Build it into the workflow and give reviewers the authority to refuse.
5. Have we completed a DPIA wherever AI processes personal data? Assume one is needed unless an assessment shows otherwise.
6. Do any significant decisions about people run without meaningful human involvement? If so, confirm the UK GDPR safeguards and take advice.
7. Could any use discriminate? Test before launch and monitor afterwards.
8. Does any AI output reach the EU? Check whether the EU AI Act applies and which dates affect you.
9. What will we monitor after launch, and who reviews it? Sample outputs, track errors and agent actions, and review quarterly.
10. Is everyone trained on the same rules, in every location? Make training a condition of access, and cover transfers first.

If part of your team works outside the UK, book a discovery call to discuss how we support the people, processes and operational infrastructure around distributed teams.

Frequently asked questions

What is an AI governance framework?
An AI governance framework is the set of rules, owners and checks a business uses to control AI. Its key components cover what data AI tools and agents may access, which decisions they may make or influence, where a person must approve and how results are monitored. For a mid-sized business it can be a short policy, a register of AI uses and named owners.

What are the five principles of AI governance in the UK?
The UK government's five cross-sector principles are safety, security and robustness; appropriate transparency and explainability; fairness; accountability and governance; and contestability and redress. They were set out in the 2023 AI regulation white paper and confirmed in 2024 for existing regulators such as the ICO to apply. They are guidance for regulators and have no statutory force of their own.

Is there an AI law in the UK?
There is no general AI statute in the UK as at September 2026, and the King's Speech of May 2026 did not include an AI bill. AI is governed through existing law, chiefly UK GDPR and the Data Protection Act 2018 as amended by the Data (Use and Access) Act 2025, and equality law: the Equality Act 2010 in Great Britain and separate legislation in Northern Ireland. Take advice on how these apply to your use of AI.

Does the EU AI Act apply to UK businesses?
It can. The EU AI Act applies to a UK business that places an AI system on the EU market or puts one into service there, or whose AI system produces output used in the EU. Prohibited practices and the AI literacy duty have applied since 2 February 2025, and most other provisions, including the transparency rules, since 2 August 2026. After amendments that took effect in July 2026, the high-risk obligations apply from 2 December 2027, or 2 August 2028 for AI in regulated products.

Can offshore team members use AI tools with UK personal data?
Yes, under the same controls as UK staff and with the transfer covered. Where the team is employed by a separate legal entity abroad, their access to UK personal data is a restricted transfer under UK GDPR. Romania is covered by UK adequacy regulations; South Africa, India and Brazil need a safeguard such as the IDTA or the Addendum, plus a transfer risk assessment. Take advice before access is granted.

Sources: DSIT, UK Business Data Survey 2026 (June 2026); Deloitte, agentic AI survey (August 2026); IBM, Cost of a Data Breach Report 2026 (July 2026); OWASP, Excessive Agency, Prompt Injection and Top 10 for Agentic Applications (December 2025); ICO, guidance plans (September 2026); European Commission, AI Omnibus (July 2026); AI Act Service Desk, AI agents FAQ; NIST, agent identity (August 2026); NCSC, agentic AI guidance (May 2026); ICO, international transfers (January 2026).