A 15-person accounting firm (8 accountants, 7 support staff) was turning away clients. Each accountant managed 35-40 clients, spending 60-70% of time on recurring bookkeeping, queries, and report generation.
The deployment
- Nathan (Bookkeeping): Pulls bank feeds, categorizes transactions using client-specific patterns, reconciles balances.
- Ethan (Client Advisory): Handles routine queries using real data from QuickBooks and Xero. Holds advisory-adjacent responses for accountant approval.
- Iris (Reporting & Deadlines): Generates monthly management accounts and advisory dashboards.
Results (120 days)
| Metric | Before | After | Change |
|---|---|---|---|
| Clients per accountant | 35-40 | 95-110 | 3x |
| Weekly bookkeeping hours | 22 hours | 6 hours (review only) | -73% |
| Client query response time | 4-8 hours | < 15 minutes (routine) | -96% |
| Month-end close time/client | 4-6 hours | 1.5 hours | -70% |
| Revenue per accountant | Baseline | +85% | Growth |
The capacity unlock
Accountants shifted from data entry to reviewing outputs, advising clients, and building relationships. The firm launched an advisory services tier that didn't exist before, commanding higher fees. Revenue per accountant increased 85%.
What didn't work perfectly
Nathan's categorization accuracy started at 82%, improving to 94% by week 8 through the feedback loop. For clients with unusual patterns, accuracy plateaued at 88%. Ethan required careful boundary management — an early query about depreciation crossed into advisory territory, prompting tightened boundary rules.
The hiring decision
After 120 days, the firm hired — but a client relationship manager instead of a bookkeeper. The AI workforce changed not just capacity but the hiring profile.
Explore what this would look like for your firm. Learn about our AI agent development approach.