Rank overdue accounts by cash impact and payment likelihood, send the right reminder or make the right call at the right time, and follow every promise to pay to its outcome.
J. Harrison, finance lead, opens the collections board. Overnight Juno ranked 212 overdue accounts by cash impact and likelihood to pay, sent 140 first reminders at A3 under the collection policy, and logged three promises that came back by e-mail, each with its date and the next action.
Nine accounts above $50,000 wait in the approval queue, each with a drafted escalation. Helvetia Retail AG is second on the list: $312,000 at 31 days, a promise broken last Tuesday, and a note that the customer’s payment run falls on Thursdays, so the draft asks for confirmation before Wednesday noon. She edits one sentence and approves eight drafts; the ninth she holds for the account manager.
Then she asks ChatLima which of the nine paid late in each of the last three quarters. The answer arrives in 4 seconds with the query on show: four of them, and Helvetia Retail AG in all three. The reminders were the agent’s, the escalations were hers, and the pattern is what she takes into Tuesday’s credit review, the A0 scene of chapter 3. Showcase estate; the roles are designed, yet to be built.
Collections is the largest working-capital lever in the chain, and the gap is measured. Hackett assessed 19 collections, dispute and deduction vendors and reports that about half of collection activity is touchless today, that the average company runs 8.5 days more delinquent and 1.1 points more bad debt than leaders, worth up to $44 million a year in write-offs (The Hackett Group, 30 March 2026) 3. At companies that redesign order-to-cash around AI, collectors spend 83% more time with customers and average delinquent days fall by 85% (The Hackett Group, 15 July 2026) 1.
It is also the step with the most named agents. SAP lists five agents for collections, for collection insights, e-mail outreach, account preparation, dunning insights and receivable account analysis (SAP, n.d.) 11, and an accounts receivable assistant generally available in the second quarter of 2026 (SAP, May 2026) 12. Oracle’s collectors workspace, one of 22 agentic applications, is built to collect cash faster, lower days sales outstanding and raise promise-to-pay conversion (Oracle, 24 March 2026) 14. BlackLine runs concurrent collection calls around the clock with a voice agent that records payment commitments (BlackLine, 14 April 2026) 27 29.
The named results are vendor-published. HighRadius names GE Healthcare with 66% fewer past dues, Nidec with $79 million of working capital released and Supply Technologies with 56% less past due (HighRadius, 10 September 2026) 24. And the demand is there: 46% of finance leaders put working capital optimisation among their priority areas for AI agents, and 63% have fully deployed AI (Deloitte, 4 February 2026) 7.
A collector pulls the aging report from the ERP each morning, sorts it by amount and days overdue, reads the account notes and the last promise, and then works down the list by e-mail and telephone. Dunning letters go out from a batch run on fixed days, whatever the customer’s history.
Promises to pay are typed into a spreadsheet or the collections worklist, and broken promises surface only when the collector returns to the account. Disputes found on the call are forwarded to sales by e-mail. Hackett finds that half of collection activity is still manual 3 and that collectors at leading companies spend 83% more of their day with customers 1.
The cost is the order of the list. A collector who works by amount and age spends the morning on accounts that would have paid anyway, while the account about to break a promise sits lower down, and the dispute mentioned on Tuesday’s call reaches sales on Friday.
The same morning, by hand and with the ranked worklist. The two figures are Hackett’s; the steps are a design.
The same six stages as chapter 6, with the receivables desk’s objects. Everything in the rows below is a design: the collections roles sit on the blueprint at A3, and the tense is “would”.
| Rank | Account | Overdue | Days | Likelihood to pay | Last contact | Proposed | Decision |
|---|---|---|---|---|---|---|---|
| 01 | Northbay Logistics | $418,000 | 18 | 84% | promise kept 14 Sep | first reminder, sent 02:10 | sent by rule · A3 |
| 02 | Helvetia Retail AG | $312,000 | 31 | 41% | promise broken Tue 15 Sep | escalation: confirmation before Wednesday noon; payment run is Thursday | edited · approved 08:41 |
| 03 to 10 | 8 more accounts above $50,000 | $602,000 | 22 to 58 | various | an escalation drafted for each | 7 approved · 1 held for the account manager | |
| 11 to 212 | 202 accounts, the long tail | $1.9M | 1 to 90 | various | 140 first reminders · 62 already contacted this cycle | by rule · A3 |
The board at 08:30: the ranking, the likelihood, the last contact and what the agent proposes on each row; the long tail went by rule, the nine above $50,000 waited for a person. Showcase estate; designed, yet to be built.
The market runs two levels at once: execute within rules (A3) for reminders, dunning e-mails and long-tail outreach, where vendors send messages and voice calls automatically and record promises, and draft-and-approve (A2) for strategic accounts, holds and escalations 11 24 27 29. Oracle describes its collections application progressing routine actions within guardrails and surfacing exceptions to people 14. Lima’s blueprint designs collections and dunning, worklist prioritisation and promise-to-pay tracking at A3, which matches the market’s level for the long tail; the value threshold in the policy is what keeps the strategic accounts at draft-and-approve inside the same agent, the way the $25,000 threshold routes lines to the buyer in chapter 5.
| Element | Setting | Set by | Changes |
|---|---|---|---|
| Objective | Bring every overdue account to payment or to a decision, with the right contact at the right time, inside the collection policy | Credit manager, in a sentence | New version |
| Data scope | Open items and aging, dunning history, promises, disputes, credit exposure, contacts and the payment prediction score in SAP S/4HANA, Oracle Fusion and the CRM | A. Novak, data lead | New version |
| Rules | The collection strategy by customer group and dunning level, and its planning table, version 1 | Credit manager with the Planner Agent | New version, simulated first |
| Trigger | The aging run at 05:00 · an invoice passing its due date · a broken promise · a dunning date · a risk signal | Credit manager | At once |
| Actions permitted | Send a reminder or dunning notice · log a promise · open a dispute case · raise a credit review flag · draft an escalation or a hold request | Credit manager, owner | New version |
| Limits | Escalations above $50,000 drafted for a person · at most one message per account per week · confidence floor 0.75 for a send · holds and payment plans always drafted | J. Harrison, finance lead | New version |
| Escalation | Above $50,000 or a key account, to the collector with the draft · a hold or a payment plan, to the credit manager · a dispute raised on a call, to the account manager | Credit manager | At once |
| Level | A3 Delegated for ranking, reminders and promise tracking; escalations, holds and plans wait for a named person | J. Harrison, finance lead | Up on evidence, down in one step |
| Feedback | Every edited draft, kept or broken promise and granted or refused hold is recorded and used in the next cycle | Always on | Fixed |
Every row is a setting or a versioned rule; the level and the limits are the ones a manager signs.
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.
Additional capability in Order-to-Cash lists the rest of the map for this chain.
| Measure |
|---|
| Touchless collection share: baseline about 50% [3] |
| Average days delinquent: 8.5-day gap to leaders; 85% fewer at leaders [1][3] |
| Bad debt: 1.1-point gap to leaders, up to $44 million a year [3] |
| Past-due balance: 56% to 66% reductions at named adopters [24] |
| Promise-to-pay conversion rate [14] |
| Collector time with customers: 83% more at leaders [1] |
| Days sales outstanding: up to 25% lower at one vendor's customers [31] |
Baseline average days delinquent first, because Hackett measures the gap in it, 8.5 days to leaders and 85% fewer at the companies that redesigned the process (The Hackett Group, 30 March 2026; 15 July 2026), and because it moves within one cycle when the ranking changes. Take it from the ERP’s aging for the last two quarters by customer group, so the long tail and the key accounts can be read apart from the first week.
In this design it writes: reminders, notices and drafted escalations. Calls stay with the collector, with the ranked list and the account’s history in front of them. A voice channel would be a separate action, permitted or withheld per agent.
The dispute case is an object the rule reads. An item in an open case is excluded from reminders by the policy, and the collector sees the case on the row.
By editing the strategy as a new version, simulated first, or by handling an account and recording why. Both are recorded and feed the next ranking.
A reminder brings a reply, and the reply says the invoice was short-paid on purpose. Whose case is that, and how does it get decided?
Short-pays, debit memos and distributor claims, turned into cases with the evidence and a validity call for a named person to approve.
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.