For most of the history of operations, leverage meant two things: more people and better systems. If you wanted to handle more orders, more customers or more cases, you hired, or you bought software that let the people you had handle more. The chief operating officer's job was to organise both so that capacity grew a little faster than cost.
AI agents add a third kind of capacity: work carried out continuously by software, inside limits a person has set. That changes what the operations leader is responsible for, and what leverage means.
What changes when agents join the team
Three things change at once.
- Capacity is designed, not only hired. A well-defined piece of work can often be given to an agent sooner than a new hire could be recruited and trained for it. But the agent only does what it has been designed, permitted and tested to do.
- The unit of management moves from task to process. A team lead who once shared tasks among half a dozen people may now own a process in which some steps are done by people and others by agents. The accountability does not change. The work of managing does: writing clear missions, setting boundaries, reviewing evidence and handling exceptions.
- Mistakes scale as well as successes. A person who misreads a policy makes a handful of errors before someone notices. An agent with the same misunderstanding can repeat it hundreds of times in an afternoon. Speed without design is not leverage; it is exposure.
The mandate, in five parts
If agents are part of the operation, the COO's mandate grows to include five responsibilities that used to sit, if anywhere, with IT or an innovation team.
- Own the operating model, not just the operation. Someone has to decide which missions exist, who owns each one, which agents and people carry it out, and how the results are recorded. That is an operating decision, not a technical one.
- Decide what stays human. Decisions about money, commitments to customers, personal data, people's roles and significant risk are the usual candidates to keep with a person. Write the list down before agents arrive, not after the first incident.
- Govern autonomy as a portfolio. Each workflow sits at a level of autonomy — the Citadel's trust model calls them observe, assist, operate and optimise — and moves between levels only on evidence and by a person's decision. The COO should be able to see that portfolio at a glance.
- Measure leverage honestly. "Hours saved" estimates are easy to produce and hard to trust. Measure what an outcome costs to complete, how often people have to intervene, and whether quality and customer outcomes held. Those numbers support decisions; estimates support slide decks.
- Build the team that runs it. The people who will thrive are those who can own a process end to end, read evidence, handle the exceptions agents escalate, and improve the system. Hiring and development should reflect that.
A new definition of leverage
The old question was: how much output can each person produce? The better question now is: how much well-governed work can each accountable person direct?
| Leverage, before agents | Leverage, with agents | |
|---|---|---|
| Source | Headcount and software licences | People, agents and the operating layer around them |
| Growth | Hire, train, retain | Design, test, govern, expand on evidence |
| Constraint | Hiring speed and cost | Clarity of missions and quality of oversight |
| Key measure | Output per person | Outcomes per accountable person, at acceptable quality and risk |
| Main risk | Burnout and turnover | Unbounded actions and unclear ownership |
The constraint moves from hiring to clarity. An organisation that cannot describe its own work — the decisions, the owners, the exceptions — will find agents hard to use well, however good the models become.
Where it connects to growth
Operations and growth are often managed as separate agendas. Agents bring them closer together. The same capacity that clears an internal backlog can shorten the time it takes to respond to a new enquiry, prepare a proposal or follow up on a renewal. Customers notice response time and consistency long before they notice the technology behind them.
That only works if the commercial side trusts what the operation is doing on its behalf. A message sent to a customer, a quote prepared, a commitment implied: each needs a clear boundary and, where the stakes are real, a person's approval. In the Guild's own managed operations the division of labour is stated simply: "Sentinel watches. Oracle interprets. Humans decide."
What the operating layer gives the COO
None of this works if the leader cannot see it. The minimum is one view that answers a handful of questions: what outcomes are we pursuing, which agents are active and within policy, where human judgement is needed before work continues, whether systems are healthy, what each mission costs, and what changed, who acted and on what evidence. In the Guild those questions map to the Citadel's modules — Missions, Agent Console, Approvals, System Health, Financials and the Activity Ledger. The tooling matters less than the principle: if you cannot see who did what and why, you are not managing the operation; you are hoping.
What the board will ask
Sooner or later, the board or the owners will ask the COO three questions about agents. It is worth being able to answer them before they are asked.
- What agents do we run, and what does each one do? A register with a mission, an owner and a status for every agent, not a list of tools someone bought.
- What can they do without a person? The actions each agent may take on its own, the limits it works within, and the decisions that always reach a named person.
- How do we know they work? The evidence: outcomes against a baseline, exceptions, interventions, incidents and what was changed as a result.
If the answers depend on asking around, the operating layer is not there yet.
The first 90 days
- List the recurring decisions and workflows that most affect results, with an owner for each.
- Write down which decisions stay with people, and why.
- Pick one workflow where an agent could take on defined work, and measure it today.
- Decide its starting level of autonomy and the evidence needed to move up.
- Agree how cost per outcome and human intervention will be measured.
- Set a monthly review of the evidence, with the authority to widen, narrow or stop.
The COO's mandate has always been to make the organisation's capacity match its ambition. Agents do not change that. They change where the capacity comes from, and they make clear ownership and good governance part of the job rather than a supporting function.
