The AI drafts. You approve.
Receipt scanning, a document from one sentence, questions about your books, Excel import and bulk receipt upload. Each one produces a draft that waits for you on screen.
Ask your books in plain language
Ask a question about the business and get an answer with numbers from your own data. The model runs four fixed queries: documents, expenses, VAT summary and top customers.
The VAT summary in the chat comes from the same computation that produces the periodic report, so the number is identical in both places.
A photo of a receipt, an expense ready to review
Photograph a receipt and get an expense draft with supplier, amount, VAT and category. The deduction percentage does not come from the model: it is read from the category catalog and stamped by the server at save time, then frozen on that expense even if the law changes later.
A whole folder of receipts at once
Drag in the month's receipts. Each one is scanned separately and lands in a review table where you can correct a field or drop a row before saving.
Saving runs row by row through the same server-side checks as a hand-typed expense.
One sentence, a full document draft
Write what happened and get a draft with line items, quantities and prices. The customer name is matched in the browser against the list already loaded, so your customer list never reaches the model.
Excel import without rearranging columns
A file from your previous system, with headers in Hebrew or English and in any order. The AI maps each column to a field and you see the mapping before you approve it.
Historical documents keep their original numbers, held apart from live numbering so they cannot collide with new documents.
What the AI does not do
Three limits enforced in server code, not in instructions to the model.
Fixed, read-only tools
The model can run four predefined queries over your data. There is no free-form SQL tool and no way to write a record through the chat.
Nothing saves itself
Every AI output is a draft on screen. Saving, issuing and sending happen on your click.
The math stays on the server
Totals, VAT and deduction percentages are recomputed server-side at save time, not taken from what the model suggested.
Common questions
Can the AI issue an invoice without me knowing?
No. No AI code path writes to documents, expenses or support messages. The output always arrives as a draft waiting for approval, and issuing itself goes through the same signing and numbering flow as a document typed by hand.
What gets sent to the model?
Only what the request needs: the receipt image for a scan, the sentence you typed for a document draft, the column headers for an Excel import. Your customer list is not sent; customer-name matching happens in the browser against the list already loaded.
Which model do you use?
Claude and Gemini, configured per feature on the platform side. If one provider is unavailable, a single feature can be moved to the other without touching the rest.