The six Overview numbers are best taught as a ten-minute morning routine: decisions taken, refusals, waiting on you, depth in use, certified, and ledger height. For each number, the team learns what normal looks like, what a change means and the action it prompts, then practises on real mornings.
Why train people to read numbers at all?
The workspace records every verdict and every human decision. The Overview condenses that into six figures. A team that reads them each morning sees a growing queue, a reviewer in trouble or a certificate about to lapse before a customer does. A team that glances at them sees only that some numbers went up.
Reading is a skill because the same movement can mean different things. Refusals rising may mean the agent is struggling, or that the business lowered a limit yesterday and the agent is doing exactly what it was told. Telling those apart needs the bound that decided, which is why training matters more than the dashboard.
The team lead owns the routine. Every reviewer learns it, because the lead is sometimes absent and the numbers do not wait.
What do the six numbers show, and what should each prompt?
- Decisions taken. Every verdict the decision chain reached in the period, whatever it was. It is the volume of work that went through the chain. A jump with no matching rise in customer contact suggests repeated work or a loop. Prompt: ask what sent the extra work through.
- Refusals. Actions the chain stopped, each one naming which of the five bounds decided: role scope, certificate, sources, the business's limits or the platform guardrails. A shift in the dominant bound matters more than the total. Prompt: ask what changed in that bound.
- Waiting on you. Held actions waiting for a named person, with the age of the oldest. A rising count with a steady oldest age means volume; a rising oldest age means something is stuck. Prompt: check who the oldest item is waiting for.
- Depth in use. How far agents are actually acting, capability by capability. A capability sitting below the depth its certificate allows may mean the business has tightened a limit. A drop overnight usually means a certificate. Prompt: find which capability moved and why.
- Certified. Certificates currently in force. Agent certificates lapse after 90 days, and a run that clears no depth drops the agent to depth 1 the same day. Prompt: confirm recertification is scheduled and plan for extra held work if it slips.
- Ledger height. Entries in the append-only record. It should grow in step with decisions taken. If decisions rise and the ledger does not, something is not being filed, and that is a question for the platform owner the same morning. Prompt: compare the two before anything else.
The steps
- Person (team lead): records a baseline: two weeks of the six numbers, to show what normal looks like for this team.
- Person (team): learns the six numbers and their prompts in a single short session using the team's own baseline.
- Agent, depth 1: presents the Overview each morning, with yesterday's changes highlighted.
- Person (rotating reviewer): reads the six numbers aloud at the morning huddle and states one action.
- Person (team lead): confirms or corrects the reading, briefly.
- Agent, depth 1: opens the underlying records for any number the team wants to examine, such as the oldest held item or a reviewer's recent approvals.
- Person (team lead): runs a scenario session monthly, with doctored Overview figures that tell a story the team must work out.
- Person (team lead): logs each day's reading and action in the team's record.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Decisions taken | Shows volume through the chain | 1 | Lead asks what drove a jump |
| Refusals | Groups by deciding bound | 1 | Lead asks the owner of that bound |
| Waiting on you | Shows count and oldest age | 1 | Reviewer chases or reallocates |
| Depth in use | Shows depth by capability | 1 | Lead finds which capability moved |
| Certified | Lists certificates in force | 1 | Lead confirms recertification |
| Ledger height | Counts record entries | 1 | Lead checks it matches decisions |
What does a scenario session look like?
Show the team a hypothetical Overview. Waiting on you has doubled overnight and the oldest item is only two hours old. Refusals are flat. Depth in use for refunds has fallen from 4 to 1, and Certified is one lower than yesterday. Decisions taken and ledger height have moved together.
The right reading: the refunds certificate lapsed or failed, the agent dropped to depth 1 for refunds, and refunds it used to issue are now arriving as held items. The action: move a reviewer onto refunds this morning and confirm when recertification will run. Teams who practised this recognise it in seconds. Teams who did not spend the morning asking whether the agent is broken.
Which controls matter, and what evidence do they leave?
- Daily reading with a stated action. Evidence: the log of readings and actions.
- Rotating reader. Evidence: names in the log, showing more than one person can do it.
- Monthly scenarios. Evidence: the scenario, the team's reading and the debrief.
- Escalation to bound owners. Evidence: messages to the owners of limits or sources when a bound shifts.
What should you measure?
Mornings with a reading and an action logged. Readers who can interpret a scenario correctly. Time from a rise in the oldest waiting item to action. Certificate lapses that arrived without warning, which should be none. And any day the ledger did not keep pace with decisions taken.
Where does it usually go wrong?
Only the lead reads the numbers. When the lead is away, the routine stops.
Teams react to totals rather than the deciding bound, and complain about the agent when the business changed its limits.
Withheld actions are logged as defects and chased with the builder. They are platform decisions; the useful question is why people keep asking for them.
Reading the Overview leads naturally into reading individual records; see training staff to read an agent's decision record. The certificate number connects to raising an agent's depth safely. Advice and lettings teams apply the same routine to their own queues; see training the advice team and training the lettings team. The guide to training teams for AI agents shows where this fits.
Questions people ask
How long should the morning reading take?
Ten minutes at most. If it runs longer, take the deep question offline and keep the huddle moving.
Should every reviewer learn this, or only leads?
Every reviewer should be able to read the six numbers. Leads also need to decide what to do about them, which takes more practice.
What if the numbers look normal but something feels wrong?
Open the records. The numbers summarise; the record shows each decision. A reviewer's instinct is worth a five-minute check.