Only 21% of firms in Deloitte's survey have mature governance for AI agents. Who owns each agent, what it may decide alone and what managers need to learn.
AI agents are moving from pilots into everyday work. They answer customer queries, update records, reconcile invoices and hand finished work to people. Once an agent can act on its own, a team has a new kind of member, and someone has to manage it.
Few organisations have settled who that is. In Deloitte's State of AI in the Enterprise survey of 3,235 business and technology leaders in 24 countries (January 2026), 74% expected their companies to use AI agents at least moderately by 2027. Only 21% said they had a mature governance model for agentic AI.
This piece is for COOs, HR and L&D directors and line managers whose teams already work with agents or soon will. It covers who answers for an agent's work, how fast agent use is growing and what managers need to learn.
When an AI agent joins a team, work starts to be done by software that takes actions, and each action needs an owner. An AI agent is software that carries out multi-step tasks across business systems: it can process a refund, change an order, draft and send a reply or move data from one system to another.
Because an agent can take the next step itself, three questions move onto the team: which tasks the agent may do alone, which need a person's approval, and who checks the result.
Our decision matrix for your first AI agent covers how to choose the first process. This article picks up after that choice, when the agent is live and people work alongside it.
The organisation that deploys an AI agent stays accountable for what the agent does, and inside the organisation each agent needs a named owner. The UK Information Commissioner's Office states in its Tech Futures report on agentic AI that "organisations remain responsible for data protection compliance of the agentic AI they develop, deploy or integrate" in their systems and processes.
The ICO also lists "determining controller and processor responsibilities through the agentic AI supply chain" as a new risk, and warns about systems with "no measures in place to secure access, monitor or stop activity". In practice, that means five questions need a name next to them before an agent goes live:
| Question | Who usually answers it |
|---|---|
Which tasks may the agent complete without a person? | The owner of the business process |
Which decisions need a person's approval first? | The process owner, with risk or compliance |
Which data and tools can the agent reach? | IT and the data protection officer |
Who reads the agent's activity log, and how often? | The line manager of the team using it |
Who can pause or stop the agent? | A named person in the team, plus IT |
Deloitte describes the same building blocks: clear limits on which decisions an agent makes on its own and which need human approval, and audit trails that record what it did. Our guide to data governance before or after AI covers the data side in more detail.
AI agent use is growing fastest in large companies, while governance is still catching up. McKinsey's State of AI survey (1,719 respondents in 97 countries, in the field May to June 2026, published August 2026) reported the following:
| Finding (McKinsey, State of AI, August 2026) | Figure |
|---|---|
Companies with more than $1 billion in revenue scaling agents in at least one function | 40%, up from 27% a year earlier |
Smaller organisations scaling agents | 22%, unchanged |
Respondents who say AI agents have reached the scaling phase across their organisation | About two in ten |
Functions where agents are most often scaled | IT, knowledge management, software engineering |
Read alongside Deloitte's figures, the gap is clear: three in four leaders expect agents at work by 2027, and about one in five reports a mature governance model for them.
Managers who lead a team with AI agents need to delegate to software in the same disciplined way they delegate to people: a clear brief, agreed limits, regular checks and a plan for when things go wrong. Microsoft's 2025 Work Trend Index (31,000 workers in 31 markets, including the UK) found that leaders expect their teams to be training agents (41%) and managing them (36%) within five years. Microsoft calls the person who builds, delegates to and manages agents an "agent boss".
In practice, that comes down to five skills:
These are management skills applied to a new kind of colleague, and HR and L&D teams can add them to existing manager development programmes.
You'll find more names on our artificial intelligence speakers page.
Sol Rashidi at TEDxNewAlbany:
How to avoid becoming irrelevant | Sol Rashidi | TEDxNewAlbany (YouTube)
An organisation should start AI agent governance with a list: every agent in use or in pilot, its owner and its limits. A one-page register is enough to begin with. It gives managers, IT and compliance one shared view, and it shows quickly where an agent has no owner.
From there, two related guides help. From AI access to AI value covers how to judge whether AI use is paying back, and managing AI vendor lock-in covers the supplier terms that shape what an agent can reach.
An AI agent in the workplace is software that carries out multi-step tasks across business systems on its own, such as processing a refund, updating a record or sending a reply. A person sets its limits and checks its work.
The organisation that deploys an AI agent is responsible for its decisions. The UK ICO states that organisations remain responsible for data protection compliance of the agentic AI they develop, deploy or integrate, so each agent should have a named owner inside the business.
AI agents need a named person who sets their limits, reviews their output and can stop them. In Microsoft's 2025 Work Trend Index, 36% of leaders expected their teams to be managing agents within five years.
Managers working with AI agents need five skills: writing a clear task, checking a sample of the output, splitting each role between people and agents, explaining the changes to their team and agreeing in advance when an agent should be paused.
About one in five organisations has mature governance for AI agents. In Deloitte's January 2026 survey of 3,235 leaders, 21% reported a mature governance model for agentic AI, while 74% expected to use agents at least moderately by 2027.
After Dinner Speakers
Contents
