From 30 manual workflows to an AI platform the firm actually trusts
An accounting firm runs 30+ operational workflows across its connected tools — reconciliation, unbilled-work detection, document follow-ups, client replies, and morning briefings — with review and approval built into the process.
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 defined the implementation priorities.
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.
When we shipped, it was behind an approval queue. Routine work could proceed automatically; higher-risk changes waited for a human. The system kept a record of each decision and its reasoning.
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 supports 30+ workflows across 9 systems, processing 500+ workflow runs 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 reviewed decision 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.”