Bookkeepers should start with Reply Faster: Client Refund Emails, a 10-minute ChatGPT lesson, because the first risk with AI in a bookkeeping practice is not a wrong number in a sheet. It is a confident email that promises a client a refund date you can't promise. After that, the path moves from client communication to the books themselves: meeting recaps, expense categories, receipt capture and reimbursement requests.
The first two lessons use ChatGPT only. The last three use n8n, and the last one adds Lovable. Every lesson keeps a bookkeeper as the reviewer.
Which lesson should you take first?
Take Reply Faster: Client Refund Emails first. You work in the Gmail inbox of Northgate Tax & Bookkeeping, a made-up firm, where a client named Greg wants to know when his refund will arrive while two tax forms are still missing.
You paste the thread into ChatGPT, then a list of facts: what the firm received, what's missing, the turnaround, the signing steps and what the firm can and can't say about refund timing. The first draft sounds calm and professional. Then Paste check builds a table of every claim, and two of the seven don't match a fact: a refund date and a promise to "file within 5 business days." You swap in the safer wording ChatGPT suggests and send.
That catch is why this lesson comes first: every AI draft to a client is unverified until checked against your facts.
What does the rest of the bookkeeper path cover?
The next four lessons move from talking to clients to keeping their records. Each one adds a tool only when the task needs it.
Step 2: How do you turn a client review into a clear follow-up?
Summarize a Client Meeting (9 minutes, ChatGPT) takes the transcript of a Q3 review with Cedar Lane Landscaping, a made-up client, and asks for a summary, the decisions, an action items table with owner and due date, and open questions.
Two due dates come back Not stated. You supply them (the truck paperwork is due Friday, and a quote goes out by October 21), then ask for a follow-up email to the client. You leave with a recap of who owes which document by when.
Step 3: Can AI sort expenses into your categories?
AI Expense Categorizer Agent (12 minutes, n8n plus the OpenAI API) builds an agent for Maple Street Bookkeeping, a made-up firm. Its system prompt allows only six categories, from Office Supplies to Needs Review, and tells it to use Needs Review whenever an expense could fit more than one category or the details are unclear.
You give it three expenses (printer paper, accounting software and a client lunch), read the categories and notes, then tell it to save them to an Expense Log sheet with Date, Vendor, Amount, Category and Note. This lesson also walks you through creating an n8n account and an OpenAI API key, so take it before the receipt lesson.
Step 4: Can AI read a receipt photo?
Yes. AI Receipt Extractor (12 minutes, n8n plus the OpenAI API) has an agent on gpt-4o read a receipt image you attach in the n8n chat, pull out the date, vendor and amount, pick a category, and append a row to an Invoices sheet in a shared Bookkeeping folder.
Receipts that arrive as phone photos become rows you can review. The lesson lets the agent invent a category; in your own version, give it your list the way Step 3 did.
Step 5: Can reimbursement requests check themselves against a policy?
AI Expense Reimbursement Form (15 minutes, Lovable plus n8n) builds an Expense Reimbursement Request form. The agent checks each request against a simple policy and picks one status: Needs receipt, Needs manager OK for $100 or more, or Approved under $100 with a receipt and a clear business purpose. It saves the request to a sheet and emails the employee the status. The system message ends with "Do not give tax advice."
It is the most complete build in the path, with a person still approving anything over the line.
How long does the path take, and what will it cost?
The five lessons total about 58 minutes. A practical pace is the two ChatGPT lessons in one sitting, then one n8n lesson a week, outside your busiest close or filing weeks.
When you run these on real work:
- Steps 1 and 2: ChatGPT only, no extra accounts. The free plan works to start; business plans add data controls many firms want.
- Steps 3 and 4: an n8n account, a Google account and an OpenAI API key. The API is pay-as-you-go and billed separately from ChatGPT plans. n8n Cloud has paid plans after a trial, or it can be self-hosted.
- Step 5: adds a Lovable account, which has paid plans.
Check each vendor's pricing page before you commit.
What should a bookkeeper do after the path?
Run one workflow on a single client's real data for a month, with you reviewing every row and every email, before you widen it. Compare the AI's categories against your chart of accounts and keep a list of the ones it gets wrong.
Prompts that fit bookkeeping: request missing documents, reply about a refund's status, categorize expenses, extract receipt details and explain a reimbursement decision. Terms worth knowing: expense categorization, chart of accounts, expense reimbursement, OCR, data extraction and PII.
The finance hub and the accounting and bookkeeping firms hub collect more examples. For further reading, see AI for Bookkeepers: Where to Start, How to Email Clients for Missing Tax Documents, Get Receipts into Google Sheets with AI and n8n and Automate Expense Reimbursement Requests with AI.
What should a bookkeeper not hand to AI?
Don't hand AI tax positions, deductibility calls, refund dates or anything you would sign. AI sorts, extracts and drafts; you decide and post.
- Tax questions. Whether an expense is deductible, how a meal is treated or when a refund will arrive are questions for a CPA or enrolled agent, not a chatbot. Have a licensed professional review anything with tax consequences.
- Posting without review. The Maple Street agent's system prompt says the audience is the bookkeeper who reviews every entry before it is posted. Keep it that way.
- Client data. Never paste Social Security numbers, EINs with names attached, full account numbers or bank login details into a chat. Mask client names where you can, use a business plan with the right data controls, and check your engagement letters and any professional rules that apply to client confidentiality.
- Reimbursement approvals. Let the form sort requests, but a person approves anything that needs a manager's OK.
Key takeaways
- Bookkeepers should start with Reply Faster: Client Refund Emails, a 10-minute ChatGPT lesson that catches a refund date the firm can't promise.
- The path then covers client meeting recaps, an expense categorizer with a Needs Review category, a receipt photo extractor and a policy-checking reimbursement form.
- The two ChatGPT lessons need no extra accounts; the n8n lessons need an n8n account and a pay-as-you-go OpenAI API key, and the last one adds Lovable.
- Give AI a fixed category list and a way to say "Needs Review" instead of letting it guess.
- Tax treatment, refund timing and posting decisions stay with the bookkeeper and a licensed tax professional.
Frequently asked questions
- Why is a client email lesson first instead of receipts?
A wrong row in a sheet gets caught at review. A wrong promise in a client email is already in the client's inbox. Learning to check AI's claims first protects you in every later lesson.
- Can the expense agent use my own chart of accounts?
Yes. In the lesson the categories are a list in the system prompt, so you can replace them with your own account names and keep a Needs Review option for anything unclear.
- Is it safe to send client receipts to an AI model?
It depends on what's on them and your plan's data controls. Remove card and account numbers where you can, use a business plan, and follow your firm's confidentiality rules.
- Do I need Lovable for the reimbursement form?
In this lesson, yes: Lovable builds the form and n8n handles the agent, the sheet and the email. The other four lessons don't use Lovable.