NHAI turns to AI and open source to preserve institutional knowledge lost to officer transfers

Marg Sarthi is NHAI’s internal AI assistant for everyday administrative and technical work. It gives officials a conversational way to search project data and policy documents instead of checking many separate sources. This matters because highway work involves complex rules, contracts, approvals and progress information.

For example, an officer managing projects in Bengaluru could ask the assistant to find unresolved problems. Marg Sarthi searched monthly progress reports, grouped problems by category and displayed them on a dashboard. It could then create a PDF report for contractors, helping the officer request explanations and track follow-up.

The assistant is designed to make its answers auditable. When asked about a possible ₹50 crore additional work proposal, it requested contract and approval details before responding. It then cited relevant rules, so officials could check the original documents. Role-based access also connects its information access to authorised project permissions.

NHAI officials draw on a large and varied information base for routine work. It includes laws, circulars from the Ministry of Road Transport and Highways, more than 1,800 internal NHAI policy circulars, and standards issued by the Indian Roads Congress. The article does not give one combined document count, but it shows that the knowledge is extensive and fragmented.

The challenge becomes sharper in complex tasks. Approving contractor bills, drafting notices, handling land acquisition and obtaining environmental clearances can require officials to consult several rules and documents. Officials often kept personal folders of relevant circulars and guidelines, updating them whenever new material appeared.

This dependence on scattered sources also created a continuity problem. When officers were transferred or sent on deputation, incoming officials had to rebuild their working knowledge. NHAI’s AI initiative aims to make information easier to find and institutional expertise easier to reuse, while citations help officers verify the rules behind an answer.

Transfers and deputations can weaken NHAI’s institutional memory. Experienced officers often build personal collections of circulars, guidelines and explanations that help them complete work efficiently. When they leave, that practical context may not remain available to the next officer. The issue is not a lack of expertise, but the difficulty of making expertise reusable across the organisation.

The article describes how incoming officials may need to consult many sources again. They must reconstruct knowledge about laws, ministry directions, internal policies and technical standards. This affects complex work, including contractor bills, notices, land acquisition and environmental clearances. Knowledge built in one office may therefore be difficult to transfer to another.

NHAI is exploring reusable workflows to reduce this loss. Experienced officers could explain a process, list its documents and approvals, and have other experts review it. Once incorporated into a system, the workflow could be used elsewhere, helping preserve context and prevent previously identified mistakes from recurring.

Reusable AI-assisted workflows would capture how recurring NHAI processes actually work. An experienced officer could explain the steps, identify required documents and approvals, and describe the relevant rules. Other experts could review the procedure before it enters the system. This would convert individual know-how into guidance that other offices can access.

The approach could support tasks such as approving contractor bills, handling land acquisition and obtaining environmental clearances. Instead of starting with scattered documents, an official could follow a reviewed workflow that points to the necessary stages and records. A workflow developed in one office could then be adapted or used in another.

The immediate value is continuity and consistency. Officers would have a clearer route through complicated administrative work, even after transfers or deputations. The system could also preserve mistakes already identified by experienced staff, helping prevent them from being repeated. The article presents this as an experiment, not a completed organisation-wide system.

Marg Sarthi combines project information with policy material through a conversational interface. For project monitoring, it can search monthly progress reports and identify unresolved issues. For administrative questions, it retrieves relevant rules and uses them to prepare an answer. The goal is not merely speed, but a response that officials can inspect and verify.

In the Bengaluru demonstration, the assistant found problems across highway projects, grouped them by category and placed them on a dashboard. It could also produce a PDF report for contractors. In a separate example, it considered whether a ₹50 crore additional work proposal could be approved for a project originally valued at ₹250 crore. It first requested contract and authority details.

After receiving the needed context, the system gave an answer with citations to relevant rules. Officials could compare that response with the original documents. The team also distinguished retrieval errors from answer-generation errors, using citations and officer feedback to identify information gaps and improve the system.

Role-based access control limits an AI assistant according to an officer’s existing permissions. This is important when the system can search project reports, contracts and internal policy information. Without matching access rules, a convenient assistant could expose data beyond an employee’s responsibilities. Access controls connect usefulness with appropriate information boundaries.

NHAI’s speakers said Marg Sarthi’s controls were linked to the authority’s existing project-management system. An officer’s access through the assistant was intended to correspond to the projects and information that officer could already access there. Thus, the assistant was not designed as one unrestricted window into every NHAI record.

This approach also supports safer adoption inside a large organisation. Officials can use conversational search while retaining the permissions already defined in the underlying system. The article does not describe every security safeguard or provide evidence of a completed deployment, but it identifies permission matching as a central design feature for handling project and contract information.

Open-source software is software developed with source code available for inspection, modification and reuse under its licence. For an organisation testing internal AI systems, ready-made open-source tools can reduce the amount of basic infrastructure that must be created first. Teams can adapt existing components, test ideas quickly and keep more control over how systems are assembled.

LibreChat could provide a conversational interface, while MCP can help connect an assistant with tools or data sources. Langfuse is commonly used to observe and evaluate AI interactions, and Forgejo can host and manage code. Together, such building blocks could support experimentation, although the supplied article text does not describe NHAI’s specific use of each named tool.

The article does state that NHAI’s internal AI team uses open-source software to build tools for administrative and technical work. Faster prototyping could help the team test workflows, retrieval, citations and access controls before wider use. These tools would not remove the need for expert review, permissions or source verification.