How it works

Deterministic work is computed. Documents are read. You decide what it means.

The question a partner actually asks is not whether the software is clever. It is whether the file will survive review. This page is the honest answer to that, including what we get wrong and what happens when we do.

Book a demo Read the security brief

The pipeline

Four stages, stated plainly.

Or watch them happen: a revenue cut-off procedure played stage by stage, from the client’s upload to a working paper waiting on its reviewer, on sample data.

Ingest

The Tally company file is connected or an export uploaded. GSTR-2B is pulled for the period. Client documents arrive by upload or through the request list. Lekhya reports what it could read and what it could not — an unreadable scan is listed, not skipped.

Output — A manifest of every file received, its type and its legibility.

Structure

Ledgers are mapped to your chart of accounts, vendor names normalised across spelling variations, documents classified and fields extracted. You confirm the mapping once. It carries forward to next year's engagement with the methodology version recorded.

Output — A structured transaction population with each field traceable to its source.

Test

Named procedures run across the full population. Each one declares the population it tested, the rule it applied and the tolerance it used. Nothing runs that is not listed in the audit programme.

Output — A result for every transaction, and an exception queue for those that failed.

Review

Exceptions arrive for a human. You accept, reject or annotate. Rejections ask for a reason, and that reason is recorded against the finding — it is part of the file, and it is what a reviewer reads when they ask why an exception was not pursued.

Output — A working paper with named sign-off and a complete audit trail.

The split that matters

What is computed, what is modelled, what is yours.

The fear is reasonable: a language model invents a number, it reaches the file, and the partner signs it. The answer is architectural. Arithmetic never touches a model.

Which parts of the audit are computed deterministically, which use a model, and which remain with the auditor
Task	Performed by	Note
Footing, casting and cross-casting	Computed	Exact arithmetic. Identical result on every run.
Three-way matching and reconciliation	Computed	Rule-based comparison on structured fields with defined tolerances.
Duplicate and variance detection	Computed	Deterministic comparison across vendor, amount, date and reference.
Date and cut-off comparison	Computed	Simple ordering of dates once they have been extracted.
Reading a field from a scanned invoice	Model	Extraction with a confidence band. Low confidence routes to a reviewer.
Classifying an uploaded document	Model	Invoice, goods receipt, dispatch note or statement — correctable by the user.
Normalising vendor names across spellings	Model	Proposed groupings, confirmed once by the engagement team and reused.
Drafting the narrative of a working paper	Model	Marked as drafted. Never included in the file without review.
Concluding whether an exception is a misstatement	Auditor	Always. Lekhya presents evidence; the engagement team decides.
Setting materiality and signing the file	Auditor	Always. Recorded against a named user with a timestamp.

Provenance

The chain, end to end.

Every finding holds a link to the procedure that raised it, the transaction it tested, the field that was extracted, the document that field came from and the page it sits on. Walk it yourself.

Finding

Potential revenue cut-off issue

₹18,40,000 recognised in FY 2025–26

Raised automatically when dispatch evidence dated after the year end was matched to an invoice recorded before it.

Step 1 of 5

When Lekhya is unsure

Confidence routes work. It does not decorate it.

You will not see a percentage. Model confidence is poorly calibrated, and a displayed “87%” invites a reviewer to trust a number that does not mean what they think it means. Three bands, each with a stated reason.

Matched

Exact amount, date and vendor across all documents. Posts to the working paper marked as drafted.

Needs review

Amount matches but vendor name differs beyond tolerance. Routed to a named reviewer's queue.

Low

Extracted from a poor scan with unreadable fields. Presented as a gap to be filled, never as a value.

A worked example of being wrong

What happens when Lekhya raises something that is fine.

It will. A tool that never raises a false exception is a tool that is missing real ones. What matters is that dismissing it is fast and that the dismissal is recorded.

PT-0119Invoice value exceeds GL posting by ₹10,800.00Dismissed by reviewer

What Lekhya said

Invoice INV-2026-0421 from Nandan Steel Co. records ₹4,12,000. The ledger entry records ₹4,01,200. The difference equals the GST component.

What the reviewer did

Opened the ledger row, confirmed input tax credit was posted separately to the GST receivable account, and dismissed the exception with that reason in one click.

What the file records

The exception, the evidence, the reviewer’s name, the timestamp and the stated reason — all of it retained in the working paper rather than deleted.

Documentation

Built to be handed to a reviewer.

SA 230 requires documentation sufficient for an experienced auditor with no prior connection to the engagement to understand the work performed, the evidence obtained and the conclusions reached. That is a description of what Lekhya produces by default rather than something you assemble afterwards.

What exports with the file

Every procedure run, with its population and parameters
Every exception raised, and how each was resolved
Evidence links preserved as document and page references
Preparer, reviewer and partner sign-off with timestamps
Every change, with the value it replaced
The methodology version the engagement was run under

The best test is a file you already closed.

Run Lekhya against a completed prior-year engagement and compare its exceptions with what your team found. You already know the answers, which makes it the only honest benchmark.

Book a demo Explore the product