Nirnay
निर्णय · decide, and defend it

Decide every claim
before you pay the supplier.

Every month, thousands of supplier invoices need an Accept / Reject / Pending call before a hard deadline — and a wrong call sends cash out on an invoice you must reverse, with 18% interest. Nirnay puts an approved, statute-cited decision in front of the payment run, so the money is held before it leaves.

The doctrine:The model proposes · Rules validate · Humans approve · The system records.

Why now · What it is · How it stays safe

Why
A hard new deadline

From 2026, makes every invoice an Accept / Reject / Pending call with a hard deadline — and the mistakes are set to lock in permanently.

What
A control layer before payment

Recommend each call with the exact law and hold the payment on what's at risk — so nothing leaves on an invoice you'll have to reverse. Every decision is kept, so a future notice becomes a drafted reply.

How
AI never touches the money

Plain code owns every calculation; the AI only judges, explains and drafts — and a person approves before anything is saved.

Matching stops at the mismatch. Nirnay acts on it — before you pay.

01
Decide, with the law

For each problem invoice an agent recommends the treatment and cites the exact section — you approve in one click.

02
Hold the money at risk

Once you approve a Reject or Pending, the payment is held before treasury runs it — so cash doesn't leave on an invoice that has to be reversed. No payment moves without a person's call.

03
Defend, from memory

Because every decision is recorded with its evidence, a becomes a drafted reply — not a scramble.

How it works

Part 1
The screens
  • Seven screens, one for each job — from the group-wide view to the analyst's day-to-day workbench.
  • A person only ever reads the reasoning and clicks approve.
Part 2
The thinking (kept in two halves)
  • Plain code does every calculation — the maths, the matching, the interest, the safety checks. Same answer every time.
  • AI only investigates, suggests and drafts — always citing the law, never the last word.
Part 3
The memory
  • Every invoice is linked to its order, delivery, supplier filing, the decision made, and any tax notice.
  • A running log that can only be added to — the evidence trail you defend a claim with later.
“Where's the AI? Isn't this just subtraction?” — the honest answershow

Fair question — and the honest answer is: the maths isn't the AI, and it shouldn't be. Working out how much tax you can claim is simple subtraction, so we do it with plain, predictable code that gives the same answer every time. You should never trust an AI to do your tax maths — and Nirnay doesn't. The AI does the harder part a calculator can't: it recommends what to do with a problem invoice, explains why in the words of the law, reads the tax notice, and writes the replies — the work that today takes a person weeks by hand.

Done by plain code (no AI)
  • Match your invoice against the supplier's filing and the goods receipt
  • Add up how much money is at risk, and the 18% interest if it's wrong
  • Check the safety rules before anything is saved
  • Wave through the clean invoices in bulk — a model never touches these
Done by the AI (judgment + writing)
  • Recommends the action — Accept, Reject or Pending — and cites the exact law for it
  • Reads a tax notice written in plain legal English and pulls out each thing being questioned
  • Writes the chase-email to the supplier and the formal reply to the tax office
  • Sorts hundreds of problems by what a senior actually needs to look at — and flags when it's unsure
A quick example. A caterer sends a genuine ₹2.4L bill. The numbers match perfectly, so a simple rule says “accept it.” But the law blocks tax credit on catering — even when the bill is real. The right move is subtle: accept the invoice (don't punish an honest supplier), but flag it to be reversed later, and cite the exact rule. A dropdown can't make that call. That judgment — across hundreds of invoices, every month — is what the AI is for.

The simple test: switch the AI off and the numbers still add up — but you lose the reasoning, the legal citations, the notice-reading, the drafted replies, and the sorting. You're back to a spreadsheet of mismatches and a person making hundreds of legal calls by hand before the deadline. Closing that gap is the product.

The fuller version, on the strategy page →

Where this is heading

A person approves every call today — on purpose. Autonomy is earned one category at a time.

Todaylive
Assisted

AI recommends, a person approves every call. The Autonomous Close screen already clears the trivial cases on its own.

Next
Supervised

The safe, low-value categories act on their own; a person reviews a daily summary. Always reversible.

Then
Policy-bounded

You set the limits — value, confidence, type. The agent works inside them and escalates the rest.

Eventually
Exception-only

People touch only the genuinely new or high-stakes. Everything routine is handled — and recorded.

What earns each rung — and what never changesshow

Autonomy isn't a switch you flip. What earns each step is the decision history itself — proof, category by category, that the machine gets it right before it's trusted to act. Two things never change at any rung: plain code still owns every rupee, and every action is still recorded.

The monthly cycle & the platform

Nirnay helps teams decide and defend GST Input Tax Credit. Every month thousands of supplier invoices need an Accept / Reject / Pending call before a hard deadline — it recommends each with the statute, a human approves, and every decision is recorded so you can defend it later.

Built for the Reliance Enterprise Intelligence assignment — a working demo, not a mockup.
Start the walkthrough