Agents add capacity only as far as depth and review allow. A COO needs to understand how far each agent may act, how much human review its held work creates, what happens to throughput when a certificate lapses, and how teams reorganise around exceptions. Get those right and agents change unit costs; get them wrong and they add a queue.
How does depth translate into capacity?
Depth is how far an agent may act, per capability. Each level shifts work differently:
| Depth | What the agent does | Effect on the operation |
|---|---|---|
| 1, read | Gathers and summarises | Saves preparation time; people still do everything else |
| 2, draft | Prepares; a person sends | Saves most effort, but every item still passes a reviewer |
| 3, execute within bounds | Completes routine work inside limits | Removes routine items from human queues |
| 4, money or hard-to-undo actions | Acts within tight limits, often held | Large savings where certified; review stays on exceptions |
| 5, coordinate other agents | Orchestrators only | Changes how work is routed |
Most operations start at depth 2 and earn depth 3 on routine work. Depth is earned, not configured: the role sets a ceiling, certification on the business’s own cases sets the actual depth, and the lower wins.
So the capacity case for an agent is not “it does the job”. It is “at depth 3 on these task types, with this share held, it removes this much work”. That number comes from evidence, not projection.
What does review effort cost?
More than most business cases assume. Every held item needs a named person with authority to decide it. If limits are set cautiously, a large share of work is held and reviewers become the bottleneck. If reviewers rush, the control weakens.
Three numbers tell a COO whether review is working: held-item age, reviewer overturn rate, and review time per item. Rising age means too few reviewers or too tight limits. A very low overturn rate may mean reviewers have stopped looking. A high one means the agent’s sources or brief need work.
Reviewers also need the power to say no and make it stick. Reviewers need authority, not just access sets out why.
What happens to capacity when certification lapses?
It drops. Certificates lapse after 90 days, and a run that clears no depth drops the agent to depth 1 the same day. An agent that cleared a thousand routine items a day at depth 3 now only reads. That work returns to people immediately.
A COO should plan for this the way they plan for system outages: know which agents re-certify when, keep cover available, and treat a failed run as an incident with a recovery plan.
How does the operating model change?
Three shifts are common.
Teams reorganise around exceptions. The routine flows through agents; people handle what is held, refused or escalated. Front-line roles become review and judgement roles, and job descriptions should say so.
New roles appear. An AI operations manager runs agents day to day. An agent owner is accountable for each agent’s role and sources. A governance lead owns the limits and evidence.
Onboarding and verification gain weight. Agents rely on sealed records. If customer or supplier data from KYC or KYB checks is wrong, agents act on it quickly. Data quality becomes an operational priority.
Who should the COO bring in early?
Compliance, to agree red lines and testing; see what a compliance team needs to know. Internal audit, to agree how the record will be sampled; see what internal audit needs to know. And team leads, who turn the plan into rotas. The full picture by role is in AI agent skills by role.
What should a COO ask before scaling an agent?
- Which capabilities is it certified for, and at what depth?
- What share of its actions were held last month, and how long did they wait?
- How often did reviewers change its drafts?
- When does its certificate lapse, and what is the cover plan?
- Are the reviewers trained, and do they hold current certificates?
Where the Academy fits
Pryme Intelligence Academy’s free readiness assessment gives a COO a quick read on a team of up to 25: the team takes the Foundations test set in its industry, and the report shows scores by concept, the gaps and the first agent to start with. Reviewers then take the Operator track; seats come with paid Pryme Intelligence plans.
Questions people ask
Should we reduce headcount when agents go live?
Measure first. Review effort, exceptions and certification cover often absorb more time than expected in the early months.
Which process should we start with?
A high-volume process with clear rules, sealed sources and modest consequences, so the agent can earn depth 3 on evidence.
Who owns the agent once it is live?
A named agent owner in the business, with operations running it day to day and governance owning the limits.