AI Without Governance Is Just One More Integration Problem

We're living through the biggest technology wave of the past decade: artificial intelligence.

These days, it's almost impossible to have a conversation with executives without AI coming up. Even when the topic has nothing to do with AI, adding "AI" to the title of a presentation is often enough to grab everyone's attention.

Over the past few months, I've had the opportunity to talk with CIOs across different industries, and I've noticed a very clear pattern.

Every organization is taking a different approach. Some are building their own AI agents using LLMs. Others are investing in Agentforce, Amazon Bedrock, Google Vertex AI, Azure OpenAI, or other platforms.

But regardless of the technology they choose, one question comes up almost every time: How do we govern the agents — and the data they consume and produce?

In my view, that's the real enterprise AI challenge.

Building Agents Isn't the Hard Part Anymore

Creating an AI agent has become surprisingly straightforward.

The real complexity begins when your organization has dozens — or even hundreds — of agents running across different teams, platforms, and business processes.

Who decides which agent should handle a particular request?

How do you prevent multiple agents from performing the exact same task, consuming unnecessary tokens, compute resources, and AI credits?

How do you control which systems each agent can access — and what data it can read or update?

And perhaps most importantly, how do you monitor and audit every one of those decisions?

From what I've seen, many organizations are solving only part of the problem.

To protect production environments, they create ETL processes that copy corporate data into a separate environment dedicated to AI agents.

It's an understandable decision.

The business wants to move fast and deliver AI capabilities. The CIO has the responsibility to protect mission-critical systems.

The result is usually a parallel environment.

The downside is that this architecture breaks one of AI's greatest strengths: continuous data enrichment.

The insights generated by the agents remain isolated. They don't flow back into the enterprise Data Lake. They don't enrich operational systems. Too often, they end up sitting in temporary databases—or worse, in someone's spreadsheet.

In other words, AI keeps evolving while the organization's data foundation stands still.

Data Is the Foundation of AI

There's one idea I've been repeating quite often lately: There is no successful AI strategy without a strong data foundation.

The organizations creating a real competitive advantage aren't necessarily the ones with the most advanced language models.

They're the ones that invested early in integration, APIs, security, governance, and reusable data assets.

That's exactly where MuleSoft becomes strategically important.

For years, MuleSoft has helped organizations connect applications through API-led Connectivity, enabling reusable APIs and integrating on-premises systems, SaaS applications, databases, messaging platforms, and event-driven architectures using Anypoint Platform.

AI simply raises the stakes.

If we used to connect applications, now we also need to connect agents—even when they're built by different vendors and running on different platforms.

Why MuleSoft Agentic Fabric Caught My Attention

This is exactly why MuleSoft Agentic Fabric (MAF) stood out to me.

Rather than treating AI agents as isolated components, MAF introduces an orchestration and discovery layer that helps organizations manage agents as part of a unified ecosystem.

Think of it as an air traffic control tower.

When a request arrives, MAF understands the user's intent and discovers the agents capable of handling that request.

Those agents might be running in Agentforce, custom enterprise platforms, or other compatible AI environments.

Instead of broadcasting the same request to multiple models, which increases latency, token consumption, and operational costs, Agentic Fabric routes the request directly to the most appropriate agent.

At the same time, access to enterprise systems follows existing security, identity, and governance policies.

In practice, MAF acts as an Agent Gateway, providing centralized agent discovery, intelligent routing, observability, and governance.

The result is a much more secure, efficient, and scalable enterprise AI architecture.

A Simple Example

Imagine a customer visiting a bank's website and asking for a loan. Agentic Fabric recognizes the customer's intent. It discovers the agent specialized in credit analysis and routes the request accordingly.

That agent consumes APIs published through Anypoint Platform to retrieve customer information, credit history, available limits, and any other data required to evaluate the request.

Every interaction follows the organization's existing authentication, authorization, auditing, and security policies.

The agent delivers the answer to the customer without needing to know where each backend system lives or how those integrations were built.

That's the beauty of abstraction through APIs.

The Next Frontier

For years, our industry has talked about API governance. Now, we're starting to talk about agent governance.

Personally, I believe this will become one of the defining conversations in enterprise architecture over the next several years.

Building AI agents will continue to get easier.

The real competitive advantage will come from integrating them, governing them, monitoring them, and making them work together at enterprise scale.

That's why I see MuleSoft Agentic Fabric as a natural evolution of MuleSoft's integration strategy for the AI era.

At the end of the day, perhaps the most important question isn't: "How many AI agents does your company have?"

It's a much simpler one: "Who's governing them?"