How AI changes the economics of tax recovery

Every large Indian company leaves some tax money unrecovered. Not because anyone decided to, but because the arithmetic did not work. A missed airline credit of ₹2,000, a TDS mismatch on a small vendor, a refund worth less than the effort of assembling it: each is too small to chase on its own, and together they add up.

What has changed is the cost of the work, not the law.

Why small claims were left behind

Tax recovery has always had two kinds of work. The first is volume: finding the invoice, downloading it, reading it, and matching it to the books and to the returns. The second is judgment: deciding whether a credit is eligible, how to respond to a notice, and when to argue.

Volume work costs about the same per document whatever the document is worth. Reading an airline invoice for ₹800 takes as long as reading one for ₹80,000. So below a certain value, the cost of recovery was higher than the money recovered, and firms and in-house teams sensibly stopped there. Specialists spent most of their day on collection and matching, and only the end of it on the questions that actually needed them.

What systems now do

Three tasks can now be handed to software with reliable results:

• Collection. Pulling invoices from airline portals, vendor systems and the GST portal on a schedule, rather than by request.
• Reading. Extracting GSTINs, invoice numbers, taxable values and tax amounts from PDFs and scans. Document AI of this kind is now a mainstream product category offered by major cloud providers.
• Matching. Comparing each document to the purchase register, GSTR-2B and Form 26AS or its successors, and flagging exactly where they differ.

Once these are automated, the cost per document falls sharply and stays roughly flat as volume rises. Claims that were not worth pursuing become worth pursuing. The chart below shows how we think about where the effort goes.

A worked example

Consider a company with 40,000 airline invoices a year, most of them between ₹500 and ₹5,000 in GST. Under a manual model, suppose a specialist can collect, read and match 60 invoices a day. Reviewing every invoice would take well over 600 working days, so the team reviews only the largest few thousand and writes off the rest. Under an AI-led model, the system collects and matches all 40,000, and a specialist reviews perhaps a few hundred exceptions. The small invoices are no longer written off; they are recovered as a matter of routine. These figures are illustrative, but the shape of the change is what matters.

Why the economics compound

The saving is not only in cost per document. Automated collection runs every month, so gaps are found while they can still be fixed: a supplier can amend an invoice, a missing GSTIN can be added, a claim can be filed inside its time limit. Under a manual model, the same gaps were usually found at year end, when some were already out of time.

There is a second effect. When matching is systematic, the exceptions it produces are cleaner. A specialist no longer sifts a spreadsheet of thousands of rows to find the forty that matter. The system shows the forty, with the documents attached. That is where an experienced reviewer adds most value, and where errors used to slip through because attention had already been spent on the routine.

The result is a different kind of engagement. Fewer hours go on assembling files, more go on the decisions that change how much is recovered.

What AI cannot do

It is worth being precise about the limits, because the claims made for AI are often loose.

Judgment stays with people. Whether a credit is blocked under Section 17(5), whether a refund falls within time, how to read an ambiguous circular: these need a specialist who understands the law and the client's business. A model can surface the question. It should not answer it unsupervised.

Representation stays with people. Replying to a deficiency memo or a show cause notice, and appearing before an officer, is professional work with professional responsibility attached.

Accountability stays with people. When a claim is filed, someone signs off. At TraCarta that is a named specialist, and the client knows who. Software errors are caught because a person reviews every exception before anything reaches the portal.

Data security is part of the design

Recovery work means handling invoices, vendor data, bank realisation certificates and portal access. That data needs the same care as any financial record. In practice that means access limited by role, credentials stored securely and never shared by email, clear retention and deletion rules, and processing that respects the Digital Personal Data Protection Act, 2023 where personal data is involved. Clients should also know which third-party models, if any, see their documents, and on what terms.

What this means for a finance team

For a finance team, a few things change in practice:

1. Revisit old thresholds

   List the categories of credit or refund you have historically ignored as too small, and estimate their total.

2. Ask for exception reports, not spreadsheets

   Expect the provider or internal team to show the items that need a decision, with documents attached.

3. Name the reviewer

   Confirm who reviews and signs each filing, and who handles notices.

4. Check data handling

   Ask where documents are stored, who can access them, and whether any external model is used.

5. Measure recovery, not activity

   Track money recovered, pending and awaiting decision each month, rather than invoices processed.

The practical change is that the threshold for "worth recovering" has moved down. A CFO who wrote off small credits as the cost of doing business should look again. The questions to ask are simple: how much sits below the old threshold, who will do the judgment work, and how will the data be protected.

Our operating model is built on this split, and clients receive recovered, pending and awaiting-decision amounts in one signed monthly statement (see what you receive). To see what sits below your threshold, get in touch.

General information only, not tax advice. Check the current law and your facts before acting.