An AI agent moves to its next depth on evidence, and much of that evidence comes from the people around it. Trained reviewers make correct decisions on held work, feed good cases back into certification and spot where the agent is ready for more. Untrained teams produce noisy evidence and stall the agent at depth 2. The fastest route to more autonomy runs through people, not settings.
What does depth mean, and how is it earned?
Depth is how far an agent may act. Depth 1 is read. Depth 2 is draft, where a person sends. Depth 3 is executing routine work within bounds. Depth 4 is moving money or taking actions that are hard to undo. Depth 5 is coordinating other agents, for orchestrators only. The full definition is in depth: what it means for AI agents.
Depth is earned, not configured. The agent's role sets a ceiling; evidence from certification sets the depth; the lower wins. Certification replays the agent against real cases from the customer's own operation and scores every answer against a floor, across gates that must each clear on their own. Certificates lapse after 90 days; a lapse, or a run that clears no depth, drops the agent to depth 1 the same day. So depth is never permanent: it is held by continuing evidence.
Where do trained people come into that evidence?
In three places. First, the cases. Certification uses the organisation's own cases, and the best ones come from reviewers: borderline items, refusals that should hold, drafts that were wrong in instructive ways. A team that tags and keeps these from the queue gives certification a far better test set than one that does not.
Second, the decisions. While an agent runs at depth 2, every draft is decided by a person. If those decisions are careful and consistent, the record becomes clean evidence of what the right outcome was, case by case. If reviewers approve in seconds and override refusals under pressure, the record is muddy, and nobody can say whether the agent would have been right on its own.
Third, the judgement about readiness. Reviewers who understand the agent see where its drafts are reliably right and where they are not. They can say, with reasons, that routine address changes are ready for depth 3 while bank detail changes are not. That judgement, backed by the record, is what a governance forum needs to approve a change.
Why do untrained teams keep agents stuck?
Because their evidence cannot be trusted. Every verdict and every human decision is filed in an append-only record, and when someone asks whether the agent's drafts were right, the answer comes from comparing them with what people decided. If people decided badly, the comparison means nothing.
Untrained teams also tend to ask for depth for the wrong reasons. The review load is heavy, so they ask for the agent to "be turned up". That request has no evidence behind it, and a careful governance forum will refuse it. The team stays at depth 2, frustrated, while a trained team in a similar operation has moved routine work to depth 3 because it could show the agent was ready.
What does a trained team do differently?
It treats the agent as a candidate for the next depth and builds the case deliberately:
- Keeps instructive cases from the queue for certification, including refusals that must hold.
- Decides held items carefully and consistently, with calibration, so the record is reliable evidence.
- Tracks where drafts are right by capability, sampled and checked, not guessed.
- Proposes depth changes with evidence, one agent at a time, starting with the most routine.
- Plans for the downside, knowing that a lapse, or a run that clears no depth, drops the agent to depth 1 the same day.
This is how the decision chain should work. How an AI agent decides what to do next explains the chain from the agent's side; how to train a team to raise an agent's depth safely covers the people's side in detail.
What is the trade-off?
Patience. Training a team well and building evidence carefully takes longer than switching a setting, which is what some vendors offer. Leaders who want the agent doing more next month will find this slow.
But configured autonomy without evidence is borrowed. It looks faster until the first serious error, after which depth is cut back and trust takes months to rebuild. Earned depth rises more slowly and stays risen. Matrices that map capabilities to depth by function, such as the bank operations depth matrix and the finance agent depth matrix, help teams plan which capabilities to build evidence for first.
What should you do differently?
Stop asking how to give the agent more authority, and start asking what evidence would earn it. Train reviewers so their decisions are reliable evidence. Collect instructive cases from the queue. Propose depth changes one agent at a time, with the record behind them. And keep training as the agent rises: at depth 3, reviewers move from deciding every item to sampling, which is a different skill. The wider approach is in training teams for AI agents.
Where the Academy fits
Pryme Intelligence Academy's Operator track trains the reviewers whose decisions become evidence, including the module on teaching the agent. The Builder track trains those who design, train and certify agents, and Certified Pryme Builder requires a capstone agent certified in Attest to depth 2 or higher.
Questions people ask
Can a well-trained team raise depth without certification?
No. Depth comes from certification evidence, within the role's ceiling. A trained team improves the evidence and the case; it does not replace certification.
What changes for reviewers when depth rises?
They review less and sample more. At depth 3, routine actions execute within bounds, and reviewers check samples and handle exceptions. Sampling well needs its own training.
Does depth ever go down?
Yes. Certificates lapse after 90 days; a lapse, or a run that clears no depth, drops the agent to depth 1 the same day. Trained teams expect this and know how to cover the work until depth is restored.