Tally AI

Overview

A financial assistant for manufacturing and distribution businesses running on Tally ERP. Operators ask the questions they actually care about — why is this month off, which customer is drifting, where did margin go — and the system decomposes each one across agents that query live ERP data and reconcile the answers.

The problem

SMBs sit on years of ERP data they cannot interrogate. The questions that matter take an analyst to answer, and these businesses do not have one. A single-shot LLM over the same data was not viable either: multi-step financial questions need intermediate results checked, and an assistant that is confidently wrong about money is a liability, not a tool.

What we built

Natural-language query layer

Questions in plain language resolved against live Tally ERP data, with no export or intermediate spreadsheet.

Multi-agent decomposition

Multi-step questions are split into sub-queries handled by specialised agents, then reconciled — so each step can be checked independently.

Discrepancy analysis

Anomalies in ledgers and receivables are surfaced without being asked for, which is where most of the value turned out to be.

Automated evaluation pipeline

A standing eval suite scores answer quality on every change, so regressions are caught before release rather than reported by a customer.

How we approached it

• Answer quality is measured, not assumed — the eval pipeline shipped alongside the product
• Decompose before answering: financial questions have checkable intermediate steps
• Read against live ERP data; a stale copy produces confidently outdated answers
• Surface the discrepancy the operator did not know to ask about

Outcome

A production assistant giving SMB operators analyst-grade answers over their own ERP data, with measured answer quality rather than anecdotal confidence.