To use AI agents for bookkeeping services, write down the firm's procedures and each client's rules, start the AI on read-only reports for one client, and keep a named person as reviewer for every entry and every sign-off. This playbook is for owners and leads of bookkeeping practices and accounting firms with many clients. It covers which work to hand over first, how to write rules the AI can follow, access, review roles, a weekly and month-end rhythm, client conversations, how to measure results and a phased rollout.
Still choosing a tool? Read the side-by-side of the best AI tools for accounting firms. This article is about running one.
What is an AI agent in a bookkeeping practice?
An AI agent here is an AI app, such as Claude or ChatGPT, connected to a client's ledger so it can read reports and transactions and, if you allow it, record entries. Because it can change the books, the controls below matter more than the prompts. New to the idea? Start with the AI bookkeeping guide.
Which bookkeeping work should you give AI agents first?
Start with work that is easy to check against the ledger and cheap to get wrong. Keep judgment and sign-off with people.
| Work | Give to the AI? | How to start | Who stays responsible |
|---|---|---|---|
| Report questions and variance explanations | Yes, first | Read only | Bookkeeper checks totals against the report |
| Duplicate bill and invoice checks | Yes, first | Read only | Bookkeeper decides what to void |
| Bank feed coding for known vendors | Yes, second | Proposals, then reviewed writes | Bookkeeper approves each batch |
| Supplier bills from documents | Yes, third | Drafts a person approves | Bookkeeper |
| Receivables follow-up | Drafts only | Aging report plus draft emails | A person sends every email |
| Reconciliation preparation | Yes | AI prepares and matches | A person finishes and signs off |
| Accruals and adjusting entries | Proposals only | From an approved schedule | Senior reviewer approves |
| Ambiguous treatment, materiality, cut-off | No | The AI raises a question | Senior reviewer |
| Tax, payroll and advice to clients | No | Not in scope | Qualified person |
| Month-end close sign-off | No | AI gathers evidence | Owner or admin |
How do you write the firm's procedures and each client's rules?
An AI agent follows what is written. Write the rules in two layers.
Firm layer, written once for every client:
- Operating procedures: how the firm codes the bank feed, what evidence every entry needs, and when the AI must stop and ask.
- Accounting policies: how the firm treats prepaid expenses, accruals, owner draws and personal expenses.
- Capitalisation policy: the threshold, currency, asset categories, method and useful life.
Client layer, written per client:
| Rule type | Example (hypothetical) |
|---|---|
| Expense rule | "Payments to Northside Fuel go to Vehicle Fuel, class Field." |
| Income rule | "Deposits from the booking platform are sales; its fees go to Merchant Fees." |
| Capitalisation threshold | "Capitalise equipment over $5,000 for this client, not the firm's $2,500." |
| Close procedure | "Reconcile the operating account and the card by business day 5. Accrue rent if the bill is missing." |
Where a client rule differs from the firm rule, the client's rule wins. Write each rule as one testable sentence, using account names exactly as they appear in the chart of accounts. Add the exception path: "If a transaction matches no rule, raise a question instead of guessing."
Before any work, test the rules themselves. The prompts use Lopez Plumbing, a fictional client:
Read only. Read Lopez Plumbing's AI Brain and skills. List any client rule that conflicts with a firm policy, and any rule that names an account missing from Lopez Plumbing's chart of accounts. Do not change any records.
How should you set access for each person and client?
Give each person the least access their work needs, for each client separately.
| Role | Typical access | Signs off |
|---|---|---|
| Owner or admin | Every client | Month-end close |
| Senior bookkeeper or reviewer | Assigned clients, read and write | Tasks where they are the reviewer |
| Bookkeeper | Assigned clients, read only at first, then read and write | Nothing on their own work |
| New staff or contractor | Assigned clients, read only | Nothing |
| An agent the firm builds | One client, through a key limited to that client | Nothing |
Review access when someone changes role or a client leaves. Before granting write access, read the AI write access safety checklist.
Who reviews what?
Name three roles for every piece of work: the assignee who does it (a person or the AI agent), the reviewer who checks it, and the person who signs off.
| Work item | Assignee | Reviewer | Sign-off |
|---|---|---|---|
| Weekly bank coding | AI agent proposes | Bookkeeper | Bookkeeper approves before recording |
| Questions the AI raises | Bookkeeper answers | Senior, for anything material | Not needed |
| Bank reconciliation | AI agent prepares | Bookkeeper | A person finishes it in the ledger |
| Adjusting entries | AI agent proposes | Senior reviewer | Senior approves each entry |
| Month-end close | AI agent runs the checklist | Senior reviewer | Owner or admin |
A useful rule: the person who asked the AI to do a piece of work should not be the only reviewer of anything material.
What does a weekly and month-end rhythm look like?
The AI acts when someone asks it to, so put the asking on a calendar. The days below are an example.
Each week, per client:
- Monday: ask the AI to read the client's AI Brain and skills, review new bank transactions, propose coding and raise questions for anything the rules do not cover.
- Tuesday to Wednesday: a person answers the questions and approves or corrects the proposals. The AI records only the approved items.
- Friday: run a duplicate check on the week's bills, review the receivables aging and have the AI draft follow-up emails for a person to send.
Each month-end:
- Business days 1 to 2: the AI prepares bank and card reconciliations and lists exceptions.
- Business days 2 to 3: the AI proposes accruals and adjusting entries from approved schedules. A senior reviewer approves or rejects each one.
- Business days 3 to 4: the AI pulls the reports and explains variances.
- Business day 5: the reviewer checks the evidence, and an owner or admin signs off.
Prepare for approval. Read Lopez Plumbing's AI Brain and skills. For October 2026, prepare the operating account reconciliation from the statement I attached, list every unmatched line, and propose an entry for each bank fee missing from the books. Do not record anything until I approve each entry.
The month-end close automation guide covers the close in more detail.
How should AI help with client communication?
The AI drafts and a person sends: receipt requests, variance explanations, year-end questions and anything about money owed.
Before sending, check the client name, every number against the ledger, and that no internal notes or another client's details slipped in.
Read only. For Lopez Plumbing, list October 2026 expenses over $75 with no attachment. Draft one short email to the office manager asking for those receipts, with date, payee and amount for each. Do not send anything and do not change any records.
What should you settle with clients before the AI touches their books?
Settle these in writing before the AI reads a client's ledger. This is not legal advice: check your professional rules, engagement letters and AI provider's terms.
| Question | Why it matters |
|---|---|
| Which AI provider and plan will the firm use? | Data handling differs by provider and by plan |
| Does that plan use conversations for model training, and is that setting off? | Ledger data returned by a tool appears in the conversation |
| How long does the provider keep conversations? | Retention is set by the provider or the workspace admin |
| May the AI read only, or also record entries? | Sets the access you grant for that client |
| Who at the firm reviews the AI's work? | The client knows who is accountable |
| How is access removed if the engagement ends? | You can show it was revoked |
Providers document this per plan. As of November 2026, OpenAI's data controls page says data sharing for model training is on by default for personal ChatGPT Free, Plus and Pro workspaces, and its enterprise privacy page says ChatGPT Business and Enterprise data is not used for training by default. Anthropic's model training article covers its consumer plans, with commercial plans documented separately. Check the page for the exact plan your staff use.
How do you measure whether it is working?
Count what reviewers find, per client and per month, and compare it with the same client's months before the AI.
| Measure | How to count it | A bad trend means |
|---|---|---|
| Review notes | Notes a reviewer raises on AI-prepared work | Rules are missing or unclear |
| Corrections after recording | Entries reclassified or voided after the AI recorded them | Writes started too early |
| Items sent back | Proposals the reviewer rejected or changed | The AI lacks evidence or rules |
| Open questions | Questions still open at month-end | People are not answering in time |
| Reconciliation exceptions | Unmatched lines at sign-off | Source documents arrive late |
A rising number of corrections is a signal to move that client back a phase.
How do you roll it out in phases?
- Phase 1, read-only reports, one client. Pick a client with clean books and a reviewer who knows them. Ask for reports on one closed month and compare every total with the ledger.
- Phase 2, proposals. Let the AI propose bank coding and draft bills, with nothing recorded. Track items sent back.
- Phase 3, reviewed writes. Allow one narrow class of work, such as recurring software bills from known vendors, recorded only after a person approves the exact fields.
- Phase 4, the next client. Reuse the firm rules, write the new client's rules, and start that client at Phase 1.
Move to the next phase only after a full month-end with no corrections after recording.
What can go wrong, and how do you review the AI's work?
- Wrong client. Start every session by asking which company is active.
- Wrong period or basis. Require the company, dates and basis in every answer.
- Duplicates. Ask for a duplicate check before every recording.
- Cut-off. Check document dates against the close period.
- Missing source documents. Require a receipt or invoice reference for every entry.
- A summary that does not match the report. Open the same report in the ledger and compare totals line by line.
- Numbers the AI invented. Ask for transaction identifiers and open them in the ledger.
A chat reply that says "done" is not evidence. The ledger record is.
Where DeepLedger fits
DeepLedger connects QuickBooks Online and Xero to Claude, ChatGPT and other AI apps at one address, https://mcp.deepledger.ai/mcp.
- One connection, every client. One connection reaches every company you can access, and the AI switches company by name. Each write must name its company, or the server refuses it and nothing is written. See one AI connection for every QuickBooks company, and the Xero MCP guide for Xero.
- AI Brain. Firm procedures, accounting policies and the capitalisation policy are written once; each company adds its own rules, and the client's rule wins. The AI agent is told to read a company's AI Brain and skills before it records anything.
- Skills. Each company gets its own copy of skills, such as month-end close and bank reconciliation, to adapt.
- Tasks. Every task has an assignee, a person or the AI agent. With a reviewer set, only that reviewer can sign it off. What the AI isn't sure about becomes a question for a person. Recurring tasks are created on their schedule, and tasks for the AI agent wait until someone asks it to work them.
- Access. Owners and admins see every company. Everyone else sees only assigned companies, with read-only or read and write access set for each person and company.
- Confirmation for what cannot be undone. In QuickBooks Online, voids, deletes and customer emails wait for your confirmation.
- Audit history. Every tool call the AI agent makes, read or write, is recorded with its result in each company. Owners and admins can export it as a CSV, so a reviewer can trace what the AI read and wrote.
DeepLedger works when you ask. The AI never changes the books on its own, and a person signs off every close.
Frequently asked questions
How do I start using AI agents for bookkeeping services?
Start with one client and read-only work: reports, variance questions and duplicate checks. Write the firm's procedures and that client's rules first, then move to proposals a person approves, and only then to reviewed writes for one narrow type of transaction.
Which bookkeeping work should stay with a person?
Judgment and sign-off stay with a person. That covers ambiguous accounting treatment, materiality and cut-off decisions, tax and payroll conclusions, anything sent to a client, approval of adjusting entries, reconciliation sign-off and the month-end close sign-off.
Can the AI agent send emails to clients?
Only sales documents, and only after you confirm. Through DeepLedger, the AI can email a QuickBooks invoice, estimate, sales receipt, refund receipt or credit memo, or a Xero invoice, to the customer once you confirm that exact send. For every other client email, use the AI to draft it and have a person check the facts, the recipient and the tone before sending it from the firm's own email.
How do I know if AI agents are working in my practice?
Count what reviewers find. Track review notes, corrections found after recording, items sent back to the AI and questions left open, for each client and each month, and compare them with the same client's months before the AI. Do not rely on a feeling of speed.
Do I need to tell clients that the firm uses AI?
It depends on your professional rules, your engagement letters and your AI provider's terms, so check all three. Settle it in writing before the AI reads a client's books: which provider, which plan and data settings, what the AI may do, and who reviews its work.
Related guides
- DeepLedger for accounting firms
- AI bookkeeping
- Month-end close automation with AI
- AI write access to QuickBooks: controls to check
- One AI connection for every QuickBooks company
Start with one client: see the supported AI apps on integrations.