Generative AI answers questions. Agentic AI acts. That one shift is rewriting what enterprises need from the systems that hold their data. According to IDC's April 2026 Worldwide Data Management Survey (n=715), 30% of organizations are already running agentic AI in production, and another 50% are investing behind an established spending plan. Four in five enterprises have moved past whether agents belong in their operations. Now they're working through the harder question: how do you run them safely?
That question lands squarely on the data platform. Agents don't reason in the abstract. They act on live business data, and the speed and trustworthiness of that data set the ceiling on what an agent can safely do. A platform built to analyze what already happened is not the same as one built to run what happens next. In the agentic era, the database moves from a passive store of record to the control point, the place where real-time context gets served, relevant information gets retrieved, and policy gets enforced the moment an agent acts. Get that wrong, and a single bad read doesn't stay a bad read. It becomes a wrong action at machine speed, and it repeats.
In this IDC Spotlight, Devin Pratt, Research Director for Database Systems for Enterprise Intelligence, AI, Automation, and Data at IDC, lays out four requirements enterprises need to treat as a single platform decision, not four separate purchases. He also examines the market forces behind them and looks at how one vendor, MongoDB, is responding.
The first requirement is a real-time operational foundation that fits AI-shaped data. AI produces and consumes information that's unstructured, fast-changing and high-volume: documents, embeddings, event streams and context that rarely fits neatly into rows and columns. Agents need to read and write that data at low latency, so the foundation must match the data's natural shape instead of just storing it for later analysis. IDC's survey backs this up: 64.6% of use cases now require real-time or near-real-time delivery, and for streaming scenarios specifically, 91.2% require end-to-end latency below 200ms. IDC also forecasts that non-schematic and data lake systems will outgrow the overall database market, even as relational systems, still the largest segment at $61.0 billion and 56.6% of 2025 revenue, keep growing but lose share.
The second requirement is elastic scale for unpredictable, agent-driven demand. Agent traffic doesn't behave like human traffic. It's bursty, hard to forecast, and can spike as thousands of agents request context at once. Traditional architectures handle that volatility by over-provisioning, or worse, by hitting a hard scaling ceiling that forces a disruptive rebuild right when a workload is succeeding. Enterprises want the opposite: room to grow on one foundation without re-architecting. That preference shows up as a structural move toward consolidating on a core platform, with public cloud now 64.6% of DBMS revenue in 2025 and rising, best suited to absorb that elastic demand.
The third requirement is accurate, always-current retrieval and context. An agent is only as good as the context it can retrieve. Feed it the wrong information, and it will act confidently and wrongly. Roughly 75% of buyers favor integrated vector databases for storing and querying embeddings over standalone options, and nearly half (47%) call a consistent semantic layer critical to safe, trustworthy AI.
The fourth requirement, a governed runtime to build and run agents in production, is often what separates a promising pilot from a real deployment. Trusting an agent with real business actions takes durable memory to retain context across interactions, identity-bound policy that can't be bypassed, and observability into what the agent did and why. Without those three things, governance teams have no basis to sign off, and agents stall out in permanent pilot mode. IDC predicts that by 2027, 80% of agentic AI use cases will require real-time, contextual, and ubiquitous data access, pushing organizations toward federated access with centralized governance instead of centralized gatekeeping.
Knowing the requirements isn't the same as meeting them. The Spotlight names three practical concerns that consistently slow enterprises down: staying deployment-flexible and avoiding lock-in as the model and framework landscape shifts monthly; governing and auditing autonomous action, since access control stops being a back-office concern the moment software can act on its own; and controlling the cost of bursty, retrieval-heavy agent workloads.
Against that backdrop, the report looks at how MongoDB is working to meet all four requirements on a single platform. IDC's 2025 market share data places MongoDB as the third-largest vendor in the non-schematic DBMS submarket, at roughly 10.9%. The Spotlight maps MongoDB's recent releases to each requirement: MongoDB 9.0, a foundation release MongoDB reports delivers up to roughly twice the throughput on large instances, addresses the real-time foundation; Atlas Infinite, which scales compute and storage independently and is entering public preview, addresses elastic scale, with MongoDB citing a customer processing roughly 50% more transactions per second with no application changes; Voyage AI, acquired in February 2025, addresses retrieval with embedding and reranking models plus automated, continuously synchronized embeddings; and the MongoDB Atlas Agent Engine, built to be LLM-agnostic, deployment-agnostic, and governance-first, addresses the need for a governed runtime with built-in memory, agent identity, policy enforcement, and observability.
The report doesn't shy away from open questions, and neither should we. Atlas Infinite and the Atlas Agent Engine are both early in their release cycles, so broad multicloud availability and a longer production track record will take time to build up, and the agentic platform category itself is crowded and moving fast. What we'd add: the advantage of doing all four on one platform is that teams aren't stitching together point solutions under pressure while agents are already in production. That's the harder problem most vendors haven't solved yet.
IDC's conclusion is direct: real-time scale, accurate retrieval, and built-in governance are no longer differentiators. They're the price of entry. IDC predicts that by 2029, 60% of enterprise data platforms will unify transactional and analytical workloads in hybrid processing architectures to power agentic AI. The database is no longer where data waits. It's where agents act. Enterprises that treat it that way will move from pilots to production. Those that don't will watch the gap widen.
Download the IDC Spotlight for the full breakdown of the four requirements for agentic AI at scale, the market forces driving them, and IDC's independent assessment of how MongoDB's platform measures up against each one.
IDC Spotlight
