Why MongoDB Says AI Agents Are Only as Good as Their Database

The AI industry is still obsessed with LLMs. Enterprises, however, are beginning to worry about something else. As AI agents move into production, databases are now increasingly determining whether these systems can deliver accurate, reliable, and low-latency responses.

In an exclusive interview with AIM at MongoDB.local Bengaluru 2026, a conference hosted by MongoDB for developers, IT executives, and founders, Sachin Chawla, VP of India, ASEAN and Japan at MongoDB, says enterprises have largely moved beyond debating model quality and are instead prioritising reliable data, search, memory, and speed to production.

“It’s the data foundation that matters,” he adds, arguing that enterprises can no longer rely on models alone to build production-ready AI agents.

Unlike traditional enterprise software, AI-native applications evolve continuously, making rigid database schemas impractical. “You need a database which is flexible. You don’t want to get bogged down with schema design because these applications are changing every day,” Chawla notes.

He believes AI agents are a natural fit for MongoDB because both operate around JSON-based structures. “Agents are JSON, and our documents are JSON. It’s a very natural fit,” he says.

Better Retrieval Tools

At MongoDB.local Bengaluru, the document-oriented NoSQL database unveiled a set of AI retrieval capabilities to address two of the biggest obstacles to enterprise AI deployments—retrieval accuracy and compliance. 

The company introduced Voyage Context 4, a new embedding model, alongside Hybrid Search and Native Reranking to improve the quality of retrieval-augmented generation (RAG) applications.

MongoDB Atlas is a fully managed, multi-cloud database service that hosts MongoDB in the cloud across platforms. Native Reranking for MongoDB Atlas reranks search results within the database, improving retrieval quality by up to 30% without requiring external APIs or additional infrastructure, according to the company.

Meanwhile, Hybrid Search combines full-text and vector search in a single database query and automatically keeps embeddings in sync with operational data, allowing AI applications to retrieve the latest information without relying on separate search systems.

“The biggest barrier to enterprise AI in production and at scale isn’t the LLM. It’s memory, retrieval, accuracy, and compliance," says Benjamin Cefalo, Chief Product Officer, Core Products, MongoDB.

He adds that most enterprises are usually held back by infrastructure that wasn’t designed to provide AI with trusted access to enterprise data. Cefalo says bolting on more systems to solve these problems only creates more vendors, latency, and points of failure.

While concerns around AI hallucinations dominated enterprise discussions last year, Chawla believes that phase is largely over. “Accuracy is table stakes. It has to be accurate. Otherwise I can’t put my agent into production."

Chawla says enterprise AI conversations have evolved over the last year. “The conversation has moved from accuracy to latency,” he says. Enterprises are now looking for platforms that combine vector search, full-text search, embeddings, and re-embedding into a single foundation capable of supporting production-grade AI applications.

Competition With Hyperscalers

MongoDB competes with database services offered by hyperscalers such as Amazon Web Services (AWS), Microsoft Azure, and Google Cloud. 

AWS offers services including Amazon DocumentDB and Amazon DynamoDB, Microsoft competes with Azure Cosmos DB, while Google Cloud provides document and NoSQL offerings such as Firestore, Bigtable, and AlloyDB as part of its broader cloud platform.

However, Chawla argues that enterprises are moving away from maintaining multiple specialised databases. “Having four databases in an environment today, when the business is demanding you to move fast, is really difficult,” he says.

Instead of stitching together multiple services, customers increasingly want a single platform that handles different workloads.

Chawla says if a single database can deliver 80% of the functionality enterprises need, there is little reason to manage separate platforms such as Pinecone and Elasticsearch. Consolidating those capabilities, he argues, takes away much of the operational heavy lifting.

He believes the industry is gradually moving away from purpose-built databases. “The era of purpose-built databases is slowly sunsetting because it’s all about how fast you can innovate.”

Multi-Cloud Strategy

Chawla says large enterprises are increasingly rejecting vendor lock-in—not only for cloud infrastructure but also for AI models.

MongoDB’s strategy, he says, is to remain infrastructure-agnostic. “We’ve always said we're going to give customers the flexibility to use different clouds, different models, and also deploy on-prem.”

He adds that several regulated industries, particularly banks, continue to prefer on-premises deployments because they do not want sensitive data leaving their environments.

Open-Source Models and Tokenmaxxing 

With enterprises increasingly worried about rising inference bills and token consumption, Chawla believes organisations will gradually stop optimising purely for token usage.

“I think the days of tokenmaxxing will slowly sunset,” he says.

Instead, companies will evaluate the entire lifecycle of AI applications—from development to production—rather than focusing only on token counts.

“You want to look at the full lifecycle of what you’re building and measure the output with the right metrics versus just measuring tokens.”

According to him, enterprises are increasingly using multiple foundation models—and in some cases smaller language models—depending on performance, latency and cost. “Open-source models are getting better and better. It will be a mix.”

India Market 

MongoDB says it serves half of India’s top 100 listed companies by market capitalisation and more than 50 unicorn startups, including Emergent, PhysicsWallah, Leena AI, and Zomato. The company has over 700 employees in India across Bengaluru, Mumbai, and Gurugram, including engineering, consulting, finance, human resources, and go-to-market teams.

Moreover, the company plans to upskill 2 million builders in India by 2030 through an expansion of its MongoDB for Academia programme.

While Chawla declines to share revenue or market share figures for India, he call the country one of MongoDB’s most important growth markets, driven by its developer ecosystem and startup activity.