Juno would score every journal as it posts against the company’s own patterns and policy: account combinations, amounts, timing, who posted, round numbers, reversals.
It is the last working day of September, 16:10, and a manual journal of $384,000 posts in SAP S/4HANA to a cost centre at plant 1210 that has carried only recurring depreciation for 14 months. The account pair has appeared twice in two years, and the preparer’s last manual journal was nine months ago. The agent holds it before it reaches the period’s ledger, lists the three reasons and the two prior occurrences, and drafts the reversal.
J. Harrison, finance lead, opens the hold from her queue. She reads the text on the journal and recognises the write-down of the die bank agreed with the plant controller on Tuesday. She releases it with a note in one line, and the journal posts under the preparer’s document number.
The hold, the three reasons, her note and the release sit on one record. The next run gives that account pair a lower weight, so the same write-down in March would post and reach her as a flag to read, while the hold is kept for the patterns the estate has yet to see.
Error and anomaly detection is the finance AI use case with the clearest survey signal: 34% of finance functions use it, third after knowledge management and accounts payable automation (Gartner, 18 November 2025) 2. The pain is measured: 18% of accountants make financial errors daily and 59% several a month, and organisations with high technology acceptance report 75% fewer (Gartner, via Accounting Today, 21 February 2024) 10.
Every ERP vendor and close suite names an agent here. Oracle’s Ledger Agent scans transactions against conditions written in plain language, flags exceptions with supporting details and drafts adjustment journals for review; autonomous execution within guardrails is named as the direction for later releases (Oracle, 15 October 2025; 1 December 2025; 20 September 2026) 16 18 19.
FloQast’s AI Detections flags ledger errors before close (FloQast, 17 September 2025) 36, Trintech scores journal risk (Trintech, 9 April 2026) 34, HighRadius runs 15 or more models over postings and claims 80% of anomalies resolved automatically (HighRadius, 1 April 2025) 37 38, Workday’s Financial Test Agent tests financials continuously (Workday, 16 September 2025) 31, and SAP’s Journal Entry Agent sits inside its Financial Closing Assistant (SAPinsider, 17 September 2026) 13.
Gartner expects AI-driven anomaly detection, continuous controls monitoring and real-time audit logging to become standard in cloud ERP finance (Gartner, 24 February 2026) 1. The customer numbers are vendor claims, which keeps the use case second in the chain. For a manufacturer the journals that matter most are the supply chain ones, inventory write-downs, scrap and yield, subcontract and consigned stock, purchase price variance, and the ontology holds the material, the plant and the production order each one refers to.
Journals post all period long from sub-ledgers, interfaces, recurring runs and people. A manual journal above a threshold gets a second approver in the ERP workflow, which checks authority more than substance. At period end the controller runs the standard reports: large postings, unusual account pairs, late entries, weekend postings, round amounts, postings by users who rarely post.
The reports go into a spreadsheet, a few lines are sampled, and the preparers are asked by e-mail. Errors that slip through surface as reconciliation differences on workday 3, as flux questions on workday 4, or in the auditors’ journal testing after year end, when the period is closed and a correction means a topside entry or a reopened period.
The cost is hours of review at each close, plus every correction found late. People: general ledger accountants, the controller, internal audit, external audit. Systems: the ERP ledger and its approval workflow, spreadsheets and, where one exists, an audit analytics tool.
The period as it runs by hand: the check comes at the end, and the error is found after the close. Illustrative; the mechanics are the note’s.
The six stages carry this use case’s objects, questions, terms, trigger and measure. Everything below is a design: the roles would be configured from the same parts as the PO cancellation agent, and the finance objects (journal entry, ledger account, cost centre, posting period, user role) would be added as an extension of the ontology, which today holds the supply chain objects.
The scene above as the run trace would record it: six steps, one record, with the hold and the release by name. Designed, yet to be built; showcase estate.
The market runs this at inform and draft-and-approve: Oracle’s agent flags and drafts adjustment journals for review 16 and names autonomous execution within guardrails as a later release 19; FloQast and Trintech flag and score 36 34. The blueprint designs the lead role at A3, holding and routing inside versioned rules while a person decides every material correction, and keeps the topside review at A0. A first release would run at A2 and move to A3 once the recorded cases show the precision the SOX team asks for.
| Element | Setting | Set by | Changes |
|---|---|---|---|
| Objective | Hold the unusual journals before the ledger closes; draft the correction; keep every flag and decision on the record | J. Harrison, finance lead | New version |
| Data scope | Journal entries and lines, chart of accounts, period status and user roles in SAP S/4HANA and the Oracle Fusion division | A. Novak, data lead | New version |
| Rules | The anomaly policy and its planning table: manual above $250,000, an account pair unseen in 24 months, a posting after cut-off | J. Harrison with the Planner Agent | New version, simulated first |
| Trigger | Every posting event (an hourly batch where the source only batches) · cut-off runs on the last working day and workday 1 · on demand | J. Harrison, finance lead | At once |
| Actions permitted | Hold a journal · route it to the approver · post an approved reversal or reclassification · log the flag and the decision | J. Harrison, owner | New version |
| Limits | Materiality $250,000 · hold budget 40 a day · confidence floor 0.85 · a correction to a closed period waits for two named people | J. Harrison with the CFO | New version |
| Escalation | A hold untouched for 4 hours on the last working day, to J. Harrison · a failed posting, to the owner · a segregation conflict, to internal audit | J. Harrison, owner | At once |
| Level | A2 Supervised with the journal approver as the named approver; A3 for holds inside the rules once four closes of recorded cases pass | The CFO | Up on evidence, down in one step |
| Feedback | Every confirm, dismiss, edit and release is recorded and reweights the pattern | 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 Record-to-Report lists the rest of the map for this chain.
| Measure |
|---|
| flags per 1,000 journals and share confirmed as errors |
| errors found after close (baseline: 18% of accountants err daily, 59% several a month) |
| manual journal count per close |
| time from posting to detection |
| topside entries and reopened periods per quarter |
| auditor journal-entry testing findings per year |
Baseline errors found after close first: the topside entries, reopened periods and auditor journal findings of the last four quarters, with the hours the controller spends on the period-end reports. Precision needs the agent to exist before it can be measured, so the simulation on the last four closes gives its first reading: how many journals the policy would have held, and how many of those the team recognises as errors.
A hold applies to the journal that broke a rule and to that journal alone. The recurring runs and the interfaces keep posting, and a pattern the team has taught as normal carries a low weight and passes.
The SOX team and internal audit read the flag-and-outcome log each quarter and approve the next rule version. A threshold change is a new version with a named approver, simulated on the last closes first.
The rules engine scores and holds; the small language model inside your estate words the reason. Row data stays inside the estate, and the cloud model sees the policy sentence at design time alone.
A held journal protects the ledger from what was posted. What about the cost that was received in September and has yet to be invoiced or posted at all?
Accruals from open purchase orders and receipts awaiting invoices, proposed with their basis and posted with automatic reversal.
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.
