---
title: "An India builder used Claude Code sub-agents to process about 2,000 invoices a day for a CA firm and replace roughly ₹1.8L a month of data entry"
person: "Archit Mittal"
handle: "@automate-archit"
profile: "https://dev.to/automate-archit"
tools:
  - "Claude Code"
  - "Tally"
  - "Postgres"
country: "India"
category: "Save money"
category_slug: save-money
proof:
  - label: "Read his post on DEV"
    url: "https://dev.to/automate-archit/i-built-a-claude-code-sub-agent-that-processes-2000-invoices-a-day-heres-the-exact-setup-30dp"
date: 2026-11-07
date_posted: 2026-04-20
as_claimed: "The result is shown as claimed in the public source linked here. It has not been checked and is not a typical outcome."
url: "https://agentstories.org/stories/archit-mittal-claude-code-invoice-subagents.html"
markdown_url: "https://agentstories.org/stories/archit-mittal-claude-code-invoice-subagents.md"
site: "Agent Stories by Solopreneur Code"
---

# An India builder used Claude Code sub-agents to process about 2,000 invoices a day for a CA firm and replace roughly ₹1.8L a month of data entry

Archit Mittal @automate-archit · India · Claude Code, Tally, Postgres

## What the agent did

Archit built three Claude Code sub-agents for a Gurgaon CA firm: one extracts GST invoice fields from PDFs, one validates tax maths and vendor data, and one prepares posts into Tally. A small Python loop runs the chain on a shared Drive inbox.

## Result, as claimed

A Gurgaon CA firm was getting 2,000+ vendor invoices a day into a shared Drive folder, with three junior accountants spending about six hours each keying them into Tally, about ₹1.8L a month in salary cost for data entry. Archit built three Claude Code sub-agents (pdf-extractor, tally-validator, tally-writer) plus a Python loop. In week one he reports about 2,047 invoices a day average throughput, 94.3% auto-posted without a human touch, 5.7% kicked to an exception queue, about ₹840 a day in compute cost, the ₹1.8L a month of manual effort replaced, and about a 2-week payback. A later prompt tweak cut the exception rate to 3.1%.

> The result is shown as claimed in the public source linked here. It has not been checked and is not a typical outcome.

## Proof

- [Read his post on DEV](https://dev.to/automate-archit/i-built-a-claude-code-sub-agent-that-processes-2000-invoices-a-day-heres-the-exact-setup-30dp)

## Try it yourself

Split invoice intake into extract, validate and post steps so each coding-agent pass stays small and checkable.

Prompt:

```text
I need a three-step invoice pipeline. Step 1 extract: from this invoice text [paste], return JSON with vendor, tax id, invoice number, date, line items and totals. Step 2 validate: check the maths and flag anything that fails my rules [paste GST or tax rules]. Step 3 prepare a post payload for my accounting tool [name the tool], but do not post it. Return pass or fail with reasons. Do not invent missing fields. Mark gaps as [missing].
```

Tools:

- [Claude Code](https://claude.ai/code): Named for the sub-agent setup that extracts, validates and prepares Tally posts
- [Tally](https://tallysolutions.com): Named as the accounting system the writer agent posts into

Steps:

1. Start with ten real invoices and write your validation rules down first.
2. Run extract alone until the JSON is reliable.
3. Add validate, and only then wire a write step behind a human approval gate.
4. Log every fail reason so you can fix the prompt instead of guessing.

Watch out: Auto-posting into live books is risky. Keep the writer behind approval until the fail rate is low on your own samples.

## More

- [This story as a web page](https://agentstories.org/stories/archit-mittal-claude-code-invoice-subagents.html)
- [More stories: Save money](https://agentstories.org/categories/save-money.html)
- [More stories from India](https://agentstories.org/countries/india.html)
- [More stories that use Claude Code](https://agentstories.org/tools/claude-code.html)
- [Every story, indexed for AI assistants](https://agentstories.org/llms.txt)
