Context: The Real Reason Agentic AI Fails In The Enterprise

Michel Tricot is Cofounder and CEO of Airbyte.

gettyAI agents have quickly moved from novelty to priority, with McKinsey finding that “88 percent of organizations are now experimenting with AI.” Agents are automating workflows, assisting employees and promising faster, more efficient operations.

But there is a growing disconnect between what AI agents can do in demos and what they actually deliver in production, with the same McKinsey research finding that 81% of companies “do not report any meaningful bottom-line gains” from AI. ​

When asked to operate on real business problems, agents often fall short with incomplete answers and failure to take meaningful action.

I’ve worked in data infrastructure technology for a long time. Many of the traditional problems I have worked on for customers, such as fragmented data and unreliable integrations, are compounding in the era of agents. We see our customers increasingly adding AI to their workflows, but the root problems remain the same: fragmented data and unreliable integrations.​​

Most AI strategies today focus on model selection: which LLM to use, how to fine-tune it, how to prompt it and optimize each model for the task at hand.

But in practice, model performance is rarely the limiting factor. The real issue is that AI agents don’t have reliable, contextualized access to internal, constantly changing business data. Even Anthropic acknowledges this in their report on AI agents. “The biggest barriers to enterprise AI value aren’t model capabilities or costs—they’re organizational readiness, especially around data accessibility and context aggregation.”​

​There is a big shift happening in how companies use data that has been brought about by AI.

In the past and still today, data systems were built for people to access to make informed decisions. Dashboards and reports provided people with the raw information that they needed, and people took action and made decisions based on that data.​

Today, organizations must shape and present the right data to agents to enable them to make decisions. A very important part of shaping that data is providing context around it to help the agent technology better understand and deliver an answer much like a person would. ​

For example, say the question is: “Are our customers unhappy with our new pricing model?” An analyst would instantly know which data sources to query and what signals to look for. An agent has to understand what “being unhappy” looks like, what signals indicate pricing changes and where to look for all of that. You can prompt your AI assistant directly and specifically to do that, but autonomous agents have to act without constant prompting.

​That means that you need more than just raw information. There is a requirement to provide context around the data. That means:​

• Making data easy to access and interpret, not locked behind complex interfaces.

These concepts are not new, but the stakes are different. The risk with fragmented data historically was slower decision-making. Today, it can cause agents to make costly mistakes in real time.​

​Many organizations are layering AI on top of existing systems without rethinking how those systems provide data. The result is predictable: agents that look capable but can’t deliver meaningful outcomes.​

AI is not just another application layer. It is a new way of interacting with data that requires not only access to the right data but also context around that data to understand and interpret the data, so that AI output is optimized. And if the underlying data systems aren’t designed to support that, the results will fall short.​

​To make AI agents work in production, organizations need to shift their focus from models to infrastructure, not only in a technical sense but also in a strategic one. Based on my experience, I recommend a three-pronged approach:

If your systems can’t access or understand relevant, real-time data, your AI initiatives will stall.

​Start by mapping the systems your agents will rely on, such as your CRM, support desk, billing platform and internal docs. For each, decide which data needs to be accessible in real time for agents at the moment they act, then separate them from data sources that can be synced in a warehouse.

Next, reconcile records that describe the same customer or account across those systems. Agents should access the data via one unified Model Context Protocol (MCP) gateway, for security and scalability.​

Agentic AI needs to do more than answer questions. It needs to safely and reliably take action, such as updating systems, triggering workflows or guiding decisions. ​

That means defining which actions an agent is allowed to take in each system. Permissions should be set by individual actions and action types, at the finest granularity, with higher-risk actions subject to human approval.

All actions should also be logged, so you can audit and reverse what an agent did.​

Early on, aim for visibility into what your agents do and how they operate. This will teach you how agents actually operate and inform your governance strategy.

Once you understand your agentic workflows, you can establish effective governance policies that meet your security requirements.​

We are still in the nascent stages of adopting AI agents. Models will continue to improve. Capabilities will expand. But without the right data behind them, AI agents will remain stuck in demo mode.​

AI doesn’t fail because it isn’t powerful enough. It fails because it doesn’t know enough and doesn’t understand the context of your business. Fix that to build the right data foundation, and you’ll see that everything starts to work.​​​

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?