Running a finance practice on AI: what to hand over and what to keep

If you run a small practice (fractional CFO, bookkeeping, a few people doing client accounting), AI changes the shape of a week more than it changes any single task. Drafting, first-pass review and explanation get cheap. Judgment and sign-off don't. The practices that do well with it are the ones that decide, on purpose, which is which.

This is how we think about it, and how we run our own books.

Hand over the drafting

AI is good at producing a first version you then correct:

• Commentary and summaries. Variance commentary, the monthly note to the owner, the paragraph on the board slide. See the owner summary and variance commentary prompts.
• Support memos. Turning your working notes into a memo a reviewer can follow: the journal entry memo.
• Client explanations. "Why don't profit and cash match?" answered with their numbers, in their words: explain profit vs. cash.
• Checklists and procedures. Reviewing a close checklist for gaps, or writing up a procedure you've only ever done from memory: close checklist gaps.

The time saved here is real, and it compounds across clients.

Hand over the first pass of review

AI is a patient second reader:

• Unmatched items in a reconciliation, sorted into timing, probable errors and questions for the client: bank reconciliation exceptions.
• Formula errors in a model before it goes to a lender: formula audit.
• Anomalies: an expense in the wrong month, a vendor suddenly billing twice as much.

A first pass finds things. It doesn't clear them. Everything it flags is a lead for a person to follow.

Keep the pen

Some things stay with a named person, and it's worth saying so to clients:

• Posting. An AI can propose an entry. A person posts it.
• Closing a period. The close is a sign-off that the numbers are right. It needs a signature, and a signature needs a person.
• Judgment calls. Estimates, reserves, revenue timing, anything where two competent accountants could disagree. AI can lay out the options; it shouldn't pick.
• Advice. What the owner should do about the numbers. The assistant drafts the words; the recommendation has to be one you'd defend.

Handle client data deliberately

Before a client's data goes near an assistant, decide:

• Which tools are approved. Business and enterprise plans usually have different data terms from personal ones. Read them, pick one, and use it for client work.
• What gets pasted. Descriptions and amounts are usually enough. Account numbers, tax IDs and anything under an NDA usually aren't needed.
• What the engagement letter says. If you use AI on client work, say so, in a sentence that names what it's used for and who reviews its output.

Tell clients plainly

Clients are going to ask. A good answer is short: "We use AI to draft and to check. A person on our team reviews everything, and nothing is posted or sent without their sign-off." That's also a better pitch than "AI-powered". It says what they get: faster work, and the same person accountable for it.

When you're ready to connect the books

Everything above works in a plain chat window, with data you paste in. The next step is an assistant connected to the books themselves, which reads the ledger instead of an export and keeps working across every client. That's RoboLedger, and it's what we teach practices to run. The connected skills show the difference. We'll walk you through it on a screenshare, with your team at the keyboard.