Routing every held action to the most senior person feels safe. It is not. One approver for everything becomes a bottleneck, then a rubber stamp, then a single point of failure with no check, and seniority is not competence on the agent's work. Route approvals by certification and depth, and keep the senior person for decisions only they should make.
Why does the instinct to escalate everything feel right?
When an organisation first deploys agents, uncertainty is high. Nobody quite knows how reliable the agent is or how the new review process will work. The natural response is to send everything the agent holds to the person with the most authority and experience. If anything goes wrong, at least the right person looked at it.
There is a similar instinct on the agent side: give one agent every task because it is simpler to supervise. The piece on whether one agent should do everything explains why that concentrates risk. The same logic applies to people.
What actually goes wrong?
Four things, usually in this order.
The queue backs up. A senior person has other responsibilities. Held items wait hours or days. Customers chase, and the business starts asking whether the agent's limits could be loosened so less is held. That moves risk from the queue to the customer.
Review becomes cursory. Faced with a long queue and little time, the senior person starts approving quickly. They trust the agent's draft more than they should because they have no time to do otherwise. Rubber-stamping by the most senior approver is worse than by a junior, because nobody checks them.
Independence disappears. If the same person sets the limits, owns the agent and approves its held actions, nobody stands outside the loop. A mistake in their judgement passes through every layer unchallenged.
Competence is assumed, not shown. Senior people often have the least hands-on practice with the workspace. They may be excellent at the underlying business and still misread a verdict, confuse a refusal with a fault, or miss that a key fact is declared rather than sealed.
How should approvals be split instead?
By what the decision needs, not by who is most senior. Each held item has a depth and a type. Match it to people who are certified for that depth and that type, and keep a separate check for the highest-risk actions.
| Held item | Who approves | Why |
|---|---|---|
| Depth 2 draft for a person to send | Any certified reviewer | Routine judgement; volume work |
| Depth 3 routine execution held over a limit | Certified reviewer with authority for that limit | Needs specific authority, not seniority |
| Depth 4 money movement or hard-to-undo action | Two certified people, one of them senior | Second check; independence |
| Change to the agent's limits or brief | Proposed by one person, approved by another | Separation of duties |
The guide on how to split review work between junior and senior staff covers how to build this into a rota. For money, a dedicated process is worth setting out on its own: how to set up two-person approval for money moved by agents, with a two-person approval procedure template to adapt.
Where should the senior person spend their time?
On decisions that need their authority and on checking the system. That means acting as second approver for the riskiest actions; approving changes to limits and briefs proposed by others; sampling junior reviewers' decisions to see whether review is real; and reading the workspace for patterns. Internal audit thinks in exactly these terms: the skills that matter most for internal auditors are about testing whether controls operate, not performing them all personally.
Senior people should still be certified on the work they approve. Seniority earns them the right to the hardest decisions; certification shows they can make them on this system.
What is the trade-off?
Spreading approvals means junior people decide things that seniors used to. Some of those decisions will be wrong. Seniors may feel they are losing control, and some will resist handing over work they believe only they can do properly.
It also takes more organising. A single approver needs no rota. A split model needs certification, clear authority per depth and type, cover, and sampling.
The response is that the single-approver model hides its failures rather than avoiding them. A senior rubber-stamping a long queue makes more mistakes than a well-certified team with a sensible split, and nobody notices until something goes badly wrong. A split model surfaces errors early, through sampling and second checks, when they are cheap to fix.
What should you do differently?
Look at who approves held items today. If one name dominates, change it. Certify a wider group of reviewers and give them authority matched to depth and type. Reserve the senior person for depth 4 second approvals, changes to the agent, and sampling. Separate whoever proposes a change from whoever approves it. And measure how long held items wait, because a growing queue is the first sign that approval is concentrated in too few hands. The guide to AI agent skills by role sets out what each role needs to know.
Questions people ask
Is it ever right for one person to approve everything?
In a very small business with low volumes, it may be unavoidable for a while. Even then, someone else should periodically sample their approvals, and limits on the agent should be set conservatively.
How do you stop juniors from escalating everything back to seniors?
Give them clear authority and the training to use it, and track escalations. Some escalation is healthy; a pattern of escalating everything means people do not feel equipped or authorised.
Should senior people be certified at a higher level?
Not necessarily higher, but certified for what they approve. If they also set limits or change agents' briefs, a deeper qualification that covers design is appropriate.