AI Agents on the Org Chart: Who Owns Them, Who Pays, Who Answers?
IN ONE MINUTE
2026 is the year AI agents move from pilot into the operating model. Barclays has announced about 5,000 job cuts, citing efficiency from AI, and Accenture now charges heavy AI use to the projects that benefit. Yet half the companies using agentic AI have not updated their governance for it. The question is no longer whether agents can do the work, but who owns them. My answer: one owner and three limits, on authority, budget and review. Budget each agent as a process cost per successful outcome, not as a salary.
THE OBJECTIVE
Most large organisations have now run agentic AI pilots. Some delivered strong returns; others are on hold. Boards want to know which agents to scale, what they will cost, and who is accountable. This issue gives a practical model for governing a workforce of people and AI agents: who approves, who checks, who pays, who answers, and what each member of the leadership team should do before the 2027 budget is set.
WHAT THE RESEARCH SHOWS
Governance lags adoption. EY's September 2026 survey of 202 senior executives at US companies with revenue of US$1 billion or more found that 91% use agentic AI, yet 49% of them have not updated their governance for it. 85% have agents acting without real-time human involvement, and 26% cannot detect unauthorised agents. Deloitte finds only one company in five has a mature governance model for autonomous agents.
Value lags use. In McKinsey's 2026 research, 80% of employees say AI has improved their own productivity, but only 37% of organisations report a positive impact on EBIT. Gartner predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027, because of rising costs, unclear value or weak risk controls.
Costs surprise at scale. McKinsey finds that moving an AI tool from pilot to production nearly quadruples its cost. One bank's projected AI spend rose from US$2–3 million a year to more than US$25 million once every employee could use every tool and model without guardrails.
The market is acting. Barclays expects to reduce its workforce by about 5,000 full-time roles, around 5%, about 60% of them in support functions. It expects more than £1.1 billion of gross efficiency savings in 2027, including fewer developers "as AI makes our people more productive"; its chief executive called AI "the icing on the cake, not the cake". Accenture now centrally funds US$29 per employee a month for Claude, and charges heavier use to the approved project or team.
The message: agents create value only when someone owns the outcome, the authority and the bill.
THE CASE: WHERE AGENTS PAY, AND WHERE THEY STALL
From 26 days to under 3
I recently led a new operating practice for digital governance approvals using AI agents. The average approval cycle fell from 26 days to under 3.
Speed was the smaller gain. The bigger ones were transparency and consistency. Every request now follows the same structured, logical sequence, against the same criteria, every time. The team can see where each request stands and why. Two hidden costs disappeared: the delay while a request waited behind other people's priorities, and the detail lost each time it was handed from one person to the next.
The agents' work doesn't stop at approval. They prepare and check each request. After the decision, they update the records to reflect what the meeting agreed and take on the skills the next step needs. Then they monitor the operation continuously, within the access they have been authorised and the guardrails set for them. They report regularly to stakeholders on KPIs, progress, outcomes and recommendations. People still make the decisions; the agents make sure the decisions are carried through and visible.
What I hear from peers
In conversations this year with senior executives across several industries, I hear the same split. Some agent pilots have delivered strong returns; others are on hold. The public evidence points to the same three reasons.
- Trust. In 2024 Klarna said its AI assistant was doing the work of 700 customer service agents. In 2025 its chief executive told Bloomberg that the AI-first approach had produced lower-quality service, and the company began hiring people again. When customers or employees don't trust the agent, the savings don't last.
- Integration. MIT's NANDA initiative found that about 95% of the generative AI pilots it studied showed little or no measurable impact on profit and loss. The cause was not model quality but a "learning gap": tools that sit beside the workflow and are never built into it.
- Economics. A poorly designed agent can burn up to 30 times the tokens of a well-designed one, McKinsey finds. Several pilots I've heard about were paused for a simple reason: the agent cost more to run than the work cost to do the existing way.
The agents that pay share three traits. They do high-volume, rules-based work, often in the back office, where MIT found the largest returns. Their workflow was redesigned before the agent was built: McKinsey describes two clients who took the same use case into production, and only the one that reimagined the workflow first saw large gains. And a named person answers for the result.
THE FRAMEWORK: ONE OWNER, THREE LIMITS
An agent can do the work. It cannot be accountable for it. So every agent needs one named human owner, and three limits that owner sets and answers for.
| Element | The question | What it looks like | How to measure it |
|---|---|---|---|
| One owner | Who answers for this agent's results? | A named business leader, not a team or a vendor, who owns the outcome the agent serves | Every agent in production has a named owner |
| Authority limit | What may the agent decide alone? | A level on the autonomy ladder below, like a financial approval limit | Decisions made within the limit; escalations; overrides |
| Budget limit | What may it spend, and is it worth it? | A cost per successful outcome compared with today's cost, plus a monthly ceiling | Cost per successful outcome; spend against ceiling |
| Review limit | When is it checked, and by whom? | A quarterly review of quality, cost and drift, and a probation period before it moves up a level | Error rate; human review time; drift; time to fix |
The autonomy ladder
Every large company already runs a delegation-of-authority matrix for people. Agents need the same. An agent moves up one level at a time, and only on evidence.
| Level | The agent may | Evidence needed to move up |
|---|---|---|
| 1. Suggest | Recommend; a person does the work | Recommendations accepted most of the time |
| 2. Draft | Prepare the work; a person reviews every item | Few corrections over a full cycle |
| 3. Act with approval | Complete the work once a person approves | Approvals rarely change the outcome |
| 4. Act, then report | Act, and report every action for review | No material errors over the probation period |
| 5. Act alone, within limits | Act within agreed limits; exceptions escalate | Continuous monitoring and quarterly review |
Most agents should stop at level 3 or 4. Level 5 is for high-volume, low-risk, reversible work.
Treat agents as a new kind of workforce. Like employees, they are hired against a job description, given only the access they need, put on probation before they scale, supervised, reviewed, and retired when they no longer pay their way. HR, Finance, Technology and Risk each own part of that lifecycle; the business owner owns the whole.
THE ORGANISATION AND THE BILL
Once agents are governed, they change how the organisation is drawn and how it is budgeted. McKinsey's October Quarterly article, "AI is changing work. Now it has to change the organization", calls the new map an "orchestration chart": which people and agents combine around an outcome, and where human judgement stays.
Could an agent cost more than your CEO?
Traditionally, organisation design starts from the work. Each role is shaped by the nature of its tasks and the size of its workload, and its pay reflects the value it creates.
That is now changing. Companies that have proven their first agent use cases are beginning to design a human–agent organisation. HR places agents in the organisation chart, each with a reporting line to a human owner. Every task is then reviewed against two questions: can an agent do it better, and how much of it is there? Roles are rebuilt around the work that remains distinctly human, and a person's pay reflects the value they add on top of the agents: judgement, relationships and accountability.
An agent's "pay" works differently. Its cost is its token consumption, and that rises with every task it handles. Put agents and people on the same budget sheet, and a high-volume agent could cost more than a senior executive. Gartner already reports individual users consuming tens of thousands of dollars of tokens in a month, and expects AI coding costs alone to match a typical engineer's monthly salary by 2028.
Arithmetically, then, yes. But it is the wrong comparison. An executive's pay reflects accountability and judgement, not task volume. An agent should be judged against the cost of the work it replaces or makes possible, per successful outcome. And its cost is variable: it rises with volume, model choice and poor design. Budget an agent like a salary and you hide that risk. Budget it as a process cost, owned by the person who owns the outcome, and you can manage it.
THE PART THAT NEEDS CARE
Agents change people's work before they change the P&L, and people notice first. When an agent takes over part of a role, the capacity it frees has to go somewhere: more output, new work, a sustainable workload, or lower cost. McKinsey calls this the productivity dividend, and argues it should be allocated deliberately rather than left to chance. Leaders who say nothing about it will find the workforce assumes the worst.
Two further risks deserve attention. Agents absorb entry-level work, which is where professionals used to learn, so organisations need new ways to build judgement, as I explored in Issue 007. And governance skipped for speed returns as an incident: 47% of executives in EY's survey say their organisation has bypassed its own AI governance for an urgent deployment.
WHAT THIS MEANS FOR CEOS AND BOARDS
Ask for the register. Request a list of every agent in production, its owner, its autonomy level and its monthly cost. If it can't be produced, governance hasn't caught up with adoption.
Name one accountable leader, and judge outcomes. Scaling agents across functions needs a senior executive with authority over funding, talent and trade-offs. Ask which agents moved a P&L line, and which should be stopped. Usage is not value.
WHAT THIS MEANS FOR CFOS AND FINANCE TEAMS
Move agent costs into the processes they serve. Give each a cost per successful outcome, a monthly ceiling and an owner, and report spend monthly by agent and by model.
Fund the base, charge the excess. A modest central allowance keeps everyone learning; heavy use is charged to the project that benefits. The people who create the value become accountable for the cost.
WHAT THIS MEANS FOR HR LEADERS
Write job descriptions for people and agents together. For every redesigned role, set out what people own, what agents do, and where judgement and approval sit. Measure people on the quality of the decisions they own, not the volume of tasks.
Plan the productivity dividend openly. Decide, and explain, where freed capacity goes. Trust is the condition for adoption.
WHAT THIS MEANS FOR TECHNOLOGY AND OPERATIONS LEADERS
Redesign the workflow before building the agent. Map the process, remove unnecessary hand-offs, then decide where an agent adds value. Not every step needs one.
Fit the model to the task, and log everything. Route routine work to smaller, cheaper models and keep frontier models for complex reasoning. Log every action, escalation and cost, so owners can see what the agent did and why.
Questions leaders ask
Who should be accountable when an AI agent gets it wrong?
A named business leader, never the agent or the vendor. An agent can do the work but cannot be accountable for it. In my framework, every agent has one owner who sets its authority, budget and review limits. Name that owner before the agent goes live.
Should we cap our employees' AI spending?
Cap the central allowance, not the use. Accenture funds a modest monthly amount centrally and charges heavier use to the project that benefits. That keeps everyone learning and puts the cost where the value is judged. Track cost per successful outcome, not tokens alone.
How should we budget for AI agents next year?
As process costs, not salaries. An agent's cost varies with volume, model and design, so put it inside the process it serves, owned by that process's leader. Set a cost per successful outcome against today's cost, add a monthly ceiling, and review quarterly.
THE LAST WORD
An agent can do the work. Only a person can own the result.
Which of your AI agents has a named owner today, and could that person tell you what it cost last month?
—
Until next week,
Ian Cheng
→ Found this useful? Forward it to a CFO, COO or HR leader.
→ Subscribe to The AI Executive Briefing, weekly in your LinkedIn inbox.
Sources: EY, "EY survey finds that autonomous AI implementation outpaces oversight, yielding an AI governance gap", 15 September 2026 (202 senior executives at US companies with revenue of US$1 billion or more, May–June 2026). Deloitte AI Institute, "The State of AI in the Enterprise", 2026 (3,235 senior leaders, 24 countries, August–September 2025). Alexis Krivkovich, Brooke Weddle, Holly Price and Vik Sohoni, "AI is changing work. Now it has to change the organization", McKinsey Quarterly, 5 October 2026. Lari Hämäläinen, Wasim Lala and Kaavini Takkar, "What does AI really cost?", McKinsey, 14 October 2026. Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027", 25 June 2025. Gartner, prediction on AI coding costs, reported by InfoWorld, 24 June 2026. MIT NANDA, "The GenAI Divide: State of AI in Business 2025", August 2025 (about 150 leader interviews, 350 employees surveyed, 300 public deployments analysed). Klarna chief executive interview with Bloomberg, reported May 2025. Barclays, Q3 2026 results call, 22 October 2026. Accenture AI spending policy, reported by Forbes, 16 October 2026. Approval-cycle results: the author's own programme data.