In my last few articles, I have questioned the enormous amounts of money being spent on LLMs, data centers and the broader AI infrastructure buildout. I have also argued that the cost of inference—the tokens we consume—is barely scratching the surface of the real economic challenge facing CEOs and CFOs.

I want to shift the conversation now. The question is no longer simply “Are we spending too much on AI?”

It is “How do we actually make AI deliver a measurable return?”

Over the last couple of years, companies have invested aggressively in AI. Customer service has an AI tool. Sales has one. Finance has one. Operations has another. Employees have copilots, and pilots are underway across the organization.

And many of these tools actually work. They summarize documents, identify anomalies, answer customer questions, help employees work faster and flag problems that might otherwise have been missed.

Yet executives continue to ask a surprisingly simple question:

Where is the ROI?

Perhaps part of the problem isn't that the AI isn't working. It is that we are evaluating each application within the narrow function it was designed to serve, rather than asking whether AI is making the business perform better.

Making our silos smarter

Think about how most companies operate. Sales, finance, operations, customer service and supply chain each have their own systems and data. Enterprises have spent decades trying to break down these organizational and information silos.

Now we are putting AI on top of them.

We may be using AI to make our silos smarter rather than making our enterprises intelligent.

Consider a simple example. Suppose a company's customer-service AI detects an increase in cancellations and determines that customers are unhappy because deliveries are late. The tool has done exactly what it was supposed to do. Management gets alerted and operations addresses the delivery problem.

But what actually caused it?

Perhaps marketing launched a promotion that produced an unexpected increase in demand. Inventory fell below required levels. Procurement substituted a supplier. Production slowed. Shipments went out late. Customers complained and eventually began cancelling.

The customer-service AI sees the end of the problem. Sales knows about the promotion. Supply chain knows about the inventory shortage. Procurement knows about the supplier change. Finance sees deteriorating margins.

No system sees the entire chain.

From detection to prevention

AI is becoming very good at finding problems. But finding a downstream problem faster doesn't necessarily prevent the upstream decisions that created it.

It is useful for AI to tell management that cancellations increased because deliveries were late.

It is considerably more valuable if management can ask why deliveries became late, what is likely to happen if the promotion is repeated, and what should change before it happens again.

That is the progression that matters:

Detection → Explanation → Prediction → Prevention

And getting there requires something many enterprise AI initiatives still don't have: Enterprise Context.

An AI system doesn't simply need access to more data. It needs to understand how the pieces of the business relate to each other:

Customer → Order → Product → Inventory → Supplier → Production → Shipment → Support → Payment → Margin

It needs to understand which systems are authoritative, what business terms mean, how processes work, what happened historically and what information someone is permitted to access.

Creating that context may require an intelligence layer connecting the company's existing systems, data, documents, processes and institutional knowledge.

This isn't about replacing specialized AI applications. A specialized product may be exceptionally good at the problem it was designed to solve.

The problem may not be that your AI tools aren't working. The problem may be that they're working alone.

Changing the economics

There is another reason this matters.

Today, every new AI initiative can become another integration project: find the data, connect the systems, establish security, define permissions, interpret business terminology, build governance and integrate the workflow. Then do much of it again for the next application.

That gets expensive quickly.

A reusable intelligence layer could change those economics. The first use case still has to justify itself, but the second, fifth and twentieth applications can increasingly reuse the same enterprise context.

The economic advantage may not be the first AI use case. It may be the declining marginal cost of every use case that follows.

The next phase of enterprise AI

The first phase of enterprise AI was understandably about experimentation. Find use cases, run pilots, deploy copilots, buy specialized applications and learn what works.

That experimentation was necessary. But perhaps the next phase is different.

The enterprise already contains enormous intelligence in its people, systems, data, processes and institutional knowledge. The problem is that much of it remains fragmented.

The CEO doesn't need ten AI systems independently reporting that ten different things happened. She needs to understand why they happened, how they are connected, what is likely to happen next and what the organization should do about it.

The next phase of enterprise AI may not be about adding more AI tools. It may be about giving the AI we already have the context to understand the business.

Which leaves a much more interesting question than simply asking about AI ROI:

Your AI knows your data. Does it understand your business?