AI for Bookkeepers in 2026: Bank Feeds, Client Questions and the Close

AI can now code most of a client's bank feed, match receipts to card charges, send clients their uncategorized transactions as questions and flag lines that look wrong. It can't tell whether a $1,200 Zelle payment was rent or a loan to the owner's brother, and it can't make a client answer you. The coding gets faster; the questions still need a person.

Most bookkeepers didn't get into the work to spend the first week of every month asking for the same statements. Yet that's where the hours go: the card that stopped syncing, the client who replies "not sure, ask my husband", the W-9 nobody collected. Below is that work piece by piece, the software that handles each part, where it goes wrong, and what to check before a client's books touch any of it.

Where a bookkeeper's month actually goes

Posting and reconciling are only part of the month; a lot of it is waiting on clients. The Labor Department's O*NET entry for bookkeeping, accounting and auditing clerks is full of recording, coding and checking: reconciling bank transactions, coding documents, matching order forms with invoices, preparing trial balances, processing payroll and submitting tax forms. In a practice with a few dozen monthly clients, the calendar looks more like this:

• The first week: statements that didn't download, feeds that disconnected, receipts still sitting in a client's camera roll.
• The uncategorized pile: every line parked in Ask My Accountant or Uncategorized Expense until the client explains it.
• Reconciliations: bank, card, loan and merchant accounts, plus the payroll clearing account that never quite zeroes.
• Closing the month: reviewing the P&L for odd swings, posting the recurring entries, setting the closing date so nobody edits March in June.
• Deadlines that aren't monthly: quarterly payroll returns, sales tax on each state's schedule, and 1099s in January.
• The portal back-and-forth: a question asked in the portal, answered by email, with the receipt attached to a text message.

The federal Occupational Outlook Handbook expects jobs in this occupation to fall 6% from 2025 to 2035, and says software has already automated many of its tasks. The data entry is going. The client questions, the judgment calls and the review are what clients keep paying for.

Bookkeeping tasks AI can take on in 2026

For a bookkeeping practice, most of the useful AI sits inside the ledger (QuickBooks Online or Xero) or in apps that connect to it. A general chat assistant can help write a client email or explain a transaction description, but it doesn't see the books.

Coding the bank feed

Transaction coding is the part of bookkeeping AI handles best, because a client's vendors repeat every month. Intuit's QuickBooks AI page lists expense categorization from its Simple Start plan, "Reconciles your books" and anomaly detection on Plus, and Payments AI that proposes reminder messages for overdue invoices. Booke AI works on top of QuickBooks Online and Xero: it categorizes the feed daily, matches lines to invoices, bills and receipts, lets you set rules such as requiring a receipt above an amount you choose, and leaves low-confidence items for you with a suggested treatment. Booke lists $129 per business per month for its AI Bookkeeper, with firm pricing on request.

Turn on one client first and read every line the AI posted for a full month before you trust it on the rest.

The Ask My Accountant pile

Uncategorized transactions are where bookkeeping AI saves the most back-and-forth, because the slow part is the client, not the coding. Uncat sends uncategorized bank lines to the client to explain, collects the receipt and writes the answer back to QuickBooks Online, QuickBooks Desktop or Xero. It lists $9 per client per month, with unlimited users, transactions and requests. Booke's page lists a client query tool for the same job.

Double (formerly Keeper) puts client messages on the close page itself, with a branded client portal and app, so a question about a transaction sits next to the account it affects instead of in a separate inbox.

Receipts and statements

Receipt capture works when clients actually send receipts. Hubdoc takes a photo, a forwarded email or an upload, reads the supplier, amount, invoice number and due date, and creates the transaction in Xero or QuickBooks Online with the document attached. Statements are harder: a disconnected feed or a credit union with no download still means asking the client for a PDF. Saving those attachments straight from Gmail into each client's folder is covered in this guide to saving Gmail attachments to Google Drive.

Reconciliation and the month-end close

Close software now does the first pass on reconciliation so you start at review. Double says its Autoclose pulls statements, codes bank feeds and reconciles accounts, with AI that catches coding errors, builds schedules, runs flux and drafts complex journal entries and accruals. On the QuickBooks side, Intuit describes its Finance AI as bringing "a more efficient month-end close" but doesn't say which plan includes that part. A flux note that says revenue fell is a prompt to ask the client why; it isn't the answer.

Payroll, 1099 and sales tax deadlines

Deadline work is mostly rules and reminders, and AI adds the most where a filing could be wrong rather than late. The dates are fixed by the IRS: Form 941 is due by the last day of the month after each quarter (April 30, July 31, October 31 and January 31), and Forms W-2 go to employees and the Social Security Administration by January 31, per the IRS's employment tax due dates. Form 1099-NEC goes to contractors and the IRS by January 31 too.

One change for January 2027 filings: the IRS's 1099-NEC instructions, which also set that January 31 date, say the reporting threshold rises to $2,000 for tax years beginning after 2025, so nonemployee compensation paid in 2026 is reported at $2,000 or more. Check that any vendor list or AI filter flagging 1099 contractors uses the new number, and collect W-9s at the first payment rather than in the last week of December. For sales tax, QuickBooks lists Sales Tax AI on its Plus plan to flag possible filing issues and suggest fixes.

Where AI gets bookkeeping wrong

AI coding fails on the transactions that need knowledge of the client rather than the vendor. A model sees "AMZN Mktp" and picks Office Supplies; you know this client buys inventory there. Watch these in particular:

• Owner money. Personal charges on the business card, owner draws and capital put back in look like ordinary expenses and income to software.
• Transfers. A move between the client's own checking and savings coded as income is the classic doubled-revenue P&L.
• Peer-to-peer payments. A Zelle, Venmo or Cash App line with a first name tells the AI nothing about whether it was a contractor, a refund or a family loan.
• Loan payments. One debit holds principal and interest, and the split comes from the lender's statement, not the bank line.
• Duplicates. A receipt captured by an app plus the same charge from the feed becomes two expenses if nobody matches them.
• Assets. Whether a laptop is expensed or added to fixed assets follows the client's capitalization policy, which no feed shows.

A closing date with a password in QuickBooks, or a lock date in Xero, matters more once AI is posting entries: it keeps an automated fix from quietly changing a month the client already filed against.

Client data: what to check before connecting a tool

Bookkeepers hold bank logins, statements with full account numbers, payroll records and Social Security numbers, so every AI tool that connects to a client's books is another place that data lives. Three things decide what you can connect:

• If you also prepare returns, the IRS's Section 7216 rules limit how tax return preparers use and disclose return information without the taxpayer's consent.
• The FTC Safeguards Rule names tax preparation firms among the businesses that need a written information security program, multi-factor authentication and oversight of service providers, per the FTC's guide. Whether it covers bookkeeping alone depends on your services, so ask counsel if you're unsure.
• Your client agreement is the place to say which software touches the books, including AI, and who can see it.

Before connecting a new app, read where it stores data, whether it trains on client data, and whether it keeps each client's data separate. Booke, for example, says it uses a separate AI model for each client and doesn't share data between them.

Merlin and the client inbox

Merlin helps with the client email that surrounds the books, not the books themselves. Installed on a Mac, with access to the Gmail account and Google Calendar you run the practice from, Merlin keeps a running record of each client: who you owe a reply, who has gone quiet on a request, and what you said you'd send.

• The clients who went quiet. Merlin's People view shows who's waiting on you and who stopped answering, so the client sitting on your March statement request doesn't slip to April.
• Reminders you approve. Request the second nudge for a missing statement or W-9 in Merlin's chat; you see the draft first, can change any line, and it leaves your outbox only after you say yes.
• Promises tracked. Merlin keeps a to-do list of follow-ups, such as the reconciled P&L you told a client they'd have by Friday.
• Prepared for the review call. Meeting Brief prepares you before each meeting on your calendar, and Merlin can record the call on your Mac. Transcripts come with the paid plans, and clients should know you're recording.

Merlin isn't bookkeeping software. It doesn't code transactions, reconcile accounts, run a client portal or send invoices, so the ledger tools above still do that work. Its access to client email deserves the same vetting as any other app's.

A first week of AI for a bookkeeping practice

Begin with the AI already in your clients' ledger, on one clean client, before buying anything.

1. Pick one client with tidy books and turn on AI categorization in QuickBooks Online or a connected app such as Booke AI.
2. Review every posted line for a month and write down the mistakes; they tell you which rules to add.
3. Move uncategorized questions out of email into a tool the client can answer from their phone, such as Uncat or your portal.
4. Update your client agreement to name the software, including AI, that will touch the books.
5. Fix the January list now: confirm which vendors cross the $2,000 1099-NEC threshold and which W-9s are missing.

For deadline reminders on your own calendar, these Google Calendar automations can put each 941 and 1099 date in front of you with time to spare.