How We Built an AI Assistant for Our Bookkeeper (And Why You Should Too)
This isn’t a thought experiment. We actually did this.
Our CFO, Michael Dean, manages books for multiple clients. He holds an EA (Enrolled Agent), PHR (Professional in Human Resources), and M.S.A. in Accounting. The man knows his craft. But even the most credentialed professional in the world can drown in data entry.
So we built him an AI assistant. We call it Ledger Blu. It connects directly to QuickBooks Online, categorizes transactions, generates reports, flags anomalies, and drafts client communications. Michael reviews everything before it goes out the door.
Here’s the full story — what worked, what didn’t, and how you could do something similar.
The Problem
Michael’s workflow before AI looked like this:
Monday: Log into QuickBooks for Client A. Download bank feed. Manually categorize 80-150 transactions. Cross-reference vendor names against the chart of accounts. Flag anything that doesn’t match. Reconcile the bank statement. Move to Client B. Repeat.
Wednesday: Generate P&L reports for three clients. Format them. Write summary emails. “Revenue is up, expenses are flat, here’s what I’m watching.” Copy-paste data between QBO and Excel. Build the narrative. Send.
Friday: Chase down missing receipts. Follow up on outstanding invoices. Check compliance deadlines. Prepare for the next payroll cycle.
The pattern was clear: 60-70% of Michael’s week was spent on tasks that required his attention but not his expertise. Categorizing a $47 Staples charge as “Office Supplies” doesn’t need an Enrolled Agent. Preparing a tax strategy does. But the Staples charge and its 150 friends were eating Michael’s day before he could get to the strategic work.
We needed to flip that ratio.
The Solution: Ledger Blu
Ledger Blu is a Claude AI instance configured specifically for bookkeeping operations. It’s not a generic chatbot with a “be a bookkeeper” prompt pasted in. It’s a purpose-built assistant with:
- Access to QuickBooks Online data via API integration
- Client-specific context about each business’s chart of accounts, vendor patterns, and financial history
- Bookkeeping-specific training in categorization rules, reconciliation procedures, and financial reporting standards
- Hard boundaries on what it can and cannot do (more on this below)
Ledger Blu sits inside a larger framework we built called AccountDean — 27 specialized AI agents across 6 divisions, all reporting to Michael. But Ledger Blu is the workhorse. It handles the daily operations.
How We Built It
Step 1: The QBO Connection
The technical foundation is an MCP (Model Context Protocol) server that connects Claude to the QuickBooks Online API. MCP is essentially a way to give an AI model access to external tools — in this case, the ability to read and interact with financial data.
Our MCP server provides tools like:
– list_invoices — Pull invoices filtered by customer, date, or status
– get_profit_loss — Generate a P&L for any date range
– get_bank_transactions — Read bank feed transactions
– search_transactions — Find specific transactions by amount, date, or memo
– get_balance_sheet — Pull a balance sheet as of any date
– get_ar_aging — Accounts receivable aging report
The server authenticates to QuickBooks using OAuth 2.0 — the same secure handshake your bank uses when you connect Mint or Plaid. Each client’s QBO instance is a separate connection with its own credentials, completely isolated.
Tech stack: TypeScript/Node.js MCP server, Intuit OAuth 2.0 client library, QuickBooks Node SDK, Anthropic MCP SDK.
Cost to run: $0 in additional infrastructure. It runs locally. The QBO API is free if you already have a QuickBooks subscription.
Step 2: The Bookkeeping Brain
Connecting to QBO gives Ledger Blu data. The next step was giving it knowledge.
We configured the AI with:
- Client profiles — Each client’s chart of accounts, common vendors, typical transaction patterns, and financial goals
- Categorization rules — “AMAZON MKTPLACE for Client A is Office Supplies. For Client B, it’s Inventory. For Client C, ask Michael.”
- Reporting templates — How Michael likes his P&Ls formatted, what metrics matter to which clients, how board reports should read
- Communication style — Professional but warm. No jargon unless the recipient speaks accounting.
This isn’t magic. It’s configuration. The AI is powerful because we gave it the right context, not because it figured everything out on its own.
Step 3: The Guard Rails
This is the part that matters most, and the part most people skip.
What Ledger Blu CAN do:
– Read financial data from QBO
– Categorize transactions (as drafts for review)
– Generate reports
– Flag anomalies and discrepancies
– Draft client emails and meeting summaries
– Research vendor history and patterns
What Ledger Blu CANNOT do:
– Move money
– Execute payments
– File tax returns
– Access bank accounts directly
– Send client communications without Michael’s approval
– Make tax strategy decisions
– Sign anything
Every output is a draft. Every action that touches real money requires a human in the loop. This isn’t a limitation we’re working around — it’s a design principle. Michael’s EA credential means he has a professional and legal obligation to review financial work. The AI makes that review faster and more thorough, not optional.
The Results
We’re still early — iterating weekly, adjusting the system, learning what works. But here’s what we’re seeing:
Time Savings: 5-10 Hours Per Week
The bulk comes from three areas:
– Transaction categorization — What took 2-3 hours per client now takes 20-30 minutes of review time
– Report generation — P&Ls, balance sheets, and cash flow statements generated in seconds instead of manually built in Excel
– Client communication — First drafts of monthly summaries, follow-up emails, and meeting prep written by AI and refined by Michael
Accuracy Improvements
The AI doesn’t get tired on Friday afternoon. It doesn’t accidentally put a utility bill under “Professional Services” because it was rushing to close the month. Its categorization accuracy after training on a client’s data runs north of 90%, and Michael’s review catches the rest.
The anomaly detection has already flagged two duplicate charges and a subscription renewal that the client had canceled. Those are real dollars saved.
The Ratio Flip
This is the big one. Michael’s time allocation is shifting from:
– Before: ~65% data processing, ~35% advisory/strategy
– After: ~35% review and data processing, ~65% advisory/strategy
That means more time for tax planning conversations. More time for financial strategy sessions. More time for the work that actually requires an Enrolled Agent — not just any warm body with a QuickBooks login.
Lessons Learned
AI Is Great at Pattern Matching, Terrible at Judgment
Ledger Blu can categorize 500 transactions in the time it takes to drink a cup of coffee. But ask it whether a home office expense should be classified under the simplified method or actual expense method for a client who also runs a daycare out of the same space? That’s Michael’s call. Every time.
The line between “pattern” and “judgment” is the line between what AI should do and what humans must do. We built our system around that line.
Always Have the Human Review
We learned this the hard way with a categorization that was technically correct but contextually wrong. A client’s payment to a contractor was auto-categorized as “Subcontractor Expense” — which is accurate — but Michael knew from a phone conversation that the contractor’s role had changed and the expense needed to be recategorized for 1099 reporting purposes.
The AI had the data. Michael had the relationship. Both are necessary.
Start With Low-Stakes Tasks
We didn’t start by having AI generate tax returns. We started with transaction categorization — the most repetitive, lowest-risk task in the workflow. If the AI miscategorizes a $12 lunch charge, Michael catches it in review and corrects it. No harm done. The AI learns from the correction.
We graduated to report generation only after categorization was reliable. We’ll move to more complex tasks only when the simpler ones are solid.
The Setup Investment Is Real
Building this wasn’t zero effort. The MCP server needed development. Client profiles needed to be written. Categorization rules needed to be configured. The OAuth connection needed security review.
For us, the engineering investment was internal — we have the team to build it. For most small businesses, the build would need to come from a provider. That’s part of what we offer at FIT.
How You Could Do This Too
You don’t need to build a custom MCP server to start using AI in your bookkeeping. Here are three practical steps, from easiest to most involved:
Step 1: Start With AI-Assisted Categorization (This Week)
Open Claude, ChatGPT, or whatever AI tool you’re comfortable with. Export your last month’s transactions from QuickBooks as a CSV. Upload it and ask: “Categorize these transactions based on the vendor name and amount. Use these categories: [paste your chart of accounts].”
Review the output. You’ll be surprised how accurate it is — and where it’s wrong will teach you what context the AI needs.
Time investment: 30 minutes.
Cost: Free (most AI tools have free tiers that handle this).
Step 2: Build a Client Profile Document (This Month)
Write up a one-page profile for each client or business entity you manage. Include:
– Chart of accounts with descriptions
– Common vendors and how to categorize them
– Recurring transactions (subscriptions, rent, payroll)
– Special rules (“Anything from Home Depot is Materials, not Office Supplies”)
– Reporting preferences
Feed this to your AI tool as context before asking it to categorize or generate reports. The quality of AI output is directly proportional to the quality of context you provide.
Time investment: 1-2 hours per client.
Cost: Free.
Step 3: Connect AI to Your Financial Data (This Quarter)
This is where it gets powerful — and where you likely need help. Connecting AI directly to QuickBooks via API means no more exporting CSVs. The AI can pull live data, run reports on demand, and work with current numbers instead of last week’s export.
Options:
– DIY: If you have a developer on staff, MCP servers for QuickBooks are buildable with open-source tools. The Intuit API documentation is solid.
– Provider: Work with an IT partner (like us) who can build and maintain the integration for you. We’re doing this for our own clients and can extend it.
– Wait: This space is moving fast. More turnkey solutions will emerge. But the early movers will have months of trained AI context that newcomers won’t.
Honest Caveats
We’re being transparent about where we are:
This is still early. We’re iterating. The system gets better every week as Michael provides feedback and we refine the configuration. Six months from now it will be significantly more capable than it is today.
AI hallucinations are real in finance. The AI can occasionally generate a plausible-sounding number that is flat wrong. This is why human review isn’t optional — it’s structural. We’ve built our workflow around the assumption that the AI will make mistakes, and Michael’s job is to catch them.
Not every bookkeeper wants this. Some bookkeepers see AI as a threat. We see it differently. The bookkeepers who adopt AI tools will handle more clients, deliver faster service, and focus on higher-value work. The ones who don’t will compete on price with a machine that works for pennies. We’d rather give bookkeepers superpowers than watch them get sidelined.
The Bigger Picture
What we’re building with Ledger Blu and the AccountDean framework is a proof of concept for something larger: small firms punching above their weight class.
Michael is one person. With AI augmentation, he operates like a team. He provides the expertise, the credentials, the judgment, and the client relationships. The AI provides the speed, the pattern matching, the tireless data processing, and the first-draft everything.
A solo EA with 27 AI agents across 6 divisions isn’t a gimmick. It’s how small firms will compete with large ones — by matching their throughput without matching their overhead.
We’re not just building this for Michael. We’re building it so we can offer it to every FIT client who needs bookkeeping support. Enterprise-grade financial operations at small-business pricing, backed by a real credentialed professional who reviews everything.
That’s the model. And it works.
Want to see what AI-augmented bookkeeping looks like for your organization?
Contact Flower Insider Technologies — we’ll walk you through exactly how it works, what it costs, and whether it’s the right fit.
Matt Stoltz is the founder of Flower Insider Technologies, providing managed IT and AI-augmented business services for small businesses and nonprofits. FIT’s CFO Michael Dean (EA, PHR, M.S.A.) leads financial operations, powered by the AccountDean AI framework — 27 specialized agents, one credentialed professional, zero corners cut.