Many UK firms have already tried a chatbot that answers questions from a prompt. The next offer is different. Vendors are selling systems that plan a task, open software, and act across a workflow that used to need a person moving between applications. That shift, usually called agentic AI, can remove dull coordination. It also raises a plainer question: who is accountable when the system does the wrong thing?
The gap is no longer theoretical. Research published in 2026 suggests that deployment plans are running well ahead of the rules firms have in place for agents. A business that switches an agent on because a competitor might, and only later asks what it was allowed to touch, is governing the incident rather than the tool.
What Deloitte measured
Deloitte's 2026 State of AI in the Enterprise report found that nearly three-quarters of surveyed companies planned to deploy agentic AI within two years, while only 21% said they had a mature model for agent governance. The survey, conducted in August and September 2025, included 3,235 business and IT leaders across 24 countries. Deloitte also found that only 25% of respondents had moved at least 40% of their AI pilots into production. Adoption intent and operational readiness are not moving together.
Those figures are from a global survey of leaders, not a census of UK small firms. They are still a useful warning. A large share of organisations expect agents soon. A much smaller share say they can govern them. Deloitte's writing on agents scaling faster than governance makes that the point of the research, and the wider State of AI series sets the same finding in the context of pilots that have not yet become day-to-day operations.
Governance is a set of decisions, not a PDF
Publishing an AI policy is not the same as running one. An agent that can use tools needs answers a policy page often leaves blank. Which actions may it take without asking? Which actions need a person to approve them first? What data may this agent see, and which mailboxes, drives and customer records stay closed? Where is the record of what it did, and who reads that record?
The awkward cases need a plan as well. An agent will eventually draft from a stale document, call a tool nobody expected, or produce a result that looks tidy and is wrong. The team should know who stops it, who tells the people affected, and what gets switched off while that is checked. Without those decisions, "human oversight" is a sentence in a slide, not a step in the working day.
- Independent actions: listed, and limited to work a person can reverse.
- Approval steps: named for anything that sends, pays, publishes or changes a live record.
- Data access: the smallest set of sources that can complete the trial.
- Logs: a record a colleague can read the next morning, including tool calls.
- Errors: a named owner, and a way to pause the agent without a project meeting.
Start with a process you can measure
A practical first step is a low-risk process with a clear edge and a result you can count. A company might ask an agent to draft an internal report from a folder of approved documents, and require an employee to review it before it is shared. That trial can be scored. Was the draft accurate? How long did the review take, compared with writing the note from scratch? Did the agent ask for a file it did not need? What errors showed up, and were they obvious?
Expand only after those answers are boring in a good way. A second workflow is justified when the first one stays inside its permissions and the reviewer trusts the log. It is not justified by a vendor roadmap. Google's Gemini agent is one of the products now offering multi-step work across workplace apps. The governance questions do not change because the logo does. The same applies to the security risks of letting an agent follow instructions it finds in a document or a webpage, covered in AI agent security for business workflows.
Oversight can make adoption easier
Governance is often described as a brake. Used properly, it is how a firm finds out which agent is worth keeping. Clear responsibilities, access controls, a test on real work, and an audit log tell you what the system is doing and where a person's judgement still has to sit. A team that can answer those points can widen the trial. A team that cannot should stay on the one workflow it understands.
For UK firms, personal data adds a legal edge that a global survey will not spell out for you. Enquiry text, staff records and customer files are personal data, and UK GDPR applies to them whether a person or an agent opens the file. Knowing the purpose, the systems and the log is part of using the tool, not a separate compliance project you start after go-live.
Put the boundary in the build
An agent bolted onto every login is harder to govern than a feature built around one job. AI development here means a defined task, approved sources and a handover to a person. Maintenance and support is how that boundary stays intact after launch, when permissions creep and the source documents change.
Web Works Rise is a UK-led team in Birmingham, with a development hub in Tunisia for delivery. If you want to try an agent on one internal workflow and keep a colleague as the person who approves the outcome, contact the team with the task and the systems it would need to see. The survey says many firms plan to move. It does not say they are ready.