CASE STUDY

Case study: 546 commands in a closed desktop ERP, zero failures

August 2026 · 6 min read · client name under NDA, numbers real

The business problem

A US financial-operations team runs its books in QuickBooks Desktop: a closed, on-premises system with no REST API. Reconciliation meant a person reading bank feeds, hunting the matching invoice, and typing the result in. Volume grew, the backlog grew faster, and every shortcut taken under pressure became an audit finding later.

Why the obvious route was rejected

Piping financial records through a public AI endpoint failed the client's own vendor review: customer names, amounts and bank data would leave through a channel nobody could audit. The requirement that shaped everything: the data stays inside a perimeter someone answers for. That is the sovereign gateway pattern, the same one from our gateway series.

The architecture

Five agents plus one orchestrator, each with a narrow job: read the payment, shortlist candidate invoices, decide the match, post the command, verify by re-reading what the ERP actually recorded.
The window layer: QuickBooks Desktop has no API, so commands go through its native automation interface, one command at a time, verified after each write.
The brain is the iSol API (OpenAI standard), so every decision runs on sovereign infrastructure with per-key budgets and one audit log.

The decision, live

This is the heart of the loop, run today against the live gateway: one payment, three candidate invoices, one JSON decision with a confidence gate. Below 0.85, a human decides:

Real reconciliation decision: INV-2201 matched at 0.95 confidence, action post — Real call, captured 2026-08-19. Note the near-miss traps in the candidate list: a $2.50 difference and a similar payer name. The model picked the exact match and said why.

Real decision run: one payment, three candidate invoices (two of them near-match traps) and the JSON with confidence and action. 7,0s medidos.

Results in production

US$ 10M already processed by the squad, independently, with 25,870 entries completed by the posting agent.
546 commands executed in the ERP, zero failures, every write verified by re-reading.
Reconciliation confidence went from 75% to 90% measured against the official ledger.
Exceptions now arrive pre-analyzed: the human reviews the model's reasoning instead of starting from a blank screen.

What made it work (the consultative part)

Confidence thresholds beat blind automation. The agent posts only above 0.85; everything else becomes a queued human review with the reasoning attached.
Re-read after every write. The agent trusts the ERP, not itself: it re-reads what was recorded and compares. That is where "zero failures" comes from.
Narrow agents, one orchestrator. Five small jobs that can each be tested beat one giant prompt that cannot.

Replicate the pattern

Wire your own flow with the gateway pattern, and if your entry point is automation, start from the n8n or Power Automate guides. The decision prompt above is reusable as-is.

Ready to try it? Your key lives in the customer portal, isollm.ai/account → API keys. The endpoint is https://tokens.4clouders.com/v1. Prefer the outcome delivered instead of building it: that is exactly what iSol Squad is.

See plans →