Offer · Bank statements posted line by line, every entry checked, your accountant has the last word
KaizoBooks AI OS for Accounting
KaizoBooks is an AI agent that posts your bank statements: for every transaction it decides which account it belongs to. It reads the MT940 statement, recognises every transaction, matches it to invoices, proposes the entry with a reason, posts it to the buffer of your accounting system and checks the balance after every line. If anything does not add up, it withdraws the whole statement and starts again. Your accountant approves, and your financial data never leaves the company.
For whom
Accounting offices and finance teams in companies with several bank accounts and hundreds of statement lines a month, starting with Symfonia Finanse i Księgowość; especially those whose payments, tax balances and due diligence documents must stay confidential.
Runs on
AI Office Box Pro with a 24–32 GB GPU in your office; AI Office Box Start for a small number of statements. The demo runs on a laptop with a graphics card. Compare AI Box packages →
System diagram
Inputs, outputs and interfaces
What goes into the AI Office Box, what comes out, and through which interface. Everything inside the dashed line runs in your network: the language model, the rules and the audit log. There are no connections to cloud AI.
Your network: your data stays here
Inputs
Bank statements
MT940 statements from each of your bank accounts: incoming payments, supplier payments, taxes, ZUS, salaries, card payments and bank fees.
MT940 files from a folder or e-mail, or statements already imported into your systemInvoices and open items
Unsettled sales and purchase invoices, receivables and payables, contractors with their tax ID and bank account.
Read-only connection to your accounting systemYour rules and posting history
Chart of accounts, categories, and 6–12 months of past statements with their entries. From them we build the rules the agent starts with.
Export from your accounting system
AI Office Box Pro in your office
- 1Start check: statement number, opening balance and totals agree with the books
- 2Rules from your history, plus a local language model (Bielik or Qwen3) for the rest
- 3Recognises the transaction type and matches it to invoices and open items
- 4Proposes the entry with accounts, confidence and a reason
- 5Posts to the buffer, checks the balance after every line, withdraws and restarts on any error
Outputs
Entries in the buffer
Every statement line posted and verified, waiting in the buffer: nothing goes into the books without a person.
Your accounting systemQuestions to your accountant
Only the lines the agent is not sure about, with its proposal and reason. Your answer can become a rule.
Review panel in your LANStatement report and audit log
Lines posted automatically and by hand, corrections, checks passed, and every action with a reason and a screenshot.
Review panel, PDF export for the auditorAnswers about your books
Tax and settlement balances, payment comparisons and an auditor's pack, asked for in plain language.
Review panel, read-only
Every correction your accountant makes becomes a rule or an example, so the share of lines posted without a correction grows from month to month.
Interfaces
| Interface | Direction | Protocol / format | What passes |
|---|---|---|---|
| Bank statements | Bank → AI Box | MT940 files from a folder or mailbox, or statements already imported through your bank connection | Statement lines, opening and closing balances |
| Accounting system: writing | AI Box → accounting system | API or file import where available; otherwise the agent works in the program's screens, like an accountant, on its own user account | Entries into the buffer only, never straight into the books |
| Accounting system: reading | Accounting system → AI Box | Read-only database connection on agreed views | Chart of accounts, open invoices and items, balances for the checks and for questions |
| Review panel | You ↔ AI Box | HTTPS in your LAN, any browser | Proposals with reasons, approvals, corrections, reports, questions in plain language |
| Remote service | Our team ↔ AI Box | VPN, under the SLA | Updates, monitoring, help with new transaction types |
How it works
The journey of one bank statement
What your accountant does today, done by the agent in minutes instead of hours. The marked stages are checks: the statement moves on only when a check passes.
- 1
Fetch
Reads the MT940 statement of each account.
- 2Check
Start check
Next statement number, opening balance equal to the books, totals that add up.
- 3
Recognise
Incoming payment, supplier, tax, ZUS, salary, bank fee or transfer between your own accounts.
- 4
Match
To invoices and open items: amount, tax ID, bank account, invoice number in the title.
- 5
Propose
Accounts, category and a reason. Not sure? It stops and asks your accountant.
- 6Check
Post and verify
Into the buffer, then reads it back: amounts, accounts and the balance after the line.
- 7Check
Closing check
Closing balance equal to the bank's, every line posted, nothing skipped.
Any check fails, a wrong category, a broken connection: the whole statement is withdrawn from the buffer and processed again. After a few failed tries it goes to a person, with the reason.
Your accountant approves the statement, and it moves from the buffer into the books.
Trust
Why your accountant can rely on it
Your data stays in the company
The language model, the rules and the documents run on the AI Box in your office. Payments and balances are never sent to cloud AI.
The buffer, not the books
The agent writes only to the buffer, on its own user account. Moving entries into the books is always a person's decision.
A wrong category cannot be posted
The validator blocks a category that does not fit the direction of the payment, inactive accounts and entries that do not balance.
Restart instead of repair
When something goes wrong the agent does not patch the books: it withdraws the statement and starts again, so no half-posted statement is left behind.
Every decision explained
“Tax ID matches, amount equals invoice FV/2026/10/112, the title contains its number.” Each line shows why, and how sure the agent is.
Ready for the auditor
Every action of the agent and of people is logged with time, input, result and a screenshot, and exported in one click.
Autonomy
You decide how much the agent does on its own
The chief accountant sets the level separately for each transaction type, and raises it only when the numbers show it is safe.
| Level | What the agent does | When |
|---|---|---|
| 0 Proposal | Proposes every entry; a person approves each line | At the start of the pilot and for every new transaction type |
| 1 Auto when sure | Posts the lines it is sure of, with a known pattern; the rest go to a person | After the backtest shows at least 98% accuracy for that type |
| 2 Whole statement to the buffer | Posts the whole statement to the buffer; a person approves only the closing | After two month-end closes without corrections |
At every level, moving entries from the buffer into the books needs a person: AI proposes, people decide.
AI OS
One AI system for the whole finance team, added step by step
We start with the task that takes the most hours, and add modules on the same AI Box once the first one has proven itself.
- 1MVP
Bank statement agent
Fetch, check, recognise, match, post and verify, as shown above, with the review panel for your accountant.
- 2MVP: basic
Ask your books
“What is the VAT to pay for September?”, “Compare supplier payments in Q3 this year and last year”, “Which invoices over PLN 10,000 are more than 30 days overdue?” Answers with numbers and the source, read-only.
- 3Phase 2
Auditor's pack
Turnover and balances of the accounts the auditor asks for, with the audit log of how every entry was made.
- 4Phase 2
Confidential data room
Due diligence and share purchase documents in one place, searchable in plain language, with access per document and per project, watermarks and an access log.
- 5Phase 3
Information flow
Who is responsible for what, who should receive a piece of information and whether they did: no more “you forgot about this person”.
Pilot
From a demo on your laptop to the agent in your office
- 15 days
Demo on your laptop
A .zip you unpack and run with a double click: a sample statement, a program in the style of your accounting system and demo data. 30 minutes together to go through it.
- 22–4 weeks
Discovery and data
Your accounting system and how to connect to it, your transaction types and posting rules, 6–12 months of statements and entries. Baseline: how long a statement takes today.
- 32 weeks
Backtest
The agent goes through your last 3 months of statements and we compare its entries with yours. Result: accuracy per transaction type.
- 41 month-end close
Shadow mode
The agent proposes, your team works as usual, and we compare. The closing balance must match the bank on every statement.
- 5From month 2
Production
Autonomy raised type by type, monthly reports, then the next modules of the AI OS.
What we measure
KPIs we measure before the pilot and prove after it
| KPI | How it turns into money |
|---|---|
| Time to post one statement | Accountants' hours: the target is at least half the time of today |
| Share of lines posted without correction | More clients or accounts without hiring |
| Corrections by your accountant | Quality and trust: the basis for raising the autonomy level |
| Errors caught by the checks | Corrections avoided, and fewer questions at the audit |
| Days to close the month | Reports for the management sooner |
“The agent does the posting, the checks catch every mistake, your accountant makes the decisions. And the books never leave your office.”
All our offers run on the same on-premise platform: an AI Box in your local network with your own models. Your data does not leave the company, which matters for GDPR, trade secrets and medical data. We take care of the hardware, updates and secure remote service (VPN, SLA).
Start with KaizoBooks AI OS for Accounting
Book a short discovery call. We measure your baseline KPIs first, so the result can be proven in numbers.
Request MVP