At A2 supervised the agent decides and drafts; a named person approves each action before anything changes in your systems.
| Line | What the agent proposes | Value | Basis | Confidence | Decision |
|---|---|---|---|---|---|
| 4500067123 · 00020 | Cancel 200 of 7741-X at plant 1210 | $8,420 | excess 11.4 wk · window Thu 24 Sep · standard | 91% | Approved 09:12 · DEC-0417 |
| 4500068004 · 00020 | Cancel 1,200 of 8802-K at plant 1210 | $24,100 | excess 9.3 wk · window Wed 23 Sep · standard | 93% | Approved 08:58 |
| 4500067990 · 00030 | Cancel 300 of 5513-B at plant 1220 | $7,160 | excess 8.6 wk · window Thu 24 Sep · standard | 88% | Edited to 200 · approved 09:03 |
| 4500068110 · 00010 | Cancel 150 of 9021-M at plant 1220 | $5,900 | excess 12.1 wk · window Fri 25 Sep · standard | 90% | Approved 09:05 |
| 4500068221 · 00050 | Cancel 80 of 4410-T at plant 1210 | $2,280 | excess 10.2 wk · window Thu 24 Sep · standard | 90% | Approved 09:10 |
| 24 more lines | each under $25,000 | $139,440 | three conditions matched on each | All approved by 09:12 |
Six steps: arrive, rank, read, decide, act, record.
A draft reaches the queue with everything a decision needs: the action in one sentence, the conditions that matched, the values they read, a confidence and the money and cover at stake.
The queue ranks the drafts, so the first row is the one that earns the most attention. Filters narrow it to one plant, one supplier or one material, and a group of related lines can be read together.
The approver decides. An approval sends the action to your system through the interface that system supports, and the answer it returns is stored beside the decision.
The record then closes with the name, the role, the time, the decision, any reason given and the answer from your system. That record is what an auditor reads, and what the learning loop reads for the next cycle.
Each of the three decisions closes the record, and each one teaches the next draft.
Edits and reasons are the most valuable thing a queue produces. In the showcase estate the buyer made 9 edits across 212 drafts, and each one told Juno something about how this team works.
One reason appeared on five of the nine edits: the supplier is on allocation. Juno proposed an exclusion for that supplier class, and j.park approved it as rule version 4.
Working through a block keeps the detail. A group of nine lines for one supplier is read once, and the nine records still carry their own values, their own confidence and their own answer from your system.
Every decision, whichever of the three it is, ends in the same record: who, when, what, why, and what your system replied.
The Monday as it ran, the escalation route the agent carries, and the clock a queue can add.
The showcase estate’s cancellation queue uses the window itself as its clock, and its escalation route has three legs: a line at or above $25,000 goes to D. Okafor as buyer, for his judgement; a failed write-back goes to the owner, L. Tan; a limit breach pauses the run and tells her.
A queue can also carry a service-level clock, set by the team that owns the decision: hours for a queue that works against a closing window, days for a safety-stock queue. A reminder goes out before the clock runs out, on the channel the approver already uses, and cover for holidays is a setting: a named colleague receives the queue while the owner is away.
When a clock runs out, the escalation route sends the item to the next desk named in advance, so the item always has somewhere to go, and both desks appear on the record.
An item that reaches the end of its window stays open. The agent raises the same finding on its next run, so the case returns while it can still be acted on.
Six limits decide whether an action goes out, waits for one person, or waits for two.
| Limit | What it does | Example from the showcase estate |
|---|---|---|
| The level | Decides whether an action needs an approval at all | A2 asks every time; A3 asks for exceptions only |
| Value threshold | Sends an action above a set value to a named approver | $25,000 on a cancellation line |
| Daily action budget | Caps how many actions one agent may take in a day | 60 actions a day for the cancellation agent |
| Confidence floor | Holds back an action whose evidence agrees less well | 0.80; below it the finding goes to a person |
| Rate limit | Spreads actions over time so a source stays healthy | 20 writes a minute to the purchasing interface |
| Two-person approval | Asks a second named person for sensitive fields | Bank details and payment terms on a supplier record |
Limits are checked before an action leaves, and every check is stored with the decision. A limit that sends an action to a person is the same mechanism as the level: it decides who signs for this one.
Two-person approval applies to the fields where a single mistake is expensive. The record then holds two names, two times and two sessions.
Market practice runs the same way. A KPMG pulse survey reports 61% of organisations requiring approval processes for agent actions (KPMG, Q2 2026).
The AWS agency scoping matrix makes a related point: as autonomy rises, oversight shifts from approving each action to auditing what was done (AWS, 21 November 2025). Juno holds both, because the level is yours to set per agent.
Showcase estate.
D. Okafor opened the queue at 08:50. Twenty-eight of the 29 drafts were approved as drafted and one, line 00030 of order 4500067990, was edited from 300 to 200 pieces and approved at 09:03. The last decision was made at 09:12: 22 minutes for the whole queue. The 9 lines at or above $25,000 reached him as buyer, for his judgement, with the reason for each.
From 09:12 the 29 changes went to SAP S/4HANA, and by 09:21 every acknowledgement was stored. After six weeks and 212 drafts, 203 approved as written and 9 edited, J. Ridley moved lines under $25,000 to A3 at the level review on Friday 30 October, and D. Okafor sees exceptions only on those.
How an approved action reaches SAP or Oracle through the interface that system supports, and how its answer is kept.
A named colleague receives the queue as cover, and every item still carries its clock. When a clock runs out the escalation route sends the item to the next desk.
Yes. Drafts that share a supplier, a material or a plant are read as a group and approved in one action, and each line keeps its own record and its own answer from your system.
The values the agent drafted, the values the approver sent, the reason and the person. The learning loop reads those pairs, so the next drafts arrive closer to your practice.
It replaces the work of assembling evidence. In the showcase estate the buyer decided 29 drafts in 22 minutes, because the reason and the facts sat on the row.
Each one builds on the one before it.
Reads your rows in place.
Connect and move ↗One record per real thing.
Master and govern ↗Apps for every decision.
Operate and decide ↗Asks and acts, at your level.
Ask and act ↗Bring two systems and one decision your planners make every week. In the first week we register the data, confirm the mapping with your team, answer your questions in ChatLima and run the first agent at the level you choose.