AI Copilot for TallyPrime — Natural Language Reports, Anomaly Detection, GST Reasoning

The Perimattic bridge runs on the same Windows machine as your Tally installation, reads voucher data through Tally's standard ODBC interface, and holds it locally. When a user asks a question, only the parsed query (the question text and the fields it needs) is sent to the AI layer; the ledger content stays on your machine. Query results are computed locally against the indexed data and returned to the user. Nothing about your voucher entries, party balances or GST filings ever leaves the machine. This matches how Busy-WhatsApp and Marg-WhatsApp deployments run — traders and MSMEs who explicitly do not want their books on the cloud.

Any question that can be answered from your Tally data. Common patterns include: revenue and expense trends by period ("compare this quarter to last quarter"); receivables and payables queries ("who owes me more than ₹5 lakh over 60 days"); GST-specific queries ("what is my ITC available for July"); vendor and customer analytics ("which vendors gave me the biggest discount last year"); anomaly-shaped queries ("show me expense entries this month that look unusual"). The copilot rejects questions it cannot answer from your data rather than making up numbers, and cites the source vouchers for every claim.

A generic chatbot answers from its training data or a public knowledge base and cannot see your Tally ledgers. AI Copilot for TallyPrime is grounded in your specific Tally data with structured retrieval — every answer comes with citations back to the exact vouchers used, respects your active company and financial year, and honours your role-based access to particular ledgers. It also understands Indian tax context: CGST/SGST/IGST logic by state, GSTIN validity checks, HSN classification, TDS thresholds, e-invoice IRN reconciliation. A generic chatbot does none of that.

Yes. Query input supports English, Hindi, and Hinglish, and the copilot auto-detects the language of the incoming question. Outputs (both textual answers and WhatsApp debtor-chase messages) can be configured per party — Hindi for parties in the Hindi belt, English or Hinglish elsewhere — using the same state-code-driven language selection that Busy-WhatsApp and Marg-WhatsApp already use. Voice input over WhatsApp is transcribed via the same pipeline and answered in the language of the question.

When a GSTR-1 or GSTR-3B submission fails at the GST Portal, the portal returns a JSON error payload identifying the offending rows. The copilot reads that JSON, correlates each error against the specific vouchers or invoices in your Tally ledger that produced them, and explains the failure in plain English along with the fix — for example, "Invoice #TSA/2026/1234 was rejected because the buyer GSTIN 07AABCT1332M1Z9 shows as cancelled on 12 June 2026. You can either re-issue the invoice to the corrected active GSTIN or move this line to B2C-Large." This is the difference between debugging your filing through a form-by-form manual review and getting a targeted fix in under a minute.

It is currently in early access. Pilot deployments are running with a small number of Perimattic customers who have Tally on-prem plus the Perimattic bridge (typically because they already use Tally-WhatsApp, Tally-Shopify or Tally-Razorpay). Access is by invitation while we tune the copilot against real customer ledgers. To join the pilot, contact the Perimattic team — we prioritise customers who have a clear operational use case (month-end reporting, GST filing pain, receivables chasing) and can commit two hours a week during the first month to feedback.

TallyPrime (all current releases) and Tally.ERP 9 Release 6.6.3 and later. The Perimattic bridge is version-aware and auto-detects which schema the copilot should query. Multi-company Tally installations are directly supported — each company gets its own query scope and audit trail. If you run a mix of TallyPrime and Tally.ERP 9 across group entities, one copilot deployment covers all of them from a single dashboard.

Every query is logged with the user identity (mapped from your Perimattic dashboard user), the exact question asked, the fields the copilot resolved against, the vouchers returned, and the timestamp — all written to your choice of storage (local S3 or on-prem bucket). Every autonomous action the copilot takes (a WhatsApp debtor-chase, an anomaly alert dispatched, an entry it recommends flagging) is logged the same way. This is the audit surface most Indian statutory auditors ask about when AI enters an accounting workflow, and it is designed to answer those questions rather than dodge them.