The promise of agentic AI is powerful. Software agents can reason, act, and collaborate with minimal human intervention, and many executives view this shift as transformative for business operations.
Research from Bain & Company highlights that companies must rethink data and system foundations to extract real value from agentic AI, noting that the biggest gap between pilot and production success is not model capability but the underlying infrastructure. Leaders at HFS Research add that data accessibility, governance, and process readiness remain the largest blockers to enterprise-scale deployment.
Yet despite the enthusiasm, many agentic AI initiatives stall once they move from proof-of-concept to production. AI agents may use cutting-edge LLMs, but they struggle to operate reliably because they lack access to the right data for AI agents at the right moment.
It’s easy to assume that the LLM is the limiting factor. In practice, the main barrier is enabling AI agents to work with operational data that is accurate, current, and contextual.
Enterprise systems often reflect years of technical debt, acquisitions, and fragmented application landscapes, resulting in inconsistent and siloed data. Even the most powerful models cannot compensate for poor or inaccessible data for AI agents.
Some agentic AI use cases can succeed with static or non-operational information. For example, an agent summarizing a knowledge base does not require live system-of-record data.
However, the highest-value enterprise scenarios depend on transactional and continuously updated information. These include:
Customer service interactions that rely on the current customer’s state
Operational workflows that update inventory, trigger fulfillment, or coordinate supply chain activities
Fraud or risk monitoring that requires immediate detection of state changes in live data
These scenarios require AI agents to read from and write to core systems, and they depend on live, reliable data for AI agents that accurately reflects the present moment. When the data is delayed or inconsistent, the agent’s decisions and actions are compromised.
When data is not fit for agentic workflows, enterprises face challenges in multiple areas:
Accuracy: Agents produce incorrect outputs when they receive irrelevant, stale, or conflicting information
Speed: Slow data retrieval leads to delayed responses and disrupts the flow of work.
Cost control: Large language models charge based on data volume, so oversized inputs increase token and compute costs
Compliance: Providing agents with unnecessary or sensitive data increases regulatory and privacy risk
A successful agentic AI strategy must address all four dimensions. And each of them depends on delivering the right data for AI agents.
Traditional architectures were not designed for real-time autonomous agents:
Data lakes are optimized for analytics and often contain stale or aggregated data rather than live operational state
APIs and model context protocol servers return raw or fragmented information, forcing agents to reconstruct context, reconcile differences, or perform integration
Direct database access is brittle, risky, and difficult to govern at enterprise scale
A recent analysis by Kearney emphasizes that agentic architectures require a different underlying data foundation – one that is scalable, safe, and cost-efficient. Data structure and data access for AI agents are identified as the key inhibitors holding enterprises back.
A practical way to improve agentic outcomes is to give agents only the essential data they need for a task. This is the idea behind “minimum viable data”. Smaller, contextual, entity-centric datasets improve accuracy, reduce token usage, accelerate response times, and limit exposure of sensitive information.
Minimum Viable Data is foundational to providing precise, high-quality data for AI agents, while keeping cost and compliance risk under control.
To realize agentic AI at scale, enterprises need a data foundation built around entity-centric design. This means organizing information around core business entities such as a customer, order, or device, rather than around applications or systems.
An effective way to implement this approach is through data products. Each data product represents a specific business entity and continuously synchronizes all relevant information from systems of record into a single, reliable, and up-to-date view. For example, rather than having an AI agent assemble customer information from multiple applications, a customer data product provides one unified and current representation of that customer.
When these data products apply the principle of Minimum Viable Data, they expose only the essential information an AI agent needs for a given task. Providing smaller, contextual datasets improves accuracy, speeds up responses, reduces cost, and strengthens compliance.
Without entity-centric data products governed by Minimum Viable Data, agentic AI deployments often encounter rising operational cost, higher risk, and slower responses. Many remain stuck in pilot because the underlying data for AI agents is not prepared for autonomous agents.
Even with the right data foundation in place, AI agents still need a reliable way to access and act on that operational data. This is where data agents come in.
Data agents work alongside AI agents, inside the enterprise’s agentic framework and serve as the bridge between AI reasoning and real enterprise systems. They retrieve the Minimum Viable Data from the relevant data products, apply governance rules, and carry out operational actions when required.
Data agents ensure that AI agents work with the right data, in the right context, and in full compliance with enterprise policies.
Agentic AI succeeds AI agents get the data they need in a form they can reliably use. Minimum Viable Data (MVD) is the principle that guides this: agents receive only the essential, contextual information required for the task, which improves accuracy, reduces cost, and limits exposure. Entity-centric data products are the method that delivers MVD. By organizing data around key business entities such as a customer or order, each data product assembles all relevant information and then exposes only the minimum subset that is appropriate for the agent’s purpose. This creates a live, unified, and governed view that is naturally aligned with MVD.
Data agents complete the trio by operationalizing MVD. They retrieve the precise data that each task requires from the data products, enforce governance rules, and execute actions on enterprise systems. In other words, they ensure that MVD becomes the working reality of data for AI agents.
Together, MVD, data products, and data agents form the foundation that allows agentic AI to operate with trust, speed, and compliance at enterprise scale. Without them, agentic initiatives tend to remain experimental and fail to handle the complexity of real business operations.
In the next post, we’ll take a deeper look at the first component of this trio: Minimum Viable Data. We’ll examine why MVD is the cornerstone of accurate, efficient, and compliant data for AI agents, and how enterprises can apply it to unlock production-grade automation