Enterprise leaders are not short on ambition when it comes to AI, they’re short on results.
By the end of 2024, enterprises had already spent more than $250 billion on AI initiatives, according to Stanford’s 2025 AI Index. Yet despite that investment, many organizations are struggling to translate experimentation into sustained business value. As the New York Times recently put it, “Billions are pouring into AI. It has yet to pay off.”
This growing gap between spending and outcomes is what MIT Sloan Management Review describes as AI strategy heartburn. AI initiatives start with excitement, stall in pilots, and ultimately leave leaders questioning what went wrong.
One emerging culprit is how enterprises are adopting Model Context Protocol, or MCP.
The appeal of MCP and why it gained traction so quickly
MCP has gained attention because it standardizes a way for AI agents to access enterprise systems, retrieve context, and take action without custom integrations for every application.
In theory, MCP allows agents to plug into tools the way peripherals plug into a computer. The appeal is obvious. Faster development, lower integration effort, and a more flexible way to connect AI models to business systems.
For experimentation, this approach works. Teams can quickly wire an agent to a handful of APIs and demonstrate impressive demos. But as MIT Sloan points out, those demos rarely survive contact with real enterprise environments.
Where raw MCP begins to break down
The problem is not MCP as a concept. The problem is relying on raw MCP as if connectivity alone is sufficient for enterprise AI.
Most MCP implementations today focus on exposing tools, not on managing processes. They allow agents to call APIs, but they do not encode how work actually flows across an organization. Enterprise processes such as order to cash, employee onboarding, or financial close involve sequencing, dependencies, exception handling, approvals, and compliance checks. MCP does not provide that structure.
As a result, agents built on raw MCP tend to automate isolated steps rather than complete outcomes. This creates fragility instead of leverage.
MIT Sloan also highlights the security implications. MCP servers often run untrusted code with broad access to enterprise systems. Without identity inheritance, role-based permissions, or policy enforcement, organizations are effectively granting agents more authority than many human users.
The risk compounds when agents operate autonomously. Enterprises must be able to answer who initiated an action, what data was used, why a decision was made, and what happened next. Raw MCP does not provide auditability, traceability, or governance out of the box.
These gaps explain why many MCP pilots fail to scale. What works in a sandbox becomes unacceptable in production.
Why tools matter more than connectivity
A key insight echoed in the MIT Sloan article is that not all tools are created equal.
Anthropic recently made this point explicitly, arguing that purpose-built tools are essential for reliable agents. Tools need to represent meaningful business actions, not just low-level API calls. A tool like process_refund or provision_user carries intent, constraints, and validation. A generic API call does not.
When agents interact with raw APIs, they are forced to reason about implementation details rather than outcomes. This increases the likelihood of errors, hallucinations, and unintended consequences. Enterprises cannot afford that level of uncertainty in core systems.
In practice, this means agents must operate through governed, reusable enterprise skills rather than direct system access. MCP can transport context, but it does not define or enforce how actions should be executed.
The missing enterprise foundation
AI failure is rarely about models, itt is about missing operational foundations.
Enterprises that succeed with AI share several characteristics. They treat AI as part of their operating model, not as an overlay. They invest in orchestration, not just connectivity. They enforce identity, security, and governance at every step. And they design AI around end-to-end processes rather than individual tasks.
Without these foundations, AI initiatives generate friction instead of returns. This is why so many organizations find themselves spending more while trusting less. The heartburn comes from realizing that faster experimentation does not equal enterprise readiness.
MCP as a component, not a strategy
Seen in this light, MCP is not the problem, but treating MCP as a complete strategy is.
In an enterprise context, MCP must sit within a broader orchestration layer that provides process context, identity-aware access, auditability, and resilience. Only then can AI agents move from impressive demos to dependable operators. This is the difference between connecting AI to systems and enabling AI to run the business.
How Workato addresses the gap
Workato was built around the reality that enterprises run on processes, not APIs.
The platform provides enterprise-grade orchestration across applications, data, and workflows, giving AI agents full operational context. Identity and access controls ensure agents act within defined permissions. Built-in governance and observability provide traceability for every action. Resilience features ensure that failures are handled safely and predictably.
In this model, MCP becomes a bridge, not a bypass. Agents interact with enterprise systems through governed skills and orchestrated workflows rather than raw endpoints.
This approach aligns with what MIT Sloan describes as the path out of AI strategy heartburn. Not more tools, but better foundations.
Putting MCP in the Right Place
Enterprises have already invested hundreds of billions of dollars in AI, yet many initiatives remain stuck in pilots with little operational impact. Model Context Protocol has helped accelerate experimentation by making it easier for AI agents to connect to enterprise systems, but connectivity alone has proven insufficient.
As the MIT Sloan analysis shows, raw MCP lacks the process context, governance, identity controls, and resilience required for enterprise-scale AI. Agents built on raw APIs can automate isolated tasks, but they struggle to operate safely across real business workflows.
The lesson is not to abandon MCP, but to treat it as a component rather than a strategy. Successful enterprises place MCP inside a broader orchestration foundation that enforces enterprise skills, security, and end-to-end process discipline.
Workato provides that foundation, enabling AI agents to move from demos to dependable operators and turning AI investment into durable business outcomes.
