Each decision an agent makes is stored with the rule version that produced it, the facts at that moment, the recommendation, what the person did with it and what happened afterwards.
Five steps run every cycle, and two of them belong to a person.
Every run of an agent leaves a record. The record holds the rule version in force, the facts read at that moment, the recommendation with its basis and its confidence, the limits that were checked and the action that followed. It is written once and kept.
What the person did sits in the same record. A buyer who approves 40 drafts in a morning leaves 40 accepted decisions. A buyer who changes a quantity leaves an edit, and the edited value is kept beside the drafted one. A rejection carries the reason the buyer chose.
Juno reads these records as a group. It looks for edits and rejections that repeat: the same supplier, the same plant, the same material class, the same week of the month. A pattern that holds across enough decisions becomes a candidate change, with the cases that support it.
A candidate is a proposal, so a person decides. Your team opens it, reads the supporting decisions, simulates the change on past data and publishes it as a new rule version. The rules engine then applies that version, and the cycle starts again.
Nine fields, written once per decision and kept for the retention period you set.
| Field | What it holds | From the running example |
|---|---|---|
| Decision reference | One identifier per decision, used in every report and export. | DEC-2026-09-21-0417 |
| Trigger | The schedule, interval, source event or request that started the run. | schedule · Monday 06:00, after the planning run |
| Rule version | The exact version of the policy that produced the recommendation. | PO cancellation · version 3 |
| Facts | The values read at that moment, with the system each came from. | read 06:01 · projected excess 11.4 weeks · window closes Thu 24 Sep |
| Recommendation | The proposed action, the basis for it and the confidence. | cancel line 00020 of order 4500067123 · 0.91 |
| Limits checked | Each limit that applied and the result of the check. | value $8,420 under $25,000 · daily budget 29 of 60 |
| Human decision | Accepted, edited or rejected, with the person and the reason. | approved as drafted by D. Okafor · 09:12, after 3 h 06 min in the queue |
| Action and acknowledgement | What was written to the source system and what it returned. | cancellation posted · acknowledgement 20260921-114 |
| Outcome | What happened by the check date, against the stated objective. | excess cover 11.4 to 6.2 weeks at day 14 |
Showcase estate. The audit trail page shows the same record as an auditor reads it.
Deloitte, 24 April 2026, names three controls that most organisations still lack: clear decision boundaries, real-time monitoring, and audit trails of the full action chain. Limits carry the first, the observability page carries the second, and this record carries the third.
Juno proposes; a named owner approves; the change arrives as a new version of the rule.
Rule discovery reads past planner behaviour, and past agent decisions, for conditions a person would recognise. It shows each candidate with the decisions behind it, the share of cases it covers and what the change would have done to last quarter.
The division of work stays the same at every step. The language model finds the pattern and explains it in plain words. The rules engine executes the published rule, the same way on every run. A named owner decides which proposals become versions.
Every approval queue is also a training set, held by your own people.
Learning pays before any agent runs on its own. At A2 a named person approves each action, so every morning brings a fresh batch of drafts and a fresh batch of evidence. The share approved as written is the plainest measure of whether the drafts are improving.
Your team sets the line that a move up the ladder requires, and the manager report shows the measure against that line each week. Lowering a level takes one step and is immediate, so the reader of this chart keeps the choice in both directions.
MIT NANDA, July 2025, found that 95% of pilots show a profit effect too small to measure, and named a learning gap as the cause: tools that retain feedback do better than tools that start each cycle fresh.
Capgemini, 16 July 2025, reports that 73% of executives say the benefits of human oversight outweigh its cost. The loop on this page is what turns that oversight into a return: the approvals a person gives today shape the drafts a person reads tomorrow.
Showcase estate.
Over six weeks from Monday 21 September the agent drafted 212 cancellations; D. Okafor approved 203 as written and edited 9, with 0 rejected. Five of the nine edits cut the quantity on lines for one supplier, on allocation and kept by the buyers on purpose. Juno grouped the five edits, found 41 earlier lines with the same pattern and proposed a condition that keeps those lines.
A. Novak simulated the proposal on last quarter: of 214 cancellations, 17 would have been held back, and each of the 17 matched a line buyers had kept by hand. The approver login j.park released it as version 4. From the next run, the agent stopped drafting those lines, and D. Okafor’s queue held only lines that needed a decision.
What the agent was asked to achieve, how the result is measured, and what happens when live data moves away from what a rule assumed.
The difference is where a correction goes after it is made.
In most teams a planner’s correction lives in an e-mail, a comment in a spreadsheet or a conversation at a desk. The next planning cycle starts from the same rule as the last one, and the same correction is made again. Knowledge stays with the person who holds it.
Some tools retrain a model on recent data, which moves the behaviour in a way that is hard to explain to an auditor. Juno keeps the change visible: a proposal, its supporting cases, a simulation, a named approver and a version number that later records point back to.
Learning happens inside Juno. The decision records stay in your estate, and what they teach arrives as a rule version your team approves and releases.
A proposal reaches production only as a published rule version, released by a named owner after a simulation. Between releases, the rules engine applies the version in force and gives the same result on the same facts.
Juno shows the number of supporting cases beside each candidate and the share of decisions it covers. Your team sets the floor; in the showcase estate the first proposal rested on 46 cases, five edits and 41 earlier lines.
Yes. At A2 each approval, edit and rejection is recorded, so the drafts improve week by week. In the showcase estate the share approved as written stood above the 95% line for the last four of the six weeks, and 203 of 212 drafts were approved as written.
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.
