Built by IIT Madras and trained on 65,000 of the CAG’s own inspection reports, the model is meant to take over the grunt work of auditing. The details officials have shared are more interesting than the headline.

CAG LLM 2.0 is the Comptroller and Auditor General’s first in-house large language model, and it is set to be introduced on Monday, November 16, at Audit Diwas. Built exclusively for the CAG by IIT Madras, it is meant to take over the repetitive groundwork that eats into auditors’ time. It is not a public chatbot, and officials are clear that it won’t be doing the audits itself.
What is CAG LLM 2.0, exactly?
Think of it as a well-read junior colleague who has only ever read the CAG’s own files. Deputy CAG Krishnan Sangaran Subramanian told Business Standard that it was trained on around 65,000 internal inspection reports, along with the CAG’s auditing standards and guidelines. Auditors type simple questions and get answers. The system sits alongside the existing audit management system rather than replacing it.
The Tribune reports three streams: compliance audit, financial audit and works audit. Business Standard calls the split “likely,” so treat the exact structure as provisional until launch details are out.
What it will do, and what it won’t
Subramanian’s example was concrete. The CAG would feed in a company’s financial statements for this year and last, the auditor’s report and the board bills. The model would handle the calculations, run the analysis and flag patterns. “The routine work will be done by the computer,” he said.
But he drew a clear line. The model is an assistant for preliminary groundwork, and the CAG has not reached the stage where it can do an audit. Auditors can only ask questions based on the reports the model was built on.
Two things that don’t add up yet
The arithmetic. Subramanian said the model will carry out the numerical calculations in full. Language models have a long record of slipping on arithmetic unless the maths is handed to a calculator or code. The CAG hasn’t said how LLM 2.0 handles this, and that matters for an institution whose reports get tabled and scrutinised.
The uploads. The same Business Standard report describes uploading financial statements, yet also says auditors won’t be able to upload material to the platform or the cloud. Perhaps uploads happen through a controlled process on the CAG’s side. Nobody has explained it.
A year in the making
- April 2025: IIT Madras’s Prof. Madhusudhanan Baskaran demonstrated LLM applications in audit at the CAG’s Strategic Leadership Conclave.
- September 2025: SAI India and IIT Madras presented plans for an LLM trained on audit and inspection reports at the Global Audit Dialogue. Days later, a CAG release listed CAG-LLM as an initiative under development.
- November 2025: At Audit Diwas 2025, CAG K. Sanjay Murthy said it was in development. The Tribune reported that, with over 20,000 inspection reports generated a year, he expected it to improve auditor performance significantly.
- November 2026: LLM 2.0 is scheduled for unveiling.
One number to hold lightly: at 20,000 reports a year, 65,000 reports is roughly three years of output. That is our arithmetic, not the CAG’s. The CAG hasn’t said what period the training reports cover.
Why it matters beyond the CAG
Business Standard reports that audits planned for 2026-27 are increasingly driven by data patterns across departments. In the states, for instance, data on direct benefit transfers, public works and digitally paid salaries is being analysed for red flags that audit teams then pursue. A tool that helps auditors sift documents faster fits that shift.
The Tribune adds that the CAG’s AI Champions will present a roadmap for wider AI use in financial audit, Income Tax audit and the finalisation of pension cases for state government employees.
There is also a design choice worth noting. Instead of a general-purpose commercial model, the CAG built a narrow one on its own archive. In our reading, that keeps the model confined to CAG material, which fits the data-integrity assurances officials gave. The trade-off is that its quality rests on that archive and on testing the public hasn’t seen.
What we still don’t know
- The model’s size, base model and context window have not been disclosed in the reporting we reviewed.
- No accuracy results have been published.
- There is no public description of how errors are caught before they reach an audit report.
- Whether the launch details match the pre-event descriptions.
FAQs
Is CAG LLM 2.0 available to the public?
No public access has been announced. Reporting describes it as a tool for CAG auditors.
Will it replace auditors?
Officials say no. They describe it as an assistant for preliminary groundwork.
Who built it?
IIT Madras, exclusively for the CAG.
⚠️ Disclaimer: This article is based on publicly available reports and official CAG releases. The CAG LLM 2.0 launch was scheduled for November 16, 2026, and details may change. aitechnews.in is not affiliated with the CAG or IIT Madras. We will update this story when official launch details are released.
📎 Sources: Business Standard · The Tribune · The Tribune (Audit Diwas 2025) · CAG release, Apr 2025 · CAG release, Sep 12, 2025 · CAG release, Sep 17, 2025
