agenor.
How it works

AI prepares it. Code computes it. Your CA approves it.

Agenor is a governed-AI tax back-office for CA firms. AI reads documents in whatever form they arrive; deterministic code does every calculation; and nothing files until a human CA reviews and signs off. This page explains exactly how that works — and why a firm can trust it with its clients.

On the roadmap

Automatic connections

Today, documents are uploaded and read securely. These connections are coming, so less is keyed by hand.

🏦
Account Aggregator
Pull bank interest and holdings with consent, as a registered FIU.
Coming soon
📤
ERI auto-file
Submit the return through the Income Tax Department's official API.
Coming soon
📥
AIS / 26AS pull
Fetch the taxpayer's prefill data via ERI, with OTP consent.
Coming soon
📈
Broker connections
Capital-gains P&L via a broker's MCP or official API, with the user's own login.
Coming soon

01What it is

Agenor turns a CA firm's tax workload into an AI-assisted, fully-audited workflow that the firm runs under its own brand. A client answers a few plain-language questions and hands over their documents however they have them — a Form 16 PDF, a broker statement, a screenshot of a figure. Agenor extracts what's there, computes the return, and presents it to the firm's CA to review, correct, and approve.

It is not a consumer self-filing app, and it is not another desktop tool. It is the layer a firm couldn't build itself: AI assistance with the governance, audit trail, and security that filing other people's taxes demands.

02The core principle

One rule shapes everything: AI extracts, deterministic code computes, a human approves. No AI ever authors a number or a tax rule.

AI extracts

The model reads documents in any layout and pulls the figures out — format-resilient, so a new Form 16 template or an odd broker export still works.

Code computes

Every rupee of tax is calculated by deterministic, version-controlled TypeScript — slabs, regimes, §24a/§24b house property, §111A/§112A gains, §71 set-off, surcharge, §87A. Auditable, repeatable, testable.

CA approves

The computed return is the CA's to verify, edit, and approve. Nothing is filed on its own. The firm's professional judgment stays in the loop on every return.

It self-checks

Extraction is cross-checked against the document's own totals; on a mismatch the model re-reads, and if it still can't reconcile it flags the discrepancy loudly rather than emitting a confident wrong number.

The result is the opposite of a black box: a figure you can trace to its source, a calculation you can re-run, and a human accountable for the filing.

03How a return flows

1
Client answers questions. Plain-language, adaptive — only what applies to them. The answers determine exactly which documents and figures are actually needed.
2
Client provides documents & figures. Upload a Form 16, a capital-gains statement, type a rent or 80C amount, add house properties. Whatever form the data is in.
3
The CA loads the client. From the firm dashboard, the CA opens any client and the return assembles itself — answers become the tax picture, and the client's documents are reused with one click (no re-uploading).
4
The estimate computes automatically. Total income, the recommended regime, refund or balance due — calculated deterministically and shown the moment the client is loaded.
5
The CA reviews & approves. Full breakdown, regime comparison, every flagged item. The CA corrects anything, then approves. Only then is the return submitted for filing.
6
The client sees the validated result. Once the CA signs off, the client sees the final, CA-approved figures on their own screen. Carry-forward losses are recorded for next year.

04Governance & audit

Filing on behalf of a client is a liability. Agenor is built so a firm can adopt AI without taking on that risk:

Every figure is sourced

Each number traces to the document or answer it came from. No figure appears without a provenance.

Approval gate

An efile_return is always gated — it cannot file without an explicit CA approval recorded against it.

Tamper-evident trail

Every governed action — extraction, compute, approval — is written to an append-only, hash-chained audit log.

Flag, don't guess

When a document doesn't reconcile, the system surfaces a precise "review before filing" warning instead of a confident wrong number.

05Security & privacy (DPDP)

The product is built by a security team, for data that deserves it. Under India's DPDP Act 2023:

·
Itemized consent before processing. The taxpayer chooses what may be done with their data, can withdraw, export (§11), or delete (§12) it at any time, with a published grievance contact.
·
Sensitive identifiers are redacted. PAN, account numbers, and Aadhaar are stripped before any data reaches an AI model and before it's persisted.
·
PAN & DOB are never stored. They're used transiently to unlock a password-protected AIS/TIS PDF on our own server, then discarded — never sent to any external service.
·
No portal passwords, ever. We never ask for a taxpayer's income-tax portal login. Data comes from documents the taxpayer provides, or via consented, government-sanctioned channels.

06Accuracy & evals

Because code does the math, the math can be tested. Agenor's tax engine is pinned by a suite of golden test cases — known inputs with hand-verified expected outputs — that run automatically before every release and on a daily schedule.

If any tax calculation drifts, the eval suite fails the build and raises an alert — the regression is caught before it ever reaches a return. The extraction layer is covered too, with its own cross-checks. Accuracy isn't a claim; it's a gate.

Coverage spans the mechanisms that actually trip people up: regime comparison and the §87A rebate, §24a/§24b house property (single and multi-property, with §71 set-off and carry-forward), §111A/§112A capital gains and the basic-exemption adjustment, intra-head loss set-off, crypto §115BBH, the surcharge bands, and Schedule AL.

07Where AI helps, and what's unique

Given that code does the math and the CA files, it is fair to ask where AI actually earns its place. In the one part nothing else does well: understanding messy, varied documents.

AI: the format-resilient part

A Form 16 from one employer looks nothing like another's; a Zerodha P&L, a Dhan P&L, a screenshot, all different. AI reads any layout and pulls the right figures, classifies each document, maps plain-language answers to tax situations, and reconciles against AIS. A parser per format would be brittle; AI generalises.

Code: the exact part

Every number is computed by deterministic code, because a hallucinated tax figure is a notice or a penalty. No AI authors a number.

The CA: the human part

Judgment, edge cases, professional accountability, and the filing itself stay with the CA.

What is unique is the boundary itself, enforced as architecture. Most AI tax tools either let AI do everything (untrustworthy, it authors numbers) or are deterministic desktop tools with no AI and all manual entry. Agenor is the only one where AI only extracts, code only computes, and the CA only approves, and that split is redacted, gated, and audited. The differentiator is not "AI files taxes", everyone claims that. It is the trust architecture that makes AI safe on regulated financial data, built by a 22-year security team.

08For CA firms

Agenor is designed to be run by a firm, under the firm's brand:

Your brand, your clients

The client-facing flow carries your firm's identity. Agenor is the infrastructure; the relationship stays yours.

Your filing channel

Returns file through your firm's own e-Return Intermediary registration — no new regulatory step to begin.

Built for many clients

One dashboard, every client, the documents they've already provided reused — designed for a firm's volume, not a single return.

Trust as the product

The governance, audit, and security layer is the point — so you can offer AI-speed without AI-liability.

Interested in a pilot for your firm? jayesh@agenor.ai.