Q&A: Inside the Multi-Agent System Serving 350,000 Cognizant Employees

A Q&A with the two architects behind 1Cognizant on building a 200-agent system that 350,000 employees actually use.

The challenge for most enterprises is not a shortage of AI tools. It is that those tools were never designed to work together. Teams adopt agents in silos, each one solving a narrow problem on its own stack, and the employee experience becomes more fragmented with every new capability added.

Cognizant ran into this problem at its own scale. With 350,000 employees spread across the globe, the company's internal intranet, 1 Cognizant (1C), had become a patchwork of disconnected tools and portals. Different departments had built their own applications on different tech stacks over the years, and employees were left navigating between them to get anything done. Booking leave meant one system, checking IT request status meant another. When AI capabilities started getting layered on top, they landed in the same silos. The tools got smarter, but the experience stayed fragmented.

Rather than continue adding capabilities onto a broken foundation, Cognizant rebuilt its digital workplace from the ground up and agentified its entire intranet. The revitalized 1Cognizant platform, built on Cognizant's open-source multi-agent framework neuro-san, now serves as a single conversational interface for every employee across web and mobile, coordinating across more than 200 agent capabilities and integrations with Agentforce, ServiceNow, PeopleSoft, and more. Within five months of launch, support ticket volume had already dropped by 50%.

On The Agent Effect podcast, I sat down with Venkatesh Balaji, AVP of Enterprise Architecture, and Dan Fink, AVP of Platform Engineering to explore the architecture behind the build, the tradeoffs the team had to make, and the lessons that only show up once something like this is running at full scale. What follows is a Q&A from that conversation. 

Q&A: Behind the Scenes of the 1C Platform

Q: What is 1Cognizant and who does it serve?

1Cognizant is Cognizant's internal digital workplace platform, essentially an enterprise app store that gives all 350,000 employees a single place to access workflows, processes, automations, news, and AI-powered assistance across web, intranet, and mobile. The goal was to replace a fragmented landscape of departmental tools with one unified interface that any employee could use to get anything done, without needing to know which underlying system owned which task.

Q: What problem prompted the move to a multi-agent system?

As Cognizant grew, different departments built their own AI tools and web applications on different tech stacks, and the result was a fragmented employee experience. People had to jump between applications constantly just to complete routine work. The move to a multi-agent system was driven by a single clear goal: unify everything behind one conversational AI interface so employees could interact with any system through a single query rather than navigating between dozens of separate tools.

Q: What is agentification and how does it apply here?

Agentification means layering a natural language interface on top of existing agents and APIs so users can interact with multiple systems through a single query, rather than navigating each tool separately. For 1Cognizant, that meant taking every major enterprise application the company runs and making it accessible through one conversational interface. The platform started with three agents at launch and has grown to more than 200 capabilities spanning applications and processes across the entire organization.

Q: What role does neuro-san play in the architecture?

neuro-san acts as the master orchestration layer for the entire platform. It receives every incoming user query, determines which agent or agents should handle it, and maintains session history throughout the interaction so the conversation stays coherent even when a single request touches multiple systems.

The routing structure is hierarchical rather than flat. An HR agent, for example, knows which sub-agents to route HR-related questions to, rather than the top-level orchestrator needing a direct line to all 200 agents simultaneously. That hierarchy is what makes the system manageable at scale and allows new capabilities to be added without requiring the entire network to be reconfigured.

Q: How is sensitive data kept secure as it flows between agents?

neuro-san uses a separate private data channel called sly-data that is built specifically for sensitive information. That channel is not injected into logs, observability systems, or chat streams, which means credentials, payroll details, SSNs, and other personal data stay protected as they move through the system. The sensitive information travels where it needs to go, but it never becomes part of the conversation artifact that could be stored, reviewed, or exposed through standard observability tooling.

Q: How do you scale a system like this to 350,000 employees?

1Cognizant runs on Azure in a multi-region distributed architecture. The design principle that makes scale possible is that neuro-san is stateless per query, meaning each request is handled independently. That allows the system to scale horizontally by spinning up additional server instances without the complexity of managing shared state. At peak load, up to 60 server pods can spin up simultaneously to handle traffic, and agent-level usage data informs capacity planning so the team can stay ahead of demand rather than react to it.

Q: What about latency? Is there a noticeable delay?

There is some lag, and it is typically measured in seconds rather than milliseconds. Multi-agent systems require each agent to have its own conversation with an underlying LLM to work through its part of a query, and that reasoning takes time. The right frame of reference is not a search engine. It is how long a human employee would take to contact twelve different applications, gather the relevant information from each, and compile a coherent response, which is far longer than a few seconds. Once users understand what is happening underneath, the response time reads as fast rather than slow.

Q: Does 1Cognizant only use internally built agents, or does it integrate with third parties?

It integrates extensively with third-party platforms. Current integrations include Salesforce, ServiceNow, PeopleSoft, Palo Alto Networks, and Cisco. When a salesperson asks about their pipeline, for example, neuro-san routes the request to the Salesforce agent, which validates the user's role-based access before making any call, then returns the result within the same conversation thread. The employee does not need to switch context or log into a separate system. The entire interaction happens within 1Cognizant.

Q: What results has 1Cognizant delivered since rollout?

Within five months of deployment, the platform delivered a 50% improvement in operational efficiency and a 50% reduction in support ticket volume. Employee engagement on the intranet rose by 35%, with 10,000 average daily users. The platform has logged more than 10 million agent interactions with a 92% positive feedback rate, and response times came in at twice the speed of the previous experience.

Q: What were the biggest surprises or lessons learned?

The speed and scale of the support ticket reduction was the biggest surprise. The team had projected improvement, but a 50% drop within five months was beyond what anyone anticipated. The explanation, in hindsight, is that a large share of tickets were employees who could not find something or did not know which system owned a particular process. When you remove that navigation overhead entirely, the tickets stop being created.

The two lessons the team would carry into any future build: define security guardrails and the right data boundaries early, before any agents go live, and invest seriously in integration planning. Multi-agent architectures expose the inconsistencies across your existing tech stack in ways nothing else does, and the varied integration methods across enterprise platforms mean that partner collaboration is not optional, it is foundational.

Q: Is it expensive to run 200-plus agents?

Not in the way most people expect. The agents run in lightweight containers as an extension of the existing platform infrastructure, so the incremental costs are lower than the scale might suggest. More importantly, the cost savings generated by the reduction in support ticket volume far outpace any token or compute costs associated with running the agent network. The economics improve significantly once the system is operating at scale.

Q: What is next for 1Cognizant?

The team is integrating with Genesys to enable voice, so when employees call support, they will be speaking with an AI that already has their context and history and can act across the full agent network. Beyond voice, the focus is on expanding the agent landscape beyond 200 capabilities and working with process owners across the business to embed agents more deeply into specific workflows. The platform is also continuing to advance its responsible AI framework as the ecosystem grows.

What This Tells Us About Enterprise Agentification

The 1Cognizant story is not just about what Cognizant built. It is about what becomes possible when an organization stops treating AI tools as isolated point solutions and starts thinking about coordination as the core design problem. The outcomes did not come from any single agent doing something remarkable. They came from 200 agents working together through a single, coherent layer that employees never have to think about.

The broader lesson is architectural. Agentification works when security, integration, and governance are built into the foundation rather than layered on after the fact. It works when the system is designed to grow without requiring a rebuild every time a new capability is added. And it works when the people responsible for the underlying platforms are treated as partners in the build, not just stakeholders to be notified. None of that is especially complicated in principle. It just requires making those decisions early, before the pressure to ship takes over.

Build Your Own Agentic Enterprise With neuro-san

The framework behind 1Cognizant is neuro-san, Cognizant AI Lab's open-source multi-agent framework, available at github.com/cognizant-ai-lab/neuro-san. Because neuro-san is both LLM-agnostic and cloud-agnostic, organizations can adopt the same orchestration model without committing to a specific vendor stack or infrastructure setup. The agent network designer lets teams prototype and iterate quickly, and the framework handles routing, session management, sensitive data protection through sly_data, and horizontal scale as the system grows.

For engineering and architecture teams thinking seriously about moving from scattered AI tools to a functioning agentic system, neuro-san provides the foundation that 1Cognizant was built on. The architecture is documented, the framework is open, and the work done at Cognizant's scale offers a detailed reference point for what this looks like in a real production environment. The path is there.

Listen to the full podcast conversation: