Yes, with a bookkeeper reviewing the result. AI can read PDF and scanned bank statements in Arabic and English, extract every line, suggest an account code from how that client's similar lines were coded before, and flag what it is unsure about. The bookkeeper approves the coding before anything posts to Xero, QuickBooks, Zoho Books or Odoo.
#Key takeaways
- The work that eats a bookkeeping practice is reading statements, not posting entries.
- AI can extract lines and suggest codes; a bookkeeper approves every suggestion before it posts.
- Test on your own clients' statements, including Arabic and scanned ones, and count the misses.
- Bank feeds remove most of the problem where they exist; AI matters for the statements that arrive as PDFs and photos.
- Measure lines reviewed per hour and corrections per statement, before and after.
#What does a bookkeeper's week look like?
A practice with dozens of small clients gets statements every month-end by email, WhatsApp and portal uploads. Some banks offer feeds, many clients send PDFs, and a few send phone photos. The bookkeeper re-keys or imports each one, codes transactions from memory of the client, and asks the client about the lines nobody recognises.
Arabic statements add a step. Descriptions mix Arabic and English, dates may follow different formats, and amounts sit in columns that a generic import tool misreads. That is where most re-keying still happens.
#What can AI do with a bank statement?
- Read digital PDFs and scans, in Arabic and English, and return each line with date, description, amount and balance.
- Check that opening balance plus lines equals closing balance, and stop if it does not.
- Suggest an account and tax code from how the same client's similar lines were coded before.
- Group unknown lines into one question list for the client instead of separate messages.
- Spot duplicates across a PDF and a bank feed for the same period.
The balance check matters most. If the extracted lines do not add up to the closing balance, the statement goes back to a person before anything else happens. That one rule catches most extraction errors. We covered field-by-field testing in can AI read Arabic invoices?.
#How do you test it before trusting it?
- Collect a test set from your own clients: clean PDFs, scans, photos, Arabic and English, several banks.
- Have a bookkeeper code them by hand, as the answer key.
- Run the AI on the same set and compare line by line: missed lines, wrong amounts, wrong codes.
- Decide which error types are acceptable with review and which stop the workflow.
- Repeat on new statements regularly; bank formats change.
Note that Gartner expected 30% of generative AI projects to be abandoned after proof of concept, citing poor data quality among the reasons (Gartner, july 2024). A test set from your own clients is how you avoid joining that number.
#When is a bank feed or an import tool enough?
If most of your clients bank where your accounting software has a direct feed, the feed plus the software's own bank rules will do most of the coding. A standard statement-to-CSV converter is fine for clean digital PDFs in English. AI is worth it when a real share of statements are scanned, photographed or Arabic, and when coding depends on knowing each client's history.
#Where does Stennir fit?
Stennir starts with AI consulting to check that statement reading is the workflow that pays back first; for some practices it is receipts, VAT prep or client queries. If it is, AI implementation builds it with real users on your clients' statements, in Arabic and English, hosted where your rules require, with every suggestion approved by a bookkeeper and logged.
Statements only go into tools your practice has approved: ICAEW has warned that client data in a public AI prompt will very likely breach confidentiality (ICAEW, february 2024). Managed AI operations then keeps watch as banks change their formats. AI training teaches the team how to review suggestions quickly without approving on autopilot. For the next step after extraction, see AI for bank reconciliation.
We are
Stennir builds statement reading and coding suggestions that a bookkeeper approves before anything posts.
We aren't
Stennir is not an accounting package or a bank feed provider, and it does not post entries without a person approving them.
#What should a practice owner do next?
Count how many of last month's statements arrived as scans, photos or Arabic PDFs. If it is a meaningful share, build a small test set. See more on AI for finance teams, then take the free readiness assessment or book a 30-minute call.
#Common questions
Which accounting software does it work with?
The extracted and approved lines can be pushed to Xero, QuickBooks, Zoho Books, Odoo or other packages through their import or API, or exported as a file you import. Which route works depends on your software and your clients' setups, and Stennir confirms that during the audit, before anything is built.
How accurate is it on Arabic statements?
It depends on the bank, the scan quality and the layout, which is why the test set comes first. Stennir measures accuracy on your own clients' Arabic and English statements, line by line, and the balance check stops any statement that does not add up before a bookkeeper sees it.
What could go wrong?
A misread amount, a missed line on a multi-page scan, or a code suggestion approved without reading. The controls are the balance check, a confidence flag on uncertain lines and a review screen that shows the original next to each suggestion. Corrections feed back into the suggestions for that client.
Where is our clients' data kept?
In the region your practice's rules and client contracts require, agreed before any statement is processed. Your practice owns the accounts, the data and the off switch, every approval is logged, and statements are never used to train a public model.
