Seed
Build two synthetic ledgers with realistic break patterns: timing differences, duplicates, orphans, fee drift.
02 / AI case study / Finance + revenue operations
Designed · not yet builtThe Reconciliation Sentinel
Two systems that should agree, reconciled nightly by agents, with every mismatch explained in plain English and routed to someone who can fix it.
PROVEN EXPERIENCE → NEW BUILD
Rooted in a reconciliation I ran by hand
At Nimbus USA I led a four-to-five-month reconciliation with our largest dealer until both organizations agreed on what had been paid and what remained due. This case study automates the pattern I ran manually, and keeps the human judgment where it mattered.
THE BRIEF
An automated reconciliation layer between two synthetic books of record: agents compare balances and transactions, classify every discrepancy, explain each one in a sentence a human can act on, and track recurring root causes over time.
The problem
When billing, CRM, and accounting disagree, someone spends nights in spreadsheets finding out why. Worse, sometimes no one does, and the gap compounds. Most reconciliation tooling flags that records differ; it cannot say why, or what to do about it.
The system
Agents run the comparison on a schedule, bucket each mismatch by pattern (timing, duplicate, missing counterpart, amount drift), draft the plain-English explanation with linked evidence, route fixable items to an owner queue, and maintain a running root-cause ledger that shows which upstream process keeps creating the same break.
- Two books of record
- Matching engine
- Discrepancy classification
- Plain-English explanation
- Owner routing
- Root-cause ledger
Claude agentsPower AutomateDataverseSynthetic ledgersPower BI
DESIGNED OUTCOMES
What “working” will mean.
- Nightly reconciliation without the nights
- Every discrepancy explained, not just flagged
- A routed work queue instead of a spreadsheet hunt
- Root-cause trends that point at the upstream fix
MEASUREMENT PLAN
How it gets judged.
- Match precision against seeded truth
- Explanation usefulness rating
- Auto-classification rate
- Time to clear a break
- Repeat root causes caught
THE BUILD PLAN
How this gets built
when it gets built.
Match
Agent matching and classification, with confidence scores and linked evidence per finding.
Explain + route
Explanation drafting, owner queues in Dataverse, and the recurring root-cause tracker.
Prove
Seeded-truth evaluation, Power BI trend views, and the written case study.
Transparency note. This is a self-directed AI case study built on synthetic data. Results will be reported as evaluation findings against seeded ground truth, not client outcomes. When the build is complete, this page gains: working reconciliation pipeline, discrepancy queue with explanations, seeded-truth evaluation set, root-cause trend report, runbook, written case study.
NEXT CASE STUDY / 03The Shared-Inbox Triage Agent
→