The skeptics have a point. McKinsey’s State of AI 2025 showed that 62 percent of companies are experimenting with AI agents. Only 23 percent have started scaling them within a single business function. And a PwC survey of 4,454 executives found that 56 percent of CEOs still don’t see financial returns on their AI investments.
Numbers like these make IT leaders question the entire project. Understandably so.
But they’re asking the wrong question.
It’s not about the quality of the agents
The question isn’t “do AI agents work?” They do. In bounded contexts, they deliver measurable impact – shorter case handling times, lower error rates, more capacity per employee.
The question is: “Can our agents actually do their work?”
And the answer is often no – not because the technology is bad, but because the organization’s data and systems were built to keep each other out.
Vendors build walls, not bridges
CIO.com quotes Nancy Gohring from IDC: the APIs from one vendor’s customer service platform don’t work with another vendor’s e-commerce solution. Vendors want to keep customers inside their own ecosystem. That’s not a technical limitation – it’s a business decision.
Salesforce Agentforce, Microsoft Copilot, ServiceNow agents. Three platforms, three closed environments. Each agent is capable within its platform. None of them know what the others are doing.
Marc Benioff admitted it indirectly in an interview: “The speed of innovation is outstripping customer adoption. These customers have to go back and modify massive architectures they have and systems they’re running.”
That’s a polite way of putting it. The more direct version: we’re selling agents to organizations that haven’t yet built the foundation that makes agents useful.
A mistake we’ve made before
This isn’t the first time we’ve been here.
When ERP systems arrived in the 90s, many companies lifted their silos straight into a new system and called it transformation. The result was a new silo system that cost more to maintain. When cloud came along, many copied their on-premises architecture to Azure or AWS and wondered why it didn’t get cheaper.
New technology on top of old architecture doesn’t change anything fundamental.
AI agents are repeating the pattern. Companies implement agents on top of existing systems – the CRM that doesn’t talk to the ERP, which doesn’t talk to the ticketing system – and wonder why the agents create more confusion than clarity.
Of course they do. They’re fighting the same data problems as the employees.
What the CIO should be asking
Not: “Should we invest in AI agents?”
But: “Do we have the data structures that allow agents to work across our organization?”
That’s the question that reveals whether you’re ready. And the answer doesn’t require a new vendor – it requires looking inward: Where does our data live? Who owns it? Can one system read what another knows?
Organizations that answer “we don’t know” to those questions will get little out of AI agents regardless of which platform they choose.
What this means in practice
This isn’t a call to wait. It’s a call to build it right.
Companies that succeed with AI agents don’t start by choosing a vendor. They start by mapping data flows: what do the systems know, what do they share, and what’s missing for an agent to act across boundaries.
It’s boring work. It’s also the only kind that works.
The skeptics are right in their observation: many AI agents don’t deliver on their promise. They’re wrong in their diagnosis: the problem is rarely the technology. It’s the silos we’ve built up over decades and are now trying to layer AI on top of.
An agent is only as good as the data it has access to. And data locked in closed systems doesn’t help any agent.
