CAG to launch in-house AI model to help auditors analyse financial data

The Comptroller and Auditor General of India (CAG) is set to introduce an in-house large language model (LLM) on Monday, bringing years of internal inspection reports, auditing standards and guidelines onto a single platform. The launch coincides with the introduction of new audit plans driven by data analysis and the impending tabling of its first integrated audit reports.

Developed exclusively for CAG by the Indian Institute of Technology Madras, the LLM 2.0 has been trained on around 65,000 internal inspection reports produced by the financial watchdog, along with its auditing standards and guidelines, Krishnan Sangaran Subramanian, deputy comptroller and auditor general, said on Thursday.

Describing the platform as a facility that allows auditors to obtain responses by entering simple queries, the official said data from inspection reports, once ingested by the model, would complement the conventional audit management system used by CAG.

“We’ll upload this year’s and last year’s financial statements, the auditor’s report, and the board bills… Once we upload this, it (LLM) will do all the numerical calculations completely. It will do our analysis. It will show a pattern. So, these are very routine works. Now, the routine work will be done by the computer,” Subramanian said.

The model is likely to have three distinct streams: one each for compliance audit, financial audit and works audit.

Apart from the model, data analysis is also changing how the watchdog prepares its audit plans. Instead of the conventional approach, which centres on the unit to be audited, data from multiple departments and patterns emerging from it are now driving the audits planned for 2026-27, according to the official.

For instance, in the states, data relating to direct benefit transfers under various social sector schemes, public works and even salary payments to public officials disbursed digitally are being analysed for red flags, which audit teams then pursue.

In the social sector, the watchdog has audited seven schemes as part of its first integrated audits, which are nearing completion and are expected to be presented soon. The official said CAG had been working on these audits for the past one to two years. The exercise also includes integrated performance audits of seven social sector schemes, as well as audits relating to ease of doing business and ease of living in 101 cities with populations of 500,000 or more. Integrated audits are also being undertaken in the areas of multimodal transport and foundational literacy.

These integrated audit reports involve multiple stakeholders, resulting in faster audits, the official said. Such audits are also planned over multiple years. The integrated audit of foundational literacy and numeracy will be followed by audit reports on school dropouts and skill upgrades in the coming years.

Addressing concerns about data integrity surrounding the LLM, the official said users — in this case, auditors — would only be able to ask questions based on the underlying reports used to train the model. They would not be able to upload any material onto the platform or the Cloud.

However, the official said the LLM would not conduct audits itself. “We are using it as an assistant. We still have not reached a stage in which it will do an audit. It will be an aid… which will do the preliminary groundwork,” Subramanian said.