The CompanyStack Tally Companion: AI for people who refuse to leave Tally

Every AI accounting pitch in India eventually says “just move your data to our cloud”. Tally users, correctly, refuse. The books are on their machine, the auditor trusts the edit log, twenty years of muscle memory lives in those keyboard shortcuts, and nobody who runs a real practice migrates a system of record because a demo looked nice.

So we built the opposite: the CompanyStack Tally Companion, a desktop app that sits next to your own TallyPrime and talks to it locally. Your data stays in your Tally, on your computer. The AI comes to it.

What it actually is

A native app (Windows, macOS, Linux) that connects to the TallyPrime already running on your machine through Tally’s own local gateway. You pick a company, and you work by conversation instead of menu-diving:

“Show me all payments to Sharma Traders this quarter.” “Export this year’s sales vouchers to CSV.” “Which ledgers have negative balances?” The agent reads your books and answers with the actual figures, rendered as tables and cards, with the underlying vouchers a click away.

That alone changes who can use Tally. The power user gets there faster; the founder who never learned Tally’s keystrokes gets there at all.

The bank statement pipeline, which is the feature people stay for

The single most tedious job in Indian bookkeeping is turning a bank statement PDF into posted vouchers. The Companion does it as a pipeline:

1. Attach the statement, PDF or image.
2. A vision model extracts every row, and then, because models misread things, a deterministic running-balance check recomputes the balance line by line and flags any row where the arithmetic breaks. An extraction error cannot slip through silently, because the balance stops adding up.
3. You get a mapping panel: pick the bank ledger, confirm the suggested counterparty for each row (the model proposes, you dispose).
4. The app posts the Payment and Receipt vouchers into Tally, behind an explicit confirmation that warns you about duplicates before anything is written.

An afternoon of data entry becomes minutes of review. The judgement stays yours; the typing does not.

Document management: files attached to obligations, not folders

Most practices keep documents in a folder tree that answers “where did I put it” and never “what is it for”. The Companion’s document layer files every document against the thing it discharges: the receipt against its expense voucher, the challan against the filing it proves, the engagement letter against the client it governs.

That inversion matters for two specific rupee reasons. GST input credit under Section 16 requires possession of the tax invoice, so an unvouched purchase is input credit at risk. And an expense without substantiation is an expense the assessing officer can disallow under Section 37. The Companion matches captured receipts to vouchers (amount, date proximity, fuzzy vendor match) and gives you the aged list of unvouched expenses, ranked by what they would cost you. Vouching stops being an audit-season scramble because the sample self-populates all year.

The trust model, which is the part we are most opinionated about

AI touching a system of record needs rules, and ours are strict:

Every action is gated. Each tool call pauses for your Approve or Deny before it runs. Nothing writes to Tally without you seeing exactly what is about to happen. Posting vouchers and installing companies get their own full-screen confirmations.

The model never sees your bulk data. Exports and statements pass through as files and summaries. When you export 40,000 vouchers to CSV, the AI sees the file path, the row count and a five-row preview, not your ledger.

Nothing leaves your machine except what you approve. The connection to Tally is local. This is not a cloud sync wearing a desktop icon.

We respect the edit log. TallyPrime’s audit trail exists so your auditor can rely on your books under Rule 11(g). A tool that writes to your ledger is writing to that record, which is precisely why every write in the Companion is explicit, confirmed and attributable to you.

For power users, an escape hatch

If you know TDL and Tally’s XML envelopes, the Companion exposes them: ask for raw gateway queries when the built-in tools do not cover your case, with the bundled TDL reference available to the agent. It is the difference between a toy that demos well and a tool a practitioner can push on.

Where it fits

The Companion is for people who do the work themselves: CA firms running client portfolios on Tally, and businesses with an in-house Tally operator. If you would rather nobody in your company did the work at all, that is CompanyStack proper, our managed service, and the two share an engine. The comparison of Tally-as-record versus the work around it is in a separate post. Either way, Tally stays exactly where it is. That was the design constraint everything else followed from.