AI that does the data entry
The biggest cost in accounting is typing. Ganav’s AI Document Inbox reads your invoices, bills, and bank statements, extracts every field, and hands you a review-ready draft journal — already coded, checked for duplicates, and anchored to the source document. When it’s confident, it drafts with zero touches.
A green gate auto-drafts a complete journal only when confidence and completeness are high. Anything below the bar stands down — with logged reasons — so automation never becomes a black box. Asset purchases, reverse-charge bills, and anything carrying a warning are never auto-posted: those are decisions.
Ganav takes a census of every GSTIN and PAN on the page, reads the buyer/seller label zones around them, and resolves who is billing whom deterministically. A bill addressed to you can’t flip into a sales invoice because the AI cross-paired two names; a vendor is matched by GSTIN, then PAN, then name — never by guess. Foreign suppliers are recognised by country, so an import of services is treated as one.
Every coding suggestion carries a frequency — “you’ve coded this vendor to this account 12 times” — so the inbox gets more automatic for each supplier over time. Memory never learns an asset-class coding, so a one-off laptop purchase can’t teach it to expense the next one.
Reference-normalized duplicate detection runs across both the inbox and your posted journals, with a red banner and a “create anyway” confirmation. The same bill can’t quietly land twice.
Independent evidence scanners read the tax and the currency off the page itself — so a bill whose GST the AI returned as zero gets its CGST/SGST restored from the printed figures, a bill with no currency token is never “invented” into USD, and a total that disagrees with its lines is corrected only when two independent anchors agree. The draft carries a note explaining what was repaired and why.
When a bill carries a GSTN e-invoice QR code, Ganav reads the signed payload directly — rendered at high resolution so even a small QR decodes — and skips the AI text call entirely, using authoritative header data (with the real IRN persisted). Verified live on IRP-signed invoices.
Bank, card and loan statements each land in a dedicated lane that understands their shape — including
legacy .xls files sniffed by content, not extension, and real Indian bank layouts. An amount-in-words
solver and GST-split repair fix OCR mangling using rupee-exact evidence — never a fuzzy guess.
Read a document as the wrong thing? Switch it — purchase bill, expense, sales invoice, receipt, payment, journal, or a statement lane — in one control. The choice re-reads the stored extraction without another OCR pass, locks once the document is booked, and is remembered for that party so the app routes the next one correctly. Warn, never force: moving into a lane the evidence doesn’t support shows you why.
If a customer paid you net of TDS, the inbox reads the amount withheld, carries it through reconciliation into the receipt, and applies Section 206AA’s 20% where the deductee has no PAN — as a warning you can overrule, never a silent override. Unapplying a scanned deduction asks you to confirm.
A capitalization policy (threshold + nudge) per company means a bill that looks like an asset — by HSN, by amount, by account class, by wording — gets a capex hint, and the Capitalize action books the bill and creates the asset in one flow, net of GST with the ITC choice made explicitly.
An import of services with no tax on its face, or a domestic vendor whose nature attracts reverse charge, gets an RCM proposal: the matching reverse-charge tax code, the category, and the evidence — always as a warning for a human to accept. A foreign supplier who wrongly charged Indian GST is routed to a vendor-correction lane instead. (Engine detail on the Core Accounting page.)
Click a field in the review pane and the source document scrolls and flashes the exact box it came from. Per-field trust dots show OCR confidence, so you know exactly what to double-check.
One-click banners write a vendor’s bank details, country, or MSME status straight to the party master, and an analytics header tracks your straight-through rate, volumes, and AI cost per document.
A CA firm sees the same vendors across many client companies. With Smart Memory reuse switched on — and only where both companies opt in and the acting user has access to both — a vendor’s identity and tax rate learned in one client’s books can be suggested in another’s. Suggestion-only, never straight-through, never account codes, never bank lines: the fail-closed share catalog decides what may cross, and the answer is “very little, and only what a GSTIN proves.”
Ganav’s AI layer is a clean, swappable seam. Claude (Anthropic) and Sarvam AI (India-hosted) are both live. Choose a global default, set it per company, or route per capability — for example, send invoice extraction to Sarvam but bank statements to Claude. You’re never locked to one vendor.
Before any document text reaches an AI provider, Ganav reversibly tokenizes the sensitive identifiers —
GSTIN, PAN, Aadhaar, TAN, CIN, IEC, UAN, bank IFSC/IBAN, card numbers — into placeholders like [[PAN_1]],
then restores them byte-exactly in the AI’s answer. Because the extraction itself needs those IDs, it’s
reversible tokenization, not blunt redaction.