AI bookkeeping that actually books entries
Most paid "AI bookkeeping" in 2026 is software plus a human bookkeeper on a monthly plan. ERPClaw is the free, open source alternative: an AI bookkeeping agent imports your bank statements and Stripe or Shopify activity, clears the routine lines with your rules, posts balanced entries, and shows how each account reconciles, all on books you host yourself. What it cannot match comes to you.
By Nikhil Jathar, Co-founder, ERPClaw · Published · Updated · Vendor prices checked
What is AI bookkeeping?
AI bookkeeping is software where an AI agent does the daily transaction work: it brings in bank and payment activity, matches it to invoices and bills, posts the journal entries, reconciles the accounts, and produces the reports. A person reviews only what the agent could not match. The difference between products is who does the rest: some add a paid human bookkeeper, some sit on top of QuickBooks or Xero, and ERPClaw runs the agent on an open source ledger you host yourself.
- ·Free. $0 for the software. No seat, entity, or transaction fees. You pay for your own hosting and AI model usage.
- ·Open source. Python under GPL v3. Read the code, fork it, fix it.
- ·Self-hosted. Your books live on your hardware. SQLite by default, with PostgreSQL support.
- ·Exceptions come to you. Anything the agent cannot match waits in a review list instead of being posted as a guess.
- ·Immutable GL. Cancel means reverse, never edit. Audit-trail integrity is structural, not a setting.
Ask it like you'd ask a person
Bookkeeping requests that work today, in plain words, on statement files your bank already exports.
"Import my January bank statement and match it to open invoices"
Hand it the statement file (OFX, CAMT.053, MT940, or BAI2). Lines that match themselves are done; the rest queue for your review.
"What is still unmatched on that statement?"
The short list a human should actually look at, with one-line fixes: match it to an open invoice, map it to an account, or leave it.
"How does the checking account reconcile as of month end?"
The as-of-date picture from your imported statements: matched, unmatched, and where the account stands.
"The customer paid $50 short; write it off in the same payment"
Payment and write-off posted together, balanced, in one request.
How an AI bookkeeping agent works, step by step
The AI understands the request and picks the action. The accounting itself is done by ordinary Python code with fixed rules, and ERPClaw makes no AI calls of its own, so the model never does arithmetic on your ledger. This is the cycle in ERPClaw.
- 1
Bring the transactions in
Hand the agent the statement file your bank already exports (OFX, CAMT.053, MT940, or BAI2). The whole file is checked before anything is saved, and importing the same statement again does not create duplicates. Stripe and Shopify activity comes in through their integrations, so a payout arrives with the charges, fees, and refunds behind it instead of as one net deposit.
- 2
Clear the routine lines with your rules
Write a matching rule once (charges from your card processor go to merchant fees, transfers from savings are ignored) and it clears those lines on every import. Rules only mark lines as matched; nothing is guessed into an account to make the list shorter.
- 3
Work the unmatched list with the agent
Whatever no rule covers waits in an unmatched list. Ask the agent about it and it proposes a fix for each line: match it to an open invoice, map it to an account, or leave it. You decide.
- 4
Post through accounting rules, not guesses
Invoices, payments, and journal entries start as drafts and post when submitted. Every posting passes the same validation first: debits must equal credits, the accounts must be real posting accounts, and nothing lands in a closed or frozen period. The same input produces the same books, every time. A wrong entry is fixed with a visible reversal, never a silent edit.
- 5
Reconcile, close, and report
Ask how an account stands as of a date and you get the ledger balance next to the statement's closing balance, with the matched and unmatched lines. Close the period, freeze it, and pull the P&L, balance sheet, trial balance, and aging reports from the same ledger.
What AI bookkeeping does in ERPClaw
Bank statement import and matching
Import the statement file your bank exports (OFX, CAMT.053, MT940, BAI2). Re-imports are de-duplicated, your matching rules clear the routine lines, and everything else sits in an unmatched list you work through with the agent.
Recurring entries and depreciation
Recurring journal templates from daily to annual, recurring invoices, and recurring bills. Fixed-asset depreciation schedules (straight-line, written-down value, or double-declining). Run them on request or on a schedule you set; recurring journals can wait as drafts for your sign-off.
Reconciliation, bank and Stripe
An as-of-date view for each bank account: ledger balance, statement closing balance, matched and unmatched lines. Each Stripe payout is checked against the charges, refunds, and fees inside it, and each charge is tied back to its customer.
Period close and locked books
Check a period before you close it, close the fiscal year to retained earnings, and freeze the books up to a date. Postings into a closed year or a frozen period are refused, so a number you already reported stays put.
ASC 606 revenue recognition
Contracts with performance obligations satisfied at a point in time or over time, variable consideration, and monthly deferred revenue schedules that post from deferred revenue to revenue. Works for Stripe subscriptions and for contracts you enter yourself.
Reports and trial balance
P&L, balance sheet, cash flow, trial balance, general ledger detail, AR aging, AP aging, sales tax summary, comparative P&L, and budget versus actual. 1099 data for the vendors you flag. Ask for a report in plain words; it comes back as structured data you can share with your CPA.
Several companies, one install
Run several companies in one database, each with its own general ledger. An intercompany journal entry posts to both companies in one step, and an intercompany invoice creates the matching bill on the other side. Group companies for consolidation: each period the intercompany elimination entries are generated for you, and you record currency translation for foreign subsidiaries as a deliberate entry. No separate instance, no per-entity fee.
Hash-chained general ledger
Every general ledger row carries a SHA-256 hash linked to the row before it, per company. One integrity check confirms the chain is unbroken and that debits still equal credits, so a posting changed behind the system's back shows up.
Where AI bookkeeping goes wrong, and how to catch it
These mistakes show up with any AI bookkeeping tool, ours included, because they come from how bank data looks rather than from any one model. Each has a quick check you can run in a few minutes.
Transfers booked as income
Money moved from savings to checking looks like a deposit. A tool that only reads one account at a time can book it as revenue and overstate your income.
Check: Ask for every transfer between your own accounts in a month and confirm each one hit two balance-sheet accounts, not the P&L.
Net payouts booked as sales
A Stripe or Shopify payout is sales minus fees minus refunds. Booking the net deposit as revenue understates both sales and fees, and your sales tax base with them.
Check: Pick one payout and trace it back to the gross charges, fees, and refunds it contains.
The same transaction twice
Importing a statement file on top of a live feed, or re-importing a corrected file, can create duplicates that look perfectly normal on their own.
Check: Reconcile the account to the closing balance on the bank statement. If the books are higher than the bank, look for doubles first.
Confident answers with no ledger behind them
A chat assistant that summarizes instead of querying the books can give you a clean-sounding number that no entry supports.
Check: Ask where the number came from. A trustworthy tool points to the entries or the report it read.
History that changes quietly
If last quarter's numbers can be edited in place, a later correction can change a period you already filed or sent to your lender, with no trace of why.
Check: Try to change a posted entry. You want a reversal and a new entry, both visible, not an overwrite.
What to check before you trust an AI bookkeeper
Categorizing a transaction is not the same as supporting it. The IRS puts the burden on the business: "You must be able to prove certain elements of expenses to deduct them" (IRS, Recordkeeping). Whatever tool you pick, ask these before it touches a live month.
- ·Can I see every entry it posted, and why? Each posting should trace back to a statement line, invoice, or bill.
- ·What happens when it is unsure? You want a review queue, not a best guess booked to "uncategorized" or the nearest account.
- ·How is a mistake corrected? A reversal plus a new entry keeps history honest. An in-place edit rewrites a period you may already have filed.
- ·Does the account tie to the bank statement? Reconcile one month to the statement's closing balance before you rely on anything else.
- ·Where do my books live, and who else sees the data? Know which company stores the ledger and which AI provider reads your transactions.
- ·Can I leave with everything? A trial balance and GL detail export your CPA can open, plus the underlying data, not just PDFs.
- ·What does the price scale with? Expenses, transactions, entities, properties, or licenses. The comparison below shows how each vendor counts.
For how ERPClaw answers the audit questions in detail, see audit-ready books and security.
AI bookkeeping software compared, with 2026 prices
Published prices from each vendor's own pricing page, checked September 27, 2026. They count different things (monthly expenses, businesses, properties, firm licenses), so compare the unit as well as the number. Prices change; follow the link before you rely on one.
| Vendor | Who does the work | Where the books live | License | Published price |
|---|---|---|---|---|
| Pilot | Essentials: AI only, no bookkeeper. Core: adds a US-based bookkeeper | Vendor cloud | Proprietary | Essentials $99/month (up to $100,000 in monthly expenses). Core starts at $299/month billed annually and rises with monthly expensesSource: Pilot pricing, checked 2026-09-27 |
| Zeni | AI bookkeeping plus a dedicated finance team | Vendor cloud | Proprietary | Starter from $549/month ($494/month billed annually). Growth from $799/month ($719/month billed annually)Source: Zeni pricing, checked 2026-09-27 |
| Bookkeeper360 | Dedicated accounting team on their software | Vendor cloud | Proprietary | Monthly plan from $399/month, weekly plan from $599/month, onboarding from $1,000 per projectSource: Bookkeeper360 pricing, checked 2026-09-27 |
| Booke AI | AI bookkeeper inside QuickBooks Online or Xero; you review the exceptions | Your QuickBooks Online or Xero account | Proprietary | $129 per business per month, on top of your QuickBooks Online or Xero subscriptionSource: Booke AI pricing, checked 2026-09-27 |
| Botkeeper | AI platform that accounting firms run for their clients | Vendor cloud | Proprietary | $53 to $149 per entity license per month, sold to accounting firms (lower per license as the license count rises)Source: Botkeeper pricing, checked 2026-09-27 |
| ERPClaw | AI agent on your own ledger; what it cannot match comes to you | Your server (SQLite or PostgreSQL) | Free & Open Source (GPL v3) | $0 for the software. You pay for your own hosting and your AI model usage |
Head-to-head pages: ERPClaw vs QuickBooks, vs Xero, vs Puzzle, vs Synder, and the wider AI finance stack.
Who runs AI bookkeeping on ERPClaw
Three profiles where a free, self-hosted agent beats a monthly bookkeeping plan today. Each one has a real mix of Stripe payouts, Shopify orders, and recurring entries, the work that the plans above charge monthly to handle.
Stripe-only SaaS founders
Stripe is your only payment processor. Subscriptions, one-off charges, refunds, disputes, and ASC 606 deferred revenue all need to land in the books cleanly. The Stripe integration (live on the Stripe Marketplace) backfills history, posts charges with the fee split out, and checks each payout against the transactions inside it. The Stripe docs are at /docs/stripe/.
Shopify store operators
Shopify orders, payouts net of fees, returns, and COGS all need to flow into the GL with the right cost basis. The Shopify integration pulls orders, payouts, refunds, and disputes directly. Inventory and bookkeeping live in the same database, so with cost tracking on and products mapped to items, each order posts its cost of goods sold from your inventory cost. Setup is in the Shopify docs, and the connected inventory deep dive is at /ai-inventory/.
QuickBooks switchers
Your books are stuck in QuickBooks with manual Stripe reconciliation eating hours every month. The AI-native upgrade is at AI for QuickBooks with the full head-to-head at ERPClaw vs QuickBooks and the weekend migration playbook at /migrate/from-quickbooks/.
For the broader category framing see AI accounting and the architecture argument at open-source AI accounting. Accountants and bookkeepers running several clients should start at ERPClaw for accountants. For mid-market consolidation comparisons see vs NetSuite, vs Sage Intacct, and vs Rillet.
Frequently asked questions about AI bookkeeping
What is AI bookkeeping?
AI bookkeeping is software where an AI agent does the daily transaction work: it brings in bank and payment activity, matches and categorizes it, posts the journal entries, reconciles accounts, and produces reports. A person reviews the exceptions. Some vendors pair the AI with a human bookkeeper you pay for monthly (Pilot Core, Zeni, Bookkeeper360); others run the AI on top of your existing QuickBooks or Xero file (Booke AI). ERPClaw runs the agent on an open source ledger you host yourself.
Is there free AI bookkeeping software?
Yes. ERPClaw is free and open source (GPL v3): the software costs $0 with no seat, entity, or transaction limits. What you still pay for is the server it runs on and the AI model the agent uses. Most other AI bookkeeping products are paid subscriptions; see the comparison above for prices checked September 27, 2026.
Can AI really do my bookkeeping without a human?
It can do the volume work: bringing in transactions, matching them to invoices and bills, posting recurring entries, reconciling accounts, and producing reports. It should not be left alone on judgment calls such as unusual transactions, owner draws versus expenses, or anything headed for an audit or a tax return. The workable setup is AI for the volume and a person for the exceptions. ERPClaw queues anything it cannot match for your review instead of posting a guess.
How much does AI bookkeeping cost?
As of September 27, 2026, published prices run from $99 a month for an AI-only tier (Pilot Essentials) to $299 to $799 a month once a human bookkeeper or finance team is included (Pilot Core, Bookkeeper360, Zeni). An AI layer on top of QuickBooks Online or Xero, such as Booke AI, is $129 per business per month plus your existing subscription. ERPClaw's software is $0; you pay only for hosting and AI model usage. Each figure links to the vendor's own pricing page in the comparison above.
What about my CPA, do I still need them?
Yes. AI bookkeeping handles daily volume, not tax strategy or audit defense. Your CPA still owns year-end review, tax planning, and entity-level decisions. What changes is what they receive: reconciled books with a full trail, instead of a shoebox of statements to clean up first. If you need bookkeeping or tax help from a person, our accounting partners can help; contact support and we will connect you.
Is open-source AI bookkeeping safe for my finance data?
Self-hosted ERPClaw keeps the books on hardware you control, in a SQLite file or a PostgreSQL database. The accounting runs as ordinary code on that machine; ERPClaw itself makes no calls to an AI service. The AI model you connect it to does see what it works with (the requests you type and the results that come back, such as names and amounts), so choose that model the way you would choose any vendor that reads your financial data. With a SaaS bookkeeping service, the ledger itself lives on the vendor's servers.
What if the AI miscategorizes a transaction?
ERPClaw's general ledger has no edit path for a posted amount. To fix a miscategorization, the entry is cancelled, which posts a mirror reversal, and the correct entry is posted in its place; both stay visible in the trail. Then fix the cause: add a matching rule so the next statement line from that payee lands in the right account.
Will my CPA accept AI-generated books for tax filing?
Your CPA reviews the books, not the tool that produced them, so the question is whether the books hold up. In ERPClaw every entry is balanced and validated when it is posted, corrections show as reversals, closed periods are locked, and the general ledger is hash-chained so an integrity check catches a changed posting. Before you switch, pull a trial balance and general ledger detail for a closed month, send them to your CPA, and ask directly.
Related: read the broader AI accounting story, the AI-native ERP framework, the QuickBooks comparison, the QuickBooks migration playbook, the nonprofit view if you keep restricted funds, or the double-entry bookkeeping for developers deep dive.