Managing AI agents: the roles and accountabilities every business needs

AI agents are software that can plan, decide and act across several steps with limited human prompting. Adoption is already broad. PwC's AI Agent Survey of 300 senior executives in May 2025 found 79% saying AI agents are already being adopted in their companies. McKinsey's state of AI survey, published in November 2025, found 62% of respondents at least experimenting with agents and 23% scaling an agentic system. The language of pilots is giving way to the language of workforce.

Scale is thinner than the headlines suggest. McKinsey reports that in any given business function no more than 10% of respondents say their organisations are scaling agents, and that most of those scaling do so in only one or two functions. PwC found that 68% of executives say half or fewer of their employees interact with agents in everyday work. Agents are present in most large organisations and central to few of them. That is the window in which management design is cheapest.

Expectations are rising quickly. Deloitte's State of AI in the Enterprise 2026 reports that 74% of respondents expect their companies to use AI agents at least moderately by 2027. PwC finds 67% of executives agreeing that AI agents will drastically transform existing roles within the next 12 months. Gartner expects at least 15% of day-to-day work decisions to be made autonomously by agentic AI by 2028, up from 0% in 2024. If those expectations are even half right, decisions now made by named employees will soon be made by software, and someone must answer for them.

For a CHRO the implication is that the workforce plan now has two kinds of worker. One is employed, appraised and developed through HR processes. The other is configured, deployed and monitored through technology processes, and rarely appears in the organisation chart. The gap between them is where accountability is lost. An agent that approves a refund, drafts a contract clause or routes a candidate has taken a decision. Who answered for it yesterday, and who answers tomorrow?

Employees are already adjusting. KPMG's pulse survey for the second quarter of 2026 found only 2% of leaders reporting employee resistance to AI agents, which suggests the barrier is no longer acceptance but accountability. People will use agents; the question is who checks them. Where HR has not defined that, managers improvise, and the improvisations differ between teams, which makes any later audit of agent decisions slow and contested.

What to do: ask technology, operations and finance for a list of every agent in use or in pilot, with its purpose, data access, the decisions it takes and the person who would be called if it failed. Expect the first list to be incomplete. Treat the gaps as the first finding, and agree a single register as the basis for everything that follows.

Gartner's June 2025 prediction is blunt: over 40% of agentic AI projects will be canceled by the end of 2027. Gartner also estimates that only about 130 of the thousands of agentic AI vendors are real, a reminder that supply is uneven. Managers should expect some agents they buy or build to disappoint, and should design the supervision model so that a disappointing agent can be stopped without drama.

Gartner's poll of 3,412 webinar attendees in January 2025 shows the shape of the market: 19% had made significant investments in agentic AI, 42% conservative investments, 31% were taking a wait-and-see approach and 8% had made none. Most organisations are therefore testing rather than committing. Testing without ownership is how pilots become permanent, unmanaged and unmeasured, and how costs accumulate without a decision.

Returns are also still limited. McKinsey reports that just 39% of respondents attribute any enterprise-level EBIT impact to AI, and most of those say AI accounts for a small share of EBIT. High performers are about 6% of respondents. PwC's adopters report gains: 66% increased productivity, 57% cost savings, 55% faster decision-making and 54% improved customer experience. These are self-reported benefits from adopters, not audited returns, but they show where value is expected to come from.

Supervision connects these facts. A cancelled project is often a project nobody owned: its value was never defined, its errors were never counted and its costs were never tied to results. A named owner with a measure and a budget is what keeps an agent from joining the group Gartner expects to be cancelled. Supervision is therefore not an overhead on top of the investment; it is part of how the investment earns its return.

A further reason to supervise early is cost transparency. Agents consume compute, licences and the time of the people who correct them. Early pilots hide those costs inside project budgets, and the true cost per task becomes visible only when the pilot is extended. An owner who is asked to report cost per completed task from the start will make better decisions about scale, and will have the evidence to defend a good agent or to retire a poor one.

What to do: for each live agent, write one sentence on the value expected, the measure that will show it and the date at which the owner will decide to scale, change or stop it. Review the list quarterly with finance. Retire agents that fail the test, and report the decision openly so the organisation learns that stopping is a legitimate outcome.

Deloitte's 2026 State of AI in the Enterprise survey found that only 21% of respondents say their organisations have a mature governance model for agentic AI. The same survey finds 74% expecting at least moderate use of agents by 2027. Deployment and control are moving at different speeds. When an agent acts without a clear human owner, the first the organisation learns of a problem may be a customer complaint.

McKinsey reports that respondents now act to manage an average of four AI-related risks, compared with two in 2022, and that 51% of respondents from organisations using AI have seen at least one negative consequence, with nearly one-third reporting consequences from inaccuracy. Organisations are getting better at naming risks. The harder step is assigning each risk to a person with the authority to act on it.

In Europe, regulation is making the point explicit. For high-risk AI systems, Article 26(2) of the EU AI Act says deployers shall assign human oversight to natural persons who have the necessary competence, training and authority. Article 14(4) requires oversight measures that let a person decide not to use the system, intervene in its operation or interrupt it through a stop button. Not every agent is a high-risk system, but the standard is a useful template for any agent that takes consequential decisions.

The principle applies outside Europe as a matter of good management. Competence, training and authority are three separate conditions. A named person without competence cannot judge the output. One without training will not know what to look for. One without authority cannot stop the agent when the evidence demands it. CHROs can influence all three, since each is a people-system issue: selection, development and decision rights.

Employment and data protection law add a people dimension. Agents that screen candidates, allocate work or assess performance touch areas where employees and regulators expect human accountability and explanation. HR should review each such agent with legal advisers before launch, and should record who may explain and, where required, reverse its decisions. Doing this at the design stage is considerably cheaper than rebuilding a process after a complaint.

What to do: take the agent register and add three columns: the person assigned to oversight, the evidence of their competence and training, and the written authority to pause or switch off the agent. Where any column is empty, treat the agent as ungoverned. Involve legal and risk early, because job descriptions and employment terms may need to change, and because some uses will carry regulatory obligations.

Our position is simple: no agent should run without a named human owner. An owner holds three things: the objective the agent serves, the budget it consumes and the right to change or stop it. Ownership is not technical. The owner need not understand the model, but must understand the work. For a recruitment-screening agent the owner is a talent acquisition leader; for an invoice-matching agent it is a finance operations manager.

Decision rights come next. For each agent, define what it can decide alone, what it can recommend for human approval, and what it must never do. The boundaries should be written in plain language and reviewed whenever the agent's tools, data or objectives change. With Gartner expecting at least 15% of day-to-day work decisions to be made autonomously by agentic AI by 2028, the boundary between deciding and recommending will become one of the most important documents a business holds.

A short charter is enough. One page per agent: purpose, owner, supervisor, data and systems accessed, decisions permitted, decisions reserved, escalation route, performance measures and review date. The charter is signed by the owner and the line manager above them. It is kept in the register and reviewed at least quarterly. The discipline matters more than the format, and a charter that is read is better than a framework that is admired.

Why not leave ownership with the technology team? Because technology can maintain a system but cannot judge whether its output is acceptable for the business purpose. Technology should steward the agent: access, logs, versions and updates. The business should own its outcomes. Mixing the two produces the familiar failure in which each side assumes the other is watching.

Decision rights also have to survive change. Agents are updated, retrained and given new tools, and each change can move the boundary of what they do. Make the owner the gatekeeper for any change that expands permissions, data access or autonomy, with the steward recording the change in the register. A simple rule works well: no new permission without a dated, owner-approved entry. It keeps the charter true and gives auditors a trail.

What to do: assign an owner to every agent within thirty days, starting with those touching customers, employees, money or regulated data. Use the charter template, require both signatures and publish the register to the leadership team. Where no suitable owner exists, pause the agent until one is found. That rule, applied once, will change behaviour across the organisation.

Four roles carry most of the load. The agent owner is accountable for outcomes, cost and risk. The agent supervisor reviews the work daily or weekly, samples outputs and spots drift. The exception handler resolves cases the agent cannot or should not decide. The agent steward, usually in technology or data teams, maintains access, logs and versions. Small organisations can combine roles; none should be left empty.

These roles change the work of existing jobs. PwC found 67% of executives agreeing that agents will drastically transform existing roles within the next 12 months. The World Economic Forum's Future of Jobs Report 2025 says workers can expect 39% of their existing skill sets to be transformed or become outdated over 2025–2030, that 85% of employers plan to prioritise upskilling and that 63% identify skill gaps as a major barrier. Supervision is a new skill set that sits inside that picture.

Capacity is the practical question. A supervisor reviewing an agent's output needs time in the diary, a sampling method and tools that show what the agent did. If supervision is added to a full workload, it will be the first task dropped. Count supervision hours in workload planning, as for any other responsibility, and reflect them in grading and pay. Roles that carry accountability for a digital worker are not clerical.

Career paths complete the design. A strong exception handler is learning the edge cases that define the work. A supervisor learns how the agent fails. Both are building the knowledge needed to become an owner. HR can formalise this by recognising agent supervision as a developmental step towards management, rather than as a task for whoever is available. People take the roles seriously when the organisation does.

Selection and development deserve attention as well. Good supervisors are rarely the most senior people; they are the practitioners who know the work well enough to see when output looks plausible but wrong. HR should identify them, give them time to supervise and protect that time. Recognition also matters: a supervisor who catches a costly error should receive credit in the same way as a colleague who wins a contract.

What to do: draft role profiles for the four roles with outcomes, required competence and training, and time allocation. Pilot them in two functions where agents are already live. Adjust grading and objectives to include supervision, and create a short training course covering sampling, escalation and stopping rules. Review after six months with the owners and supervisors themselves.

If an agent is a worker, it needs a performance review. The review has three parts. Performance covers accuracy, completion rate, rework and speed. Cost covers the compute, licence, supervision and exception-handling time per completed task, compared with the human or previous process. Risk covers incidents, near misses, breaches of limits and complaints. Each has an owner and a threshold at which action is required.

The benefit side deserves honesty. PwC's adopters report increased productivity (66%), cost savings (57%), faster decision-making (55%) and improved customer experience (54%). These are perceived benefits. Few organisations yet measure per-task cost including supervision, so reported savings may overstate net gains. A supervision-inclusive cost per task is the figure a CFO will trust, and it makes the case for investing in the supervision roles in the first place.

Include human factors. KPMG's Q2 2026 AI pulse found only 2% of leaders reporting employee resistance to AI agents, with 30% reporting a mixed response. Low resistance does not mean high understanding. Employees who work alongside agents need to know what the agent does, how to challenge it and whom to tell when something looks wrong. Include a simple measure of that in engagement surveys.

Set stopping rules in advance. A threshold for error rate, for complaint count or for a data breach should automatically pause the agent, with the owner deciding how and when it restarts. Rules agreed in calm conditions are followed in stressful ones; rules invented during an incident are argued over. Test them: switch the agent off in a drill and measure how long the process takes to recover.

A final point on culture. Reporting agent errors must be safe. If supervisors fear blame for what an agent got wrong, they will under-report, and the organisation will lose the learning it needs. Treat incidents as system failures first and individual failures only when negligence is clear. Share anonymised incident summaries across owners each month, and credit teams that surface problems early. Candour is the control that makes the other controls work.

What to do: build a one-page dashboard for each agent and review all of them in a monthly forum chaired by the executive sponsor. Track cost per task with supervision included, errors caught and missed, escalations and incidents. Report trends to the executive team each quarter. Use the same dashboard to decide which agents to scale, redesign or retire.

A sequence of ninety days is enough to start. In the first thirty, build the register, identify agents without owners and set the pause rule. In the next thirty, draft charters and role profiles for the highest-risk agents, train the first supervisors and set up the monthly forum. In the final thirty, run a stop drill, review the first metrics and decide which roles need formal grading changes. The aim is a working system, not a perfect one.

Choose where to begin by consequence and autonomy. Agents that take customer-facing or employee-facing decisions with little human check deserve first attention. Agents that draft material for human review can follow. The matrix in this report sets out how autonomy and consequence combine, and what level of supervision each combination warrants. Use it to rank the register rather than to debate each agent separately.

Involve people early. Employees will ask whether agents are replacing them, and avoiding the question damages trust. Be honest about what is known and unknown, describe which tasks agents will take and describe the new roles that supervision creates. Invite the people closest to the work to help write the charters, since they will be the first to see where an agent goes wrong.

Keep the programme proportionate. A small firm with a handful of agents needs a register, a one-page charter for each, named owners and a monthly check; it does not need a committee. A large enterprise with hundreds of agents needs the same principles with tiering, automation of the register and an assurance function. The test is the same in both: if an agent fails tomorrow, can we name the person who is accountable?

What to do: appoint an executive sponsor, usually the COO or CHRO jointly with the CIO. Publish the ninety-day plan, hold the first forum within two weeks and report progress at the next leadership meeting. Karter McKinley supports AI and technology implementation, including registers, charters and role design, working alongside your legal, risk and HR teams.

Karter McKinley supports AI and technology implementation, including agent registers, charters and supervision models. We help HR, operations and technology teams agree who owns what, then document it in a form people use. Legal and regulatory judgements remain with your own advisers.