The Agent as Chief of Staff: Coordination Can Be Delegated, Accountability Cannot

The problem

In February 2018, 1,259 executives in 91 countries told McKinsey how they spent their working time. On average they put 37% of it into making decisions, and 61% said most of that decision-making time was used ineffectively¹ (self-reported; survey published April 2019). Only 48% agreed that their organisations made decisions quickly, and just 37% said their decisions were both high in quality and fast¹ (self-reported). The load rises with seniority: 14% of C-suite respondents said decisions took more than 70% of their time¹ (self-reported). These figures are more than three years old; the research found no more recent survey of comparable scale.

Much of that time is not the decision itself but the work around it. Senior executives report the most exposure to what the survey calls cross-cutting decisions: a chain of smaller choices made by different groups, such as a pricing change¹. The organisations that handled these well concentrated on coordinating stakeholders and running decision meetings properly¹. Structure matters too: 61% of respondents in organisations with one to three reporting layers said decisions were made quickly, against 38% in organisations with seven or more¹ (self-reported). Put plainly, a leader's time is consumed by convening, sequencing, chasing and recording, and the more hands a decision passes through, the slower it goes.

McKinsey turned its averages into a thought experiment: for managers at an average Fortune 500 company, ineffective decision making could mean more than 530,000 lost working days and roughly $250 million of wasted labour costs a year¹ (projected by the source from survey averages and stated assumptions, not measured).

The same pattern appears in government. In a UK cross-government trial of an AI assistant built into office software, respondents at Grade 6 and Senior Civil Service level said it saved them most time on collaboration, whereas most other grades found it most effective for drafting documents² (self-reported survey, trial run 30 September to 31 December 2024).

The gap in the evidence. No recent, independent measurement was found of how much of a senior leader's week goes on coordination specifically. The survey measures decision-making time, which includes coordination but is not the same thing. The evidence used here comes from a global executive survey and from the UK, the EU and Canada; nothing comparable was found for Asia or Africa.

What already works

Summarising and drafting. Two UK government trials in the last quarter of 2024 are the most detailed public evidence on AI assistants in office work. The Government Digital Service gave licences to 20,000 staff across 12 organisations; users estimated an average saving of 26 minutes a day² (self-reported survey, 7,115 responses, published June 2025). The report itself warns that these savings are self-reported². In the Department for Business and Trade's own pilot of 1,000 licences, transcribing or summarising meetings was the most common use, and drafting written reports showed the largest saving, at 1.3 hours per task³ (self-reported diaries adjusted for unused outputs and new work, no control group, October to December 2024).

In plain terms: tools that read and write save people some time, and summarising meetings, the raw material of coordination, is what people use them for most.

Arranging things: the weak spot. In the Department for Business and Trade pilot, scheduling took longer with the assistant than without it: after adjustment, each scheduling task cost an average of 0.6 hours more, and 36% of recorded scheduling uses were rated as dissatisfying³ (self-reported diaries, November 2024; only 14 scheduling cases, so the figure is fragile). The cross-government survey, by contrast, reported a saving of 9 minutes on communication tasks such as scheduling meetings² (self-reported). The Department for Business and Trade found no evidence that time savings had led to improved productivity, and colleagues outside the pilot had not noticed any improvement³.

What this means: the evidence on the arranging side of coordination is thin and contradictory, and the one evaluation that adjusted for wasted outputs found it slower.

Agents that act on their own. An agent goes further than an assistant: it takes a sequence of actions, such as reading a thread, booking a room and sending a follow-up, without a person approving each step. In a benchmark that simulates a small software company, with internal websites, data and simulated colleagues to message, the best agent tested completed 30% of tasks autonomously⁴ (measured in a simulated benchmark without a control group; September 2025 version; abstract). A separate research group measured the length of task, in skilled human working time, that frontier models could complete half the time: about 50 minutes for models tested in early 2025, with that length doubling roughly every seven months since 2019⁵ (measured on software and research tasks; abstract). The authors caution that it is unclear how far the results carry over to real work⁵.

What these figures mean: agents can now finish some short, self-contained office tasks unaided, but most attempts at longer, multi-step work still fail, and coordination is long, multi-step work.

What does not exist. No deployment was found, in either the public or the private sector, in which an agent carries a senior leader's coordination end to end, across teams and over weeks, with a published evaluation. The building blocks exist separately; the system would have to be assembled, and its hardest parts invented.

The possibility

Take the chief operating officer of a mid-sized organisation, public or private. One way this could work is that the officer gives an agent standing authority over the coordination layer of the job. The agent would keep a live register of pending decisions, each with a named owner and a deadline. It would assemble the papers for each decision, work out which people a decision touches, propose and book the meetings, draft and send follow-ups, chase overdue actions, and record what was decided and why. When a decision stalled beyond an agreed window, it would escalate to the officer. The officer would keep direction: setting priorities, making the calls and answering for them.

Which parts depend on AI. Generative AI and agents would do the reading and judging: extracting decisions and actions from meeting transcripts and email threads, drafting papers and messages, inferring who needs to be involved, deciding when and how to chase, and planning sequences of steps. Much of the rest is ordinary software that exists today and needs no AI: calendars, task tracking, access permissions, audit logs and a decision register. The remainder is organisational: written decision rights, delegation limits and the officer's willingness to let go. A good share of any gain would come from the register and the decision rights, which an organisation could adopt without any AI at all.

Where accountability would sit. The evidence points one way. Under UK company law a company must have at least one director who is a natural person⁶, and a tribunal has already rejected an organisation's attempt to treat its chatbot as answerable for itself⁷. If an agent acted in the officer's name, the officer and the organisation would answer for what it did. The evidence supports an agent that coordinates on behalf of a leader; it does not support an agent that is itself accountable.

Coordination versus accountability. Coordination is work an agent could, in principle, carry out in someone's name. Accountability is a position a person holds. On current law, the first could be delegated and the second could not.

What would have to be true

For the agent to give a leader time back, rather than create new work, each of these conditions would need to hold.

1. Arranging tasks become reliable. Scheduling is where the trial evidence is weakest³. An agent that invited the wrong people or misread priorities would spend the leader's time instead of saving it.
2. Agents can sustain long, multi-party sequences. This is the AI capability the whole idea rests on. A decision may take weeks and pass through many hands, while current results show agents completing under a third of simulated office tasks⁴ and succeeding half the time on tasks that take a skilled person about 50 minutes⁵.
3. Decision rights are written down, and are not treated as the whole answer. An agent can only route a decision if someone has defined who decides what. Yet the survey found that clear roles, accountability or decision rights on their own were not enough to set the best decision makers apart; process and well-run meetings mattered more¹.
4. Every authority the agent uses belongs to a named person. Each action would need to trace back to someone who could answer for it⁶⁷, with logs, and limits above which the agent could only propose, not act.
5. Oversight meets the rules where they apply. If the agent allocated work based on people's behaviour or traits, or monitored and evaluated staff, the EU AI Act would class it as a high-risk system⁸, and the requirements for such systems include human oversight⁸.
6. Staff accept being coordinated by an agent. In the Department for Business and Trade pilot, how much people used the tool was shaped by the attitudes of colleagues and line managers³. Being chased by a system on a senior leader's behalf would be a different experience from receiving a summary.
7. Saved time goes to direction. Neither UK trial could show what the time saved was used for²³. Unless the leader's diary actually changed, faster coordination would only move work around.

What could stop it

What the evidence shows

Organisations answer for their systems. In February 2024 a British Columbia tribunal heard Air Canada argue that it could not be held liable for information given by its website chatbot. The tribunal called this a remarkable submission, found that the chatbot was simply part of the airline's website, held that the airline had not taken reasonable care to make it accurate, and ordered it to pay C$812.02⁷. The case concerned information rather than actions, but it shows an organisation cannot place an automated system between itself and responsibility.

The law expects a human at the top. A UK company must have at least one director who is a natural person⁶.

Employment uses carry extra obligations in the EU. The EU AI Act classes AI systems used to allocate tasks on the basis of individual behaviour or traits, or to monitor and evaluate people at work, as high-risk⁸. In July 2026 the EU adopted an amending regulation to simplify the Act's implementation, which entered into force on 27 July 2026⁹.

Errors are real and checking is uneven. In the Department for Business and Trade pilot, 22% of diary respondents said they had encountered false or misleading information presented as fact, and the evaluation found that checking of outputs was inconsistent across people and tasks³ (self-reported, 300 diaries, November 2024). Scheduling had the largest share of cases, 36%, in which the output was not checked or checking was not comparable³ (self-reported).

Structure drives speed. Respondents in organisations with fewer reporting layers reported faster and better decisions¹ (self-reported, 2018).

Reasoning

The challenge question is what would make this possibility less likely, less useful or more harmful than it first appears. Several answers stand out.

A simpler fix may do more. If slow decisions come mainly from too many layers and unclear processes, an agent would speed up the paperwork of a slow structure without changing its outcome. Removing a layer, writing down decision rights, keeping a decision register or appointing a capable chief of staff are older remedies, and the survey evidence points more towards structure than towards clerical load.

Accountability could blur inside even if it cannot blur outside. The tribunal case suggests customers and courts would hold the organisation responsible. Internally, however, "the agent scheduled it" could become a way of avoiding ownership, and decision records written by an agent could make it harder, not easier, to see who actually decided.

The risk falls on people who did not choose it. The leader would carry responsibility for messages they never read; staff would receive instructions and reminders from a system; and a single error could reach dozens of people at once, because coordination is by nature broadcast.

Coordination is also how leaders learn. Chasing a stalled decision often reveals why it stalled: a quiet objection, a resourcing gap, a political problem. If an agent absorbed that work, the leader could lose the informal signals that make direction possible.

Incentives favour adoption over outcomes. Both UK trials measured satisfaction and self-reported time²³; neither measured whether decisions became faster or better. Without outcome measures, an agent could look successful while decisions stayed as slow as before.

Someone has to pay. The Department for Business and Trade said further evaluation was needed to assess value for money³. An agent with standing authority would also need integration, logging, oversight and training, costs that sit on top of licences.

Signs to watch

• Government or company evaluations of agents that use a control group and measure decision speed or quality, rather than self-reported time saved.
• A published value-for-money assessment of the Department for Business and Trade pilot, which its evaluation said was being conducted³.
• The date from which the EU AI Act's high-risk obligations for employment systems apply, following the July 2026 amendment⁸⁹.
• New workplace-agent benchmark results on tasks that require coordinating several simulated colleagues⁴.
• Court or tribunal rulings on liability for actions taken by an AI agent, as distinct from information given by a chatbot⁷.
• The bringing into force of the 2015 amendment listed as omitting section 155 of the Companies Act 2006, which would change the rules on company directors⁶.