MongoDB (NASDAQ:MDB) said enterprises moving artificial intelligence applications and autonomous agents into production need unified data architectures that combine operational information, historical records, business rules and security controls.
Speaking during The 2026 Six Five AI Summit, Ashish Kumar, MongoDB’s senior vice president and technical fellow, said the focus of enterprise AI development has shifted from selecting foundation models to providing those models with trusted, current context.
According to Kumar, enterprises cannot effectively operate AI in production when vector data, operational data and security rules are maintained in separate systems. Instead, he said companies need a unified platform that can provide AI applications with data and governance in the same environment.
Context, memory and operational data
Kumar described foundation models as sophisticated reasoning engines that lack knowledge of an individual company’s operations without relevant context. He defined useful context as the combination of real-time operational signals, historical records and explicit business rules that an AI system can consume and trust.
Without that context, he said, models may make educated guesses based on raw text rather than decisions grounded in current business conditions. Kumar cited AT&T as an example, saying the telecommunications company combines real-time network signals and historical outage data to help AI determine where repair crews should be sent.
Kumar said that effort avoided 3.1 million unnecessary dispatches and saved $12 million in downtime. He emphasized that poor context can create material consequences when AI agents are used in physical operations or interact with billing systems.
The MongoDB executive also highlighted the importance of statefulness and memory for agentic applications. An agent that cannot retain information from earlier steps, understand user objectives or remember preferences over time is “really just an expensive chatbot,” he said.
For complex processes involving multiple steps and approvals, agents need to preserve their state over extended periods, Kumar said. He characterized that need as a core data challenge, particularly at large scale.
Kumar said an unnamed frontier AI laboratory moved more than 50 billion conversations from Postgres to MongoDB Atlas in four weeks to address conversational-state requirements. The system involved hundreds of petabytes of conversational state, with sub-millisecond reads and no reported downtime, according to Kumar.
Flexibility and performance
MongoDB customers are increasingly moving beyond experimental chatbot projects and integrating agentic workflows into mission-critical business processes, Kumar said. Based on his discussions with roughly 100 MongoDB customers over the prior nine months, he identified real-time performance and architectural flexibility as two key requirements.
AI agents need access to conversational history, enterprise context, semantic search and full-text search capabilities on current operational data, he said. Kumar added that an operational data layer must support spikes in reads and writes while maintaining low latency.
He cited Emergent Labs, a MongoDB customer, as an example of the value of flexible data modeling. Kumar said the company evaluated Postgres before selecting MongoDB Atlas because agents building applications require data models that can change frequently. He said Emergent Labs has powered about 2 million agentic applications on MongoDB.
Kumar also pointed to Macquarie, which built a retail business payments platform requiring continuous availability. He said MongoDB Atlas gave Macquarie portability across environments as it scaled to support millions of transactions, rather than tying the operation to a single cloud provider.
That flexibility is increasingly important because companies may need agents to operate near data held in other cloud environments or in on-premises, air-gapped networks, Kumar said. Organizations should be able to run the same agent functionality across those settings without rewriting it, he added.
Advice for enterprise AI deployments
Kumar advised businesses to avoid adding more specialized point products simply to launch AI pilots quickly. He said each additional database or single-purpose tool can add synchronization delays, security risks and technical debt.
Rather than replacing legacy databases, Kumar recommended building an operational data layer in front of existing systems. Such a platform can consolidate and enrich data with metadata, create vector embeddings where data resides, and provide AI agents with controlled access to context, security rules and guardrails, he said.
He added that companies can connect a unified data layer to AI models and enterprise data through MCP and other available standards. The objective, Kumar said, is to create a practical foundation for AI agents that require resiliency, high throughput, low latency and trusted information.
About MongoDB (NASDAQ:MDB)
MongoDB, Inc is a software company best known for developing MongoDB, a general-purpose, document-oriented database designed for modern application development. The company’s platform is built to support high-performance, scalable data storage and retrieval for use cases such as cloud-native applications, mobile backends, real-time analytics, and content management. MongoDB offers a mix of open-source software, commercial server distributions, and subscription-based services that include technical support, training and professional services.
The company traces its origins to 2007 when it was founded as 10gen by Dwight Merriman and Eliot Horowitz; it later adopted the MongoDB name and completed a public listing in 2017.
