Autonomous Reconciliation at the Sub-Ledger
A closed-loop agent matches transactions across four sub-ledgers inside explicit authority bounds — exceptions reach a named reviewer with full provenance, not a queue.
The challenge
Cash, securities positions, corporate actions, and fee accruals were reconciled against the general ledger by hand, one spreadsheet tie-out at a time. That left the close exposed to:
- Roughly 45,000 line items a month spread across four sub-ledgers, each matched manually
- Eleven analysts working parallel spreadsheets with no shared matching logic
- No consistent audit trail linking a match decision back to its source records
- High-value and related-party items reviewed under the same time pressure as routine ones
- A close window that regularly slipped into the following month
At an average 3 minutes per line, manually tying out all 45,000 monthly items consumed roughly 2,250 analyst-hours — more than the 1,848 hours eleven analysts could realistically supply inside a 21-day close window.
How it works
Matching within explicit limits, not blanket automation
The agent doesn't get general authority to reconcile — it gets specific, auditable permission to act inside defined bounds:
- 01
Transaction feeds from all four sub-ledgers and the GL are ingested and normalized daily
- 02
A matching model is trained on three years of historical resolved exceptions, not just rule logic
- 03
Explicit authority bounds are set: no auto-clearing above a per-item dollar cap, no touching related-party or write-off items, no restating closed periods
- 04
Items scoring above the confidence threshold are auto-matched and cleared same-day
- 05
Everything below threshold routes to a named senior reconciliation lead with a provenance packet — source records, matching logic, confidence score, and comparable past resolutions
- 06
Every reviewer resolution is fed back into the model to retrain matching confidence monthly
- 07
Authority bounds and thresholds are recalibrated quarterly against actual reviewer override rates
What we built
Key capabilities
Bounded autonomy, not blanket authority
The agent's authority is defined item by item — dollar caps, excluded categories, and closed-period locks are enforced before a match is even attempted.
Provenance travels with every exception
A flagged item never arrives at a reviewer's desk bare — it carries the source records, the model's confidence score, and comparable past resolutions.
A named reviewer, not a queue
Every routed exception has an accountable owner, closing the gap between 'flagged' and 'someone is responsible for this.'
Retrains on its own exceptions
Reviewer decisions on the hardest 6% of items become the next month's training data, so the confidence threshold keeps tightening.
Before vs after
What changed at close
- Line items needing analyst review
- 45,000 → 2,700
- Analyst hours per monthly close
- 2,250 → 270
- Close completion
- Slipping into next month → inside the 21-day window
- Analysts assigned to these four sub-ledgers
- 11 → 4
Business impact
What it changed
1,980 analyst-hours reclaimed monthly
2,250 hours of manual tie-out fell to 270 once 94% of items cleared autonomously — hours redirected to the exceptions that actually need judgment.
Every exception is provenance-backed
The four analysts still on this desk review 2,700 items a month, each arriving with the evidence needed to decide in minutes, not re-derive from scratch.
Close finishes on schedule
The four sub-ledgers now close inside the 21-day window every month, ending the rollover backlog that used to eat into the next cycle.
“Autonomous doesn't mean unsupervised — it means the boundaries are explicit, the exceptions are named, and the humans left in the loop know exactly why they're there.”
Keep reading