Agentic AI Chapter 3 of 9

How much an agent may do is your decision, agent by agent.

Every agent in Lima Juno carries one setting: its autonomy level, from A0 to A4.

ONE SETTING PER AGENTfive designed roles from the blueprint · the dial is yours, per agentA0Human-ledCredit hold /release decisionOrder-to-Cash · financeA0A1A2A3A4PERSONJ. Harrison decides andsigns the holdA1AssistedShortage &allocationnegotiationProcure-to-Pay · purchasingA0A1A2A3A4PERSONL. Tan decides who gets theparts and actsA2SupervisedPO change &cancellationProcure-to-Pay · purchasingA0A1A2A3A4PERSOND. Okafor approves eachdraft by nameA3DelegatedStatistical baselineforecastingPlan-to-Produce · planningA0A1A2A3A4PERSONL. Tan handles theexceptions onlyA4AutonomousCash application &matchingOrder-to-Cash · financeA0A1A2A3A4PERSONJ. Harrison audits a sampleeach weekblueprint · designed, yet to be built · the A2 role is the one in production
Five designed roles from the blueprint, one at each level, each with its own dial. The dial sits on the agent, so a company can hold four roles at four different levels on the same day. Blueprint: designed, yet to be built; the A2 role runs in production.
5
levels, from human-led to autonomous, set on each agent separately
product
866 of 960
designed roles keep a person in or on the loop, at A1, A2 or A3
blueprint · designed, yet to be built
203 of 212
drafts approved as written in six weeks: the record that moved one agent towards A3
showcase estate
1 step
lowers a level, at once, by the owner or a manager
product
So far
  1. What it isYou have seen the eight parts of an agent, and that two of them, limits and escalation, depend on how much it may do alone.
  2. Why nowYou have seen that unclear limits are one of the four reasons earlier AI stalled.
The five levels

Each level states what the agent does and what the person does

Five sentences say how one piece of work is shared between an agent and a person.

LevelWhat the agent doesWhat the person doesExample
A0 Human-led
The work stays with people; the blueprint lists the step so the map stays complete.The human performs the work.Credit hold / release decision · Order-to-Cash
A1 Assisted
The agent informs.The human decides and acts.Shortage & allocation negotiation · Procure-to-Pay
A2 Supervised
The agent decides and drafts.A named human approves each action.PO change & cancellation · Procure-to-Pay · designed at A3, started at A2
A3 Delegated
The agent decides and executes within versioned rules.The human handles exceptions only.Statistical baseline forecasting · Plan-to-Produce
A4 Autonomous
The agent executes and self-corrects.The human audits after the fact.Cash application & matching · Order-to-Cash

The two middle columns together are the definition. Each example is a designed role from the blueprint, told as a scene below. blueprint

The last chapter left two of an agent’s eight parts open: limits, how far an action may go, and escalation, where a case goes when it passes them. Both depend on how much the agent may do alone. The ladder answers that in five levels, set on each agent by name, with a date.

The advantage, in one sentence

The ladder converts 'how much AI' from an open-ended technology question into a per-agent governance setting the client controls, agent by agent.

Unclear limits were one of the four reasons earlier AI stalled; a level with an owner and a date is that limit made explicit, one agent at a time.

The whole gamut

One worked scene at each level, on the estate you already know

Five blueprint roles across finance, purchasing and planning, so the range shows on work you recognise.

WHAT THE AGENT DOESWHAT THE PERSON DOESHANDOFFA0Human-ledCredit hold / releasedecisionThe work stays with people. ChatLima answers J.Harrison’s questions in seconds: the overduebalance, the open orders, the history.J. Harrison decides the hold, signs it and records itin the ERP under her own name.a question · thefactsA1AssistedShortage & allocationnegotiationAs designed, ranks the week’s 14 shortages byrevenue at risk and shows the sales orders eachone hits.L. Tan decides who gets the parts and makes theallocation change in the planning suite herself.a ranked listA2SupervisedPO change & cancellationDrafts 29 cancellations under policy version 3,each with its basis and a confidence; asks D.Okafor as buyer on 9 more.D. Okafor approves each line by name; the write-backfollows and the ERP confirms it.a draft · thego-aheadA3DelegatedStatistical baselineforecastingAs designed, refreshes the baseline for 4,737item-locations inside the versioned error band andwrites it to the planning suite.L. Tan handles the 63 item-locations that breachedthe band, and only those.63 exceptionsA4AutonomousCash application &matchingAs designed, matches 12 of 14 invoices and posts;re-tries the 2 open lines on a second pass;re-routes what stays open.J. Harrison audits a sample each week: 20 of 340postings, each read from its record.the record,afterwardsshowcase estate · five designed roles from the blueprint · the A2 role is the one in production
Five levels, two lanes. The agent’s lane grows from an answer to a draft to an execution; the person’s lane moves from deciding to auditing. The label on each arrow is what passes between them. Showcase estate.
JH
Tuesday 10:15 · finance · the credit review

Credit hold / release decision · Order-to-Cash · A0 Human-led J. Harrison, finance lead, has Helvetia Retail AG on her list. She asks ChatLima for the overdue balance, the open orders a hold would stop and the payment history. The answer arrives in seconds from three systems, with the query on show: $312,000 overdue at 31 days, a promise broken last Tuesday, three open orders worth $448,000, 47 of the last 48 months paid on time.

She decides that the Thursday shipment holds until the customer’s payment run clears, and records it in SAP S/4HANA under her name. A released hold is a credit decision with a signature, so the step sits at A0; the platform’s part is the facts.

LT
Monday 08:30 · Penang · the shortage list · as designed

Shortage & allocation negotiation · Procure-to-Pay · A1 Assisted As designed, the shortage agent reads the 1,240 exception messages L. Tan opened at 07:40, finds the 14 shortages that bite this week and ranks them by revenue at risk, $1.1M in all, showing the sales orders each one hits. First is material 7741-X at plant 1210: 600 units short against Helvetia Retail AG’s order, with Acme Components GmbH confirming 400 for Wednesday.

L. Tan decides who gets the parts: Helvetia Retail AG takes the 400, Northbay Logistics’ smaller line moves a week, and D. Okafor calls the supplier about the rest. She makes the allocation change in the planning suite herself. Her screen is the ranked list, one row per shortage; the agent’s work ends there.

DO
Monday 21 September 2026 · 06:00 run · 09:12 approvals

PO change & cancellation · Procure-to-Pay · A2 Supervised At 06:00 the cancellation agent ran on policy version 3, the sentence L. Tan wrote on the Friday: excess above eight weeks of demand, inside the cancellation window, under $25,000, supplier class standard; at $25,000 and above, ask the buyer. 38 lines matched, worth $412,000.

29 drafts worth $187,300 reached D. Okafor’s approval queue, each with its basis and a confidence; the 9 lines of $25,000 and above reached him flagged for his judgement, with the reason. By 09:12 he had cleared the queue: 22 minutes, approve, edit, or reject with a reason, line by line. The ERP confirmed each write-back in 9 minutes. The blueprint designs this role at A3; the showcase estate started it at A2.

LT
Thursday 09:00 · Penang · cycle day · as designed

Statistical baseline forecasting · Plan-to-Produce · A3 Delegated As designed, the forecasting agent scores 4,800 item-locations overnight against version 2 of its rules. For 4,737 the refreshed statistical line sits inside the error band versioned with the rule, so the agent writes the new baseline to the planning suite and keeps each job status. 63 breach the band and go to L. Tan’s exception queue.

For material 7741-X at plant 1210 the last three overrides raised the line by 18% and actuals came in below it each time, so the agent proposes to withdraw the October override. L. Tan works the 63: 58 accepted, 3 edited, 2 rejected with a reason. Her screen is the exception queue, one row per breach: the band, the miss, the proposed correction.

JH
Tuesday 06:30 · the cash queue · overnight run · as designed

Cash application & matching · Order-to-Cash · A4 Autonomous As designed, the cash application agent takes the overnight remittance: Northbay Logistics has paid $184,300 against 14 invoices and the remittance arrives as a PDF. On the first pass the agent matches 12 invoices, applies a 2% discount inside terms and posts the receipt in SAP S/4HANA; the document number goes on the record.

On the second pass it re-reads the PDF against the customer’s payables portal: the $6,200 line carries a freight claim number, so it becomes a case on the deductions desk; the $410 line sits inside the write-off tolerance and goes to J. Harrison for the write-off decision. On Friday she audits a sample, 20 of the week’s 340 postings, from their records. ‘Self-corrects’ means the second pass and the re-route; the audit stays hers.

Showcase estate throughout. The A2 scene ran on the Monday and is the running example of chapter 1. The A1, A3 and A4 scenes are told as designed, on the same estate and people, for roles yet to be built; the A0 scene is a human decision helped by ChatLima.

At A0 the platform helps by answering, and the decision stays human on purpose. At A4 ‘self-corrects’ means a second pass and a re-route, and the person still audits. Between them the handoff changes shape at every step; the record underneath stays the same.

Designed, yet to be built

All five are roles on the blueprint; one of them runs

The map proposes a starting level for every role; the setting on a built agent is yours.

Lima’s agent blueprint is a design of 960 agent roles across five end-to-end chains. It is the map of what Juno can be configured to do, and the first agents are built from it one at a time, the way the PO cancellation agent was. The five scenes above are roles on that map.

One of the five runs: the PO cancellation agent is in production at a high-tech manufacturer, where planners author the policy in plain language, the rules engine executes it on federated ERP data, write-back waits for human approval, 100% of actions trace to a rule and run time costs zero tokens. The other four are designed, yet to be built. production

The blueprint’s designed starting levels

A0 Human-led 37A1 Assisted 68A2 Supervised 311A3 Delegated 487A4 Autonomous 57

The blueprint’s proposed start for its 960 designed roles: 866, or 90%, keep a person in or on the loop at A1 to A3. Your settings replace these agent by agent. blueprint

Moving up

A level rises when the recorded cases, the owner and the escalation route agree

Evidence first, then a name, then the route for the cases the agent hands back.

1
Recorded cases pass the testSix weeks of lines run through the agent again and its drafts sit beside what D. Okafor did: 212 drafted, 203 approved as written, 9 edited.
96%
2
The owner approves by nameL. Tan approves by name as the owner; J. Ridley, who sets the level, signs the change. The test result is attached to both signatures.
signed
3
The escalation route is liveAbove the value threshold, a line reaches D. Okafor as buyer, with the reason; a failed write-back reaches the owner, L. Tan; a limit breach, the daily budget or the confidence floor, pauses the run and tells the owner.
ready
4
The level changes, on the audit trailLines under $25,000 move to A3 from the next run; lines of $25,000 and above stay at A2. Who asked, who approved, the evidence and the date are kept together.
A3 in part
DRAFTS APPROVED AS WRITTEN, BY WEEK90%92.5%95%97.5%100%95% line for a move to A3 · four consecutive weeksWk 1 · 21 SepWk 2 · 28 SepWk 3 · 5 OctWk 4 · 12 OctWk 5 · 19 OctWk 6 · 26 Octapproved as writtenlevel reviewshowcase estate · drafted 61 · 42 · 34 · 29 · 24 · 22 = 212 · 203 as written · 9 edited
The share of drafted cancellations D. Okafor approved as written, week by week. The green band is the 95% line J. Ridley set for a move to A3, to be held for four consecutive weeks; weeks 3 to 6 held it. Drafted by week: 61, 42, 34, 29, 24 and 22. Showcase estate.

Raising a level takes three things together: a passing test on recorded cases, a named owner’s approval and a live escalation route. The test runs the agent again over cases that already happened and sets its drafts beside what the person did.

After six weeks the record read 212 drafted, 203 as written, 9 edited: 96%. The line J. Ridley had set for a move to A3 was 95% approved as written, held for four consecutive weeks, and weeks 3 to 6 held it. Lines under $25,000 moved to A3, where D. Okafor handles the exceptions only; lines of $25,000 and above stay at A2.

Moving down

A level comes down in one step, and it applies from the next run

The quick way down is what makes the way up safe to take.

Lowering is deliberately lighter. The owner or a manager sets the lower level in one step, it applies from the next run, and the reason is recorded with the change. The agent keeps working at the lower level while the cause is looked at.

What prompts itWho actsWhat happens next
A run of edits or rejectionsThe owner, L. TanDown one step; each action waits for an approver again.
A source change or a new policyThe data lead with the ownerDown one step while the rule is re-simulated and released as a new version.
A period close or an auditThe managerThe agent informs for the period; the level returns when it ends.
At every level

Limits, the audit trail and confirmed write-back hold throughout

The level moves the decision; the controls around the decision stay in place.

LimitsValue thresholds, a daily action budget, a confidence floor, rate limits and two-person approval on sensitive fields.
Access rightsEach agent reads its facts under the same role-based access a person has, with its own least-privilege permissions.
Versioned rulesEvery action names the rule version that produced it. Restoring an earlier version creates a new version.
The audit trailSignal, reasoning, recommendation, limits checked, approval, action, acknowledgement and outcome, for every run.
Confirmed write-backA write reaches the ERP through a supported interface, and the ERP’s acknowledgement is kept with the record.
An escalation routeAnything outside the limits reaches a named desk, with a clock that passes it to the next one.

These six hold at A1 as firmly as at A4. The level changes who acts on a finding; the controls around the finding stay in place, so an agent at A4 leaves the same record a buyer would have approved at A2.

Published ladders

The five levels line up with the ladders analysts and buyers already use

One row per level, so an analyst can place Juno on the ladder they already use.

Juno levelKnight Institute user roleAWS agency scopeDriving automation analogy
A0 · Human-ledThe role list starts above this rungThe scope list starts above this rungThe driver does all of it.
A1 · AssistedOperator and collaboratorScope 1 · the person keeps the agencyDriver assistance: the car warns, the driver steers.
A2 · SupervisedConsultant, moving to approverScope 2 · prescribed agencyPartial automation: hands ready on the wheel.
A3 · DelegatedApproverScope 3 · supervised agencyConditional automation: the car drives and asks the driver to take over.
A4 · AutonomousObserverScope 4 · full agencyHigh automation: the car drives, and the trip is reviewed afterwards.

Crosswalk checked 19 September 2026; published level numbers are written out in words.

Four things are Lima’s own: a rung for the work that stays with people, so coverage is reported honestly; a named approver at A2; versioned rules at A3; and a setting that sits on the agent, in your hands. The rest is deliberately familiar.

Lima’s earlier data-quality pages describe autonomy in four short steps: the first lines up with A1, the second and third with A2, the fourth with A3. A4 adds self-correction, with the audit after the fact.

Terms on this page
Autonomy levelThe setting on one agent that states what it does and what the person does.
Level ownerThe named person accountable for an agent’s level and for each change to it.
Recorded-case testA re-run of the agent on past cases, set beside what people did at the time.
Designed starting levelThe level the blueprint proposes for a role before it is built; your setting replaces it.
Questions
Where should a first agent start?

Most first agents start at A2, where the agent drafts and a named person approves each action. Those approvals build the record that supports a later move to A3.

Can two agents run at different levels on the same day?

Yes. The level sits on the agent. One company can run credit decisions at A0, cancellations at A2 and cash matching at A4 on the same day, each with its own owner.

What does A0 add, if the work stays with people?

A place on the map for work people keep on purpose, so coverage is reported honestly. The platform still helps there: ChatLima answers the questions the decision needs, in seconds.

The question this leaves

At every level a person is in the picture. What exactly does that person see, decide and set?

Answered next

People and agents

The four ways people and agents work together, the screens each person sees, and the nine settings that shape every agent.

In the platform

Where Juno sits among the four capabilities

Each one builds on the one before it.

One shared foundationCatalogLineageAccess policyWrite path

See Juno answer and act on your own data

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.