From 30 manual workflows to an AI platform the firm actually trusts
An AI operations platform running inside an accounting firm handles 30+ operational actions across its connected tools — reconciliation, unbilled-work detection, document chasing, client replies, morning briefings — every one approved by a human before it happens.
The problem
The firm’s work was spread across email, QuickBooks, spreadsheets, document storage, and chat. Staff lost hours every day moving context between tabs: triaging a shared inbox, chasing missing receipts, coding transactions, and hunting for unbilled work at month-end.
They didn’t want a chatbot that answered questions about their books. They wanted the manual work itself to get done — safely, without handing a black box the keys to client data and money.
The approach
We started by mapping the firm’s systems and workflows: where context got lost, where people re-keyed numbers, where revenue slipped. That became the action list.
Every agent then ran in shadow mode for weeks — reporting what it would do, without doing it — so the team could judge its decisions on real client accounts before anything went live. Trust was built before a single action ran.
When we shipped, it was behind an approval queue. Low-risk actions could auto-complete; high-risk ones waited for a human. Every action, and the reasoning behind it, was logged.
What it runs today
Reconciliation, proposed: the agent matches transactions and proposes entries; the accountant approves or rejects with the reasoning in front of them.
Unbilled-work detection: work with no matching invoice is surfaced before it’s written off, and turned into a draft invoice for approval.
Document chasing: outstanding receipts and forms are tracked per client, with follow-ups drafted for review.
Morning briefings: every day starts with what needs attention — transactions to code, clients to chase, deadlines approaching — ranked, with sources cited.
The result
The platform now handles 30+ distinct actions across 9 systems, processing 500+ actions a week with 95% routing accuracy. Intent is classified in under 200ms, and when the system isn’t sure, it asks instead of guessing.
Most importantly, the firm trusts it — because it was never asked to. Shadow mode, an approval queue, and a full audit trail meant trust was demonstrated, one action at a time.
“The firms that win with AI won’t be the ones with the fanciest demos. They’ll be the ones whose systems are boring enough to trust.”