Building a Finance Function Around an AI-Enabled Accountant

The Situation

Snyder made a minority investment in a regional managed IT services company. Part of what we brought was operating discipline, and the most urgent place to apply it was the books.

When I assessed them in April, the picture was common for a small, founder-run company:

• Cash-basis books, with all revenue in a single income account
• No accounts payable and no vendors set up
• Five bank and credit accounts, none of which had ever been reconciled
• Uncleared items going back four years, and a balance sheet that showed negative cash
• Sales tax not filed

I wrote a bookkeeping plan with a four-phase approach and named the biggest risk: the reconciliation backlog. The question was who would do the work.

Hiring for a Different Job

We decided not to hire a bookkeeper to key transactions. We wanted an accountant who could clean up the books and then put automation in place, someone who would orchestrate and audit rather than type. I screened candidates and coached the finalist before the executive interviews, and we hired an offshore staff accountant with about nine years of US bookkeeping and tax experience. He knew QuickBooks well and had never used Git or AI tools.

Day One, With AI

The training plan had two stages: first show what the tools can do somewhere safe, then graduate to working the way the rest of the team works.

A local repository first. From day one he worked with AI developer tools in a local repository of his own. We spent two hours together setting it up and working through real tasks: building a knowledge base in Markdown, turning meeting transcripts into notes and tasks, and asking the AI to read and reason across the files. The point of starting local was to let him see what the tools are capable of, and make mistakes freely, before anything he did touched shared work.

Then a shared repository with Git. Once that way of working was second nature, he moved into the shared executive repository. That meant learning Git: branches, commits, and opening pull requests for review. Within about six weeks, an accountant who had never used version control was submitting his work as pull requests, with policies, procedures, analyses and agent skills going through the same review as everything else in the repository.

To support both stages, I handed him a package of about 90 files:

• How to work: when to use a conversational assistant, AI developer tools in the repository, or an AI spreadsheet assistant; raw notes, then summaries, then overviews; documenting inside the spreadsheet because chat history disappears
• Rules for AI assistants working in his knowledge base
• Rhythm and cadence from daily to annual
• Systems and access for both companies, plus a glossary
• Finance policies adapted from Snyder’s own policy set, with a tracker of the procedures still to write

Access was staged to match. He started in his local repository, received access to the shared repository at the end of his first month, and opened his first pull requests about six weeks in.

The Program

My rule for pacing was to deliver outputs leaders can make decisions with before starting any big cleanup. In week one he produced a provisional P&L and a weekly cash summary. The Monday cash snapshot was routine by week three.

Then came the maturity work. He did most of it and I scoped and reviewed it:

Area	What was done
Audits	Receivables compared invoice by invoice between the billing system and the books; payroll mapping reviewed line by line, with corrections handed to the bookkeeper; every invoice matched to a contract; sales tax exposure analyzed; a note payable reviewed
Books cleanup	All five accounts reconciled through August, with every stale item dispositioned and a before-and-after letter for leadership
Process	SOPs for invoicing, forecasting, cash flow, check deposits and mail handling
Forecasting	Revenue built from recurring, open orders, quotes and ad-hoc work; a monthly loop of forecast, close, variance trace and refine; a Friday revenue check-in
Executive reporting	The accountant presents the financials for the first part of the weekly executive meeting; I keep the running agenda, whose finance section comes from his backlog

The first forecast cycle landed within about 8% of actual, and the variance trace found the miss was mostly timing.

Guardrails

Giving an accountant AI tools and repository access needed explicit limits:

• Everything is a draft. Nothing is posted to the books without the company’s bookkeeper, and every AI output is marked for human review.
• Scope in writing. Before the accounting system was connected to AI, he wrote a plan of what he would and would not use AI for.
• No binary office files in the repository. I converted the working spreadsheets into dated CSV snapshots with a manifest, so the finance work is readable by AI and reviewable in diffs while the live spreadsheets stay in their own systems.
• Least access. Admin rights to systems like payroll stayed with the people who own them.

Within those limits he built a library of agent skills for routine accounting work, such as suggesting journal entries, payroll mapping, forecast variance bridges and revenue forecast builds. Our CEO reviewed them.

Current Status

About three months after his first day, the books are current and reconciliations are part of the monthly close. The chart of accounts is being redesigned next. The reconciliation pack from the cleanup is now the template for the first diagnostic at the next company we invest in.