In India, public procurement has long faced challenges such as corruption, lack of transparency, vendor collusion, price manipulation, and bureaucratic inefficiencies. Ensuring fair and efficient procurement across central, state, and local levels is complex. In this context, the Government e-Marketplace (GeM) model becomes especially important.
Pankaj Dixit, CTO at Government e Marketplace (GeM), told AIM on the sidelines of DES 2025 that GeM began using AI for fraud and collusion detection well before the advent of Generative AI.
The Conventional Way
“At the time (before GenAI), we relied on conventional AI, primarily machine learning and deep learning. Our approach began with analysing data from our internal storehouses, focusing on the information, metrics, and metadata surrounding buyer and seller behaviour,” he said.
This included metrics like location (inferred through IP addresses), device usage and bidding patterns. They also checked if users were using the same machines, software or language, and whether their bids featured identical product specification and pricing.
“These metrics are fed into our ML and deep learning models to identify potential red flags. The insights generated from these conventional AI models help surface patterns that may indicate non-transparent transactions,” he said.
The Government e-Market (GeM) portal was launched in 2016 for online purchases of goods and services by all central government ministries and departments.
He recalled how, rather than taking direct action, GeM used this information to support and guide buyers. “If a purchase appears suspicious or not entirely transparent, we flag it for review,” Dixit said.
This approach, he noted, is meant to be collaborative and educational. “Often, discrepancies may be inadvertent, and once made aware, buyers become more cautious. Over time, this led to improved decision-making and a gradual reduction in such incidents.”
Cut to the Present
GeM, the government’s digital procurement platform, has revived how public funds are spent and is now sprinting toward becoming the largest government procurement system in the world, second only to South Korea.
Dixit said GeM has democratised that entire experience, adding that “today, a seller in Assam can bid for a contract in Tamil Nadu or Delhi, provided they have an internet connection and the relevant credentials to access the website.”
He remarked that inclusivity isn’t just a happy accident, noting that about 40% of GeM’s business is driven by MSMEs and SMEs—entities often overlooked or pushed out by larger players in the tender ecosystem.
“It’s something the Prime Minister himself has emphasised—the need to empower small and medium enterprises,” Dixit added. “In the financial year 2024–2025 alone, GeM facilitated transactions worth around $65 billion, a giant leap from the $48 billion the year before and $24 billion the year before that,” he shared.
Referring to academic studies, he said that every transaction through the GeM portal saves the government about 8–10%. “That translates to approximately $15 billion saved so far. All of that is taxpayers’ money—our money—being preserved and redirected for other national needs,” said Dixit.
The AI game
Behind this scale and savings lies a robust technological backbone, one that is now being revamped with an AI-first vision. “While the 1.0 version continues, we are building a new platform with AI at its core,” he revealed.
GeM has already implemented an AI chatbot to improve its cost efficiency and compatibility. This multilingual assistant handles basic queries, converts voice to text, and allows users to create support tickets or check their statuses.
Currently, the only post-login features are ticket creation and status check. With the post-login feature enabled, he said, “We would also like to provide additional information such as my orders, incidents, and where my brand or OEM tool is pending.”
“The next step is to reduce incoming human-handled calls,” he added, noting the team has completed some POCs on conversational virtual interfaces and graph generation. “We are also using generative AI for code analysis and generation,” he revealed, explaining that Generative AI helps us understand code functionality.
He noted that this significantly benefits the development teams and accelerates the time-to-market. Additionally, Dixit mentioned that it is being utilised for secure and innovative code generation.
The team has also established a Retrieval-Augmented Generation (RAG) system and a knowledge base, directing the LLM to respond solely to approved content.
“We’ve added input and output guardrails. Input guardrails prevent non-contextual or mischievous queries, and Output guardrails ensure the generated answers stay within acceptable and accurate bounds,” Dixit noted.
