AI Agents for Accounting & CPA Firms

How the money actually works here

Realisation rate — what you collect against what you booked — is the number that decides whether a firm is healthy, and it dies in preparation rather than in review. Every hour a bookkeeper spends chasing a client for a bank statement is an hour billed at a rate that assumed it would be spent on the return. The gap compounds during season, when the same chase is running across every client at once and the staff doing it are the ones you need for the work. Preparation drag is not an overhead. It is where the margin goes.

For scale: the median firm of this size does $137,766 per employee a year, on 2022 Census figures. The revenue per employee benchmark has the same figure split by headcount band. These are averages rather than medians, so the largest firm in the band pulls them upward.

The rhythm this has to fit

This is one of the few industries where the calendar decides everything. Nine months of manageable load and then a compression season where the constraint is not skill or capacity but the fact that clients respond slowly and deadlines do not move. That shape argues for a specific approach: build for the chase, and build it before the season rather than during it. An agent introduced in February is a distraction. The same agent running since October is the reason February is survivable.

Where the repetitive load sits

Three roles absorb most of it. Median wages are from the US Bureau of Labor Statistics, grossed up by the 1.43× private-industry loaded-cost ratio:

That is roughly $261,475 of annual compensation covering work that is, in large part, the same sequence repeated. 41% of that — the share Slack’s Workforce Lab found desk workers spend on work that is “low value, repetitive or lacks meaningful contribution to their core job functions” (n=10,281, six countries) — is about $107,205 a year.

Read across those three roles, the work itself is accounts payable, accounts receivable, financial reporting, order management, quote to cash and compliance reporting. Each of those is a loop with its own failure points, and the page for it says where the loop breaks rather than where the job title sits.

Against BreathingRoom at $30,000 a year, recovering just 28% of that covers the engagement.

What we would build first

• Client document chasing before deadlines
• Transaction categorisation with a review queue
• Recurring close schedule assembly

Which goes first is decided in the audit rather than in advance. Anyone who tells you the order before looking at your volumes is guessing.

How this goes wrong in accounting & CPA firms

Transaction categorisation is the obvious build and the quiet risk is a confidently wrong posting that compounds. An agent categorises a client's vendor correctly for eleven months, the vendor changes what it sells, and the agent keeps filing it the old way because the name did not change. Nobody catches it until a return is being prepared. The guard is not a higher confidence threshold — it is a periodic sample audit of what posted automatically, not only of what landed in the review queue. A queue shows you the decisions the agent already knew were uncertain.

When you should not do this yet

If your client base is small enough that partners know every account personally, the arithmetic is thin. This works on volume and on repetition — the same chase, the same categorisation, across many clients. A firm with fifteen clients and a partner who knows all of them will find the review overhead eats the saving. The threshold is not a client count so much as a question: is anyone doing the same administrative thing more than a hundred times a month?