AI for Accountants in 2026: Close, Clients and Confidentiality

AI can now match bank feeds, code receipts, draft flux commentary, summarize long client email threads, pull audit evidence into Excel and flag unusual journal entries. It can't sign a review report, exercise professional skepticism or decide what a client's numbers mean. Use it for the preparation, keep the judgment, and check where client data goes first.

For most firms the hard part of the month isn't the accounting. It's the missing bank statement, the client who answers a three-question email with one answer, the flux explanation nobody wants to write and the engagement letter that never got updated when the scope grew. Below: that work task by task, the software that handles each piece, and the confidentiality rule that decides which of it you may use.

Where the hours go in an accounting practice

The federal government's Occupational Outlook Handbook entry (published by BLS, the Labor Department's statistics agency) describes public accountants as handling a broad range of accounting, auditing, tax and consulting tasks, and notes that accountants may already use AI and robotic process automation to work faster. O*NET, the Labor Department's occupation database, gets more granular in its accountants and auditors entry: reviewing accounts for discrepancies and reconciling differences, preparing adjusting journal entries, reporting audit findings and conferring with company officials about financial and regulatory matters.

In a firm with a book of monthly clients, that turns into a recurring list:

• Bank and card reconciliations for every client, including the uncategorized lines nobody explained.
• Receipts and bills to capture and code before the books close.
• Month-end close: accruals, adjusting entries, the checklist and the flux explanations.
• Chasing clients: statements, payroll reports, loan schedules, the answer to "what was this $4,000 transfer?"
• Engagement letters and scope changes, and billing for the extra work.
• Audit, review and compilation work: tying out schedules, documenting evidence, clearing review notes.
• Advisory: cash forecasts, the quarterly check-in and the follow-up email afterwards.

Accounting tasks AI handles today, and the software behind them

For an accounting firm, useful AI lives in four places: the general ledger (QuickBooks Online, Xero), close software, practice management software and audit tools. A general chat assistant can help with writing, as long as no confidential client data goes into it without the safeguards covered below.

Bank feeds and coding

Bank reconciliation is where AI in the general ledger is furthest along. Xero's auto bank reconciliation page says JAX, its AI, only reconciles a bank line when it has high confidence; otherwise it suggests a match or category and leaves the line for you. Each decision is labeled Match, Rule, Memory (how you coded similar items before) or Prediction (how other Xero users coded them), and plan availability varies by country. Intuit's QuickBooks Online Accountant page says Intuit Intelligence matches client transactions and flags discrepancies.

Give Prediction matches a second look: how other businesses code a payment says nothing about this client's chart of accounts.

For the paper side, Dext reads receipts and bills that clients send by email, app, WhatsApp or upload, extracts the supplier, line items and due date, and publishes to QuickBooks Online, Xero, Sage and other ledgers. Its AI Assist proposes automations for supplier, category and split decisions, which you approve.

Month-end close and flux analysis

Close software now drafts the commentary that used to take the last day of the close. Numeric describes automatic balance sheet reconciliation where the team reviews only the exceptions, transaction monitors that flag anomalies before close, a shared close checklist, and "agentic flux" that pulls detail from payroll and CRM systems into variance explanations.

For firms on QuickBooks, Intuit lists an agentic month-end close, handling month-end prep and delivering books for approval, as coming soon in QuickBooks Online Accountant. The Core plan is free; Intuit says its Accelerate plan is free through January 2027 and then $149 a month. A draft flux explanation is a starting point. The reason revenue dropped 18% is something the client tells you, not something software infers.

Chasing clients for documents and answers

Document chasing is mostly a scheduling problem, and much of the software that solves it isn't AI at all. Financial Cents' client tasks request documents, e-signatures or answers through a portal and send automatic email and SMS reminders until the client completes them. That's a fixed reminder schedule, and for a missing bank statement it's often all you need.

AI earns its place on the reply side. Karbon's AI page lists summaries of long email threads, drafted and rewritten emails with tone and length controls, emails drafted from a task's progress, and a Client Summary that pulls emails, notes, work and billing into one brief. Karbon says its AI is included for all customers at no extra cost and that admins can turn it off firm-wide.

Proposals, engagement letters and pricing

An engagement letter that matches the work is your best protection when scope creeps, and it's also where a client can be told how their data will be handled. Ignition handles proposals and engagement letters for accounting and tax firms, collects payment details in the proposal, bills when scope changes and offers AI Price Insights for pricing. An AI price is one input; the state of the client's books is another.

Audit, review and assurance work

In assurance work, AI helps most with evidence: reading documents, tying them to workpapers and looking at every transaction instead of a sample. DataSnipper pulls evidence from the systems teams already use into Excel, extracts and cross-references data from documents, and runs workflows such as revenue testing or test of details. DataSnipper says it doesn't train on customer data. MindBridge says it analyzes 100% of journal entries with statistical models, business rules and unsupervised machine learning, and helps teams prioritize higher-risk items so testing goes where it matters.

Advisory and client insights

Intuit says QuickBooks Online Accountant surfaces AI insights at the client, portfolio and industry level, lets clients ask Intuit Intelligence for cash forecasts, and, on major decisions, prompts the client to contact their accountant and sends you the context. The AI starts the advisory conversation; you still have it.

Confidentiality: the AICPA rule that decides which AI you can use

Under the AICPA Code of Professional Conduct, a member in public practice can't disclose confidential client information without the client's specific consent (the Confidential Client Information Rule, 1.700.001). An AI tool that receives client data is a third-party service provider, and the Journal of Accountancy's summary of interpretation 1.700.040 gives two routes before sharing: a contract in which the provider agrees to keep the information confidential, with reasonable assurance that it has procedures to prevent unauthorized release, or the client's specific consent.

A second, 2022 Journal of Accountancy article adds that under interpretation 1.150.040 the firm should tell the client, preferably in writing, that a third party may be used on the engagement, before confidential information goes to it, and that this can be done with a provision in the engagement letter. In practice that means:

• Read the data terms of every AI tool before client files go into it: who stores the data, whether it trains on it, and whether you can sign a contract that covers confidentiality.
• Update your engagement letter template so clients know third-party software, including AI, may be used on their work.
• Keep a free consumer chatbot away from client data unless its terms meet the same bar as your other vendors.
• Check your state board's rules too. Your state board of accountancy may add its own requirements.

Tax return information carries separate federal consent rules (Internal Revenue Code Section 7216) on top of the AICPA Code, so check those before any return data goes into an AI tool.

What AI can't do in an accounting practice

AI can't take responsibility for a number. It doesn't know that the client's "consulting income" is really a shareholder loan, that the inventory count was done on the wrong day or that the owner's explanation for a variance changed between calls.

• Sign-off stays with the CPA. A compilation, review or audit report carries your name and license. No tool shares that.
• Review stays human. A PCAOB staff report from July 2024 found audit firms' use of generative AI focused mainly on administrative and research activities. As CFO.com reported, the report adds that human involvement remains essential to review the output.
• Skepticism stays human. A model that flags an anomalous journal entry can't interview the bookkeeper who posted it.

Where Merlin fits for accountants

Merlin covers the part of the job that happens in your inbox and calendar: the controller waiting on your answer about an accrual, the quarterly check-in you need to prepare for. Running on your Mac against the firm's Gmail and calendar, it keeps a record of each client and what you've committed to, and writes the emails you request.

Merlin isn't accounting software. It doesn't reconcile accounts, run a close, produce engagement letters or send invoices, and it doesn't replace practice management software or a client portal.

• Unanswered client mail. Merlin's People panel has a Needs you list that shows a client when a message they sent you personally has gone 2 to 30 days without an answer from you. It reads Gmail only, so a voicemail from the same client won't put them there.
• Promises in one place. The To-do panel gathers commitments Merlin spots in your mail, such as "I'll send the adjusted trial balance Monday", next to the items you type in yourself.
• Nothing leaves without you. When you request a reply to a client in chat, Merlin shows the draft for your edits and sends it only once you approve.
• Before client meetings. Meeting Brief prepares a sourced note roughly ten minutes ahead of each calendar event, with past discussions and open commitments.
• Recorded client calls. A client call held on your computer can be recorded in Merlin; speaker-labeled transcripts need a paid plan. Tell clients when you record.

Connecting a Gmail account full of client correspondence makes Merlin a third-party service provider like any other, so apply the same vendor checks and engagement letter language described above before you do.

Five first steps with AI for an accounting firm

Pick the task that eats the most staff hours across your client base, and switch on the AI inside your existing ledger and practice software before buying anything new. For most bookkeeping-heavy firms that's bank reconciliation.

1. Check your ledger's AI. Turn on JAX in Xero or the AI matching in QuickBooks Online Accountant for two clean clients, and review every auto-reconciled line for a month.
2. Fix the engagement letter. Add language that third-party software, including AI, may be used on the engagement.
3. Vet one tool properly. Read its data terms, confirm whether it trains on customer data and get confidentiality in writing.
4. Automate the chase. Put recurring document requests on an automatic reminder schedule.
5. Review every output. Treat AI-drafted flux notes, emails and workpaper extractions like a new staff accountant's work.

Firms that hold a lot of client calls can compare meeting recorders in this roundup of AI note takers, and Gmail labels can split client mail by engagement before any AI touches it.