We're Spending Billions on AI. Does Anyone Actually Know What We're Getting Back?

Everyone seems to agree that artificial intelligence will transform business. Companies are investing billions, employees are getting faster, and AI agents are finding their way into almost every department… but there's a question that keeps getting lost in all the excitement; Are we actually building better businesses, or are we just getting better at doing the same things faster?

Imagine you're the CEO of a company with 5,000 employees, over the past two years, you've invested millions of dollars in AI, your engineering team is using coding assistants, customer support has deployed AI agents, marketing is pumping out content faster than ever, and your finance department has automated reporting that previously took days.

Every department is reporting improvements, productivity is up, tasks are being completed faster, and your employees seem genuinely excited about what these new tools can do… You've attended conferences, listened to experts, hired consultants, and invested in what everyone keeps telling you is the most important technological transformation of our generation.

You walk into your next board meeting feeling pretty good about the progress you've made, then someone asks a simple question "How much money have we actually saved?" You turn to your CFO, they look at the CTO, the CTO starts explaining infrastructure costs, model consumption, and implementation expenses, then someone mentions that productivity has improved by 30% in certain departments, another executive points out that the company has launched 47 AI initiatives in the past year.

Impressive numbers! Except nobody has answered the question… 

Now this is an imaginary scenario, but I suspect it's not particularly difficult for many executives to relate to it.

I recently came across a Bain & Company report that made me think about this exact situation, Bain surveyed 951 companies and found that while many were targeting cost reductions of 11% to 20%, nearly 40% of those measuring outcomes achieved savings of less than 10%. And yet, 90% of the surveyed companies planned to increase their AI budgets.

There's something almost comical about that… we're spending enormous amounts of money trying to make our organizations more efficient, but when the results fall short, our response is to spend even more.

Of course, I'm not suggesting companies should stop investing in AI, quite the opposite, I believe we're still in the earliest stages of what this technology will eventually become, but I do wonder whether we've become so obsessed with what AI can do that we've neglected to ask what it actually needs to succeed.

And that brings me to something I've been thinking about for a while, for decades, we've been building companies around departments, sales has its systems, engineering has its tools, customer support has its platforms, and somewhere in the middle, product managers are desperately trying to connect all the pieces… We've accepted this fragmentation as a normal part of doing business, people spend their days in meetings, sending messages, updating documents, and asking colleagues questions that someone else in the organization has probably already answered. Now we're introducing AI agents into this environment, and we're expecting them to magically make everything work better. 🤦🏻‍♂️

If your sales team doesn't know what your support team knows, and your engineering team doesn't understand what customers are struggling with, giving each department its own AI agent doesn't solve that problem, you've simply introduced another participant into an already fragmented system, in fact, you might have made the problem worse.

Because unlike employees, who might eventually bump into one another at lunch or discover a contradiction during a meeting, AI agents can operate independently, continuously, and at a scale that makes those little misunderstandings much more consequential.

And that's where I think the conversation about enterprise AI needs to change.

Let's stay with our fictional company for a moment, because I think it's worth looking at what happens after that board meeting.

The CEO asks the leadership team to come back in 30 days with a clearer picture of the company's AI investments, nothing unreasonable, just a breakdown of what they're spending, what they're saving, and where the biggest opportunities are.

The CTO begins collecting infrastructure expenses, finance starts comparing departmental budgets, HR looks at employee productivity surveys, department heads prepare presentations showing how much faster their teams have become.

Meanwhile, the head of customer support reports that AI has reduced the average time required to respond to customers, the product team reports that it can now produce PRDs in a fraction of the time, engineering reports that developers are completing certain tasks faster with coding assistants. By the end of the month, everyone has a beautiful presentation… there's just one problem… They're all measuring different things.

Customer support measures response times, engineering measures development velocity, finance measures operating expenses, and product management measures how quickly requirements move through the pipeline, each department can demonstrate improvements, but nobody can confidently explain what those improvements mean for the business as a whole.

And that's before anyone starts asking whether the work being completed faster was the right work to begin with.

Imagine that customer support is now handling complaints twice as quickly, but customers keep complaining about the same confusing feature, and the product team, meanwhile, has accelerated its development process and is shipping more features than ever, but unfortunately, it still hasn't discovered that the existing feature is frustrating customers enough to make some of them leave. Both teams are more productive, but neither has solved the underlying problem. 

The company has succeeded in making two disconnected processes faster without improving the connection between them.

This is where I think the distinction between productivity and intelligence becomes important, productivity is about how efficiently we perform work, intelligence, at least in an organizational sense, is also about understanding what work matters, why it matters, and how different pieces of information influence one another.

An organization can become more productive without becoming any better at understanding itself, and if that sounds like a philosophical distinction, consider what happens when the company starts automating decisions rather than simply helping employees complete tasks.

The customer support agent notices that hundreds of customers are struggling with the same feature, it categorizes their complaints, responds politely, and closes the conversations, the product management agent is busy analyzing usage data and generating specifications for upcoming features, the engineering agent is reviewing tickets and suggesting implementation approaches, and each agent is doing exactly what it was designed to do.

But unless someone has built a way for those systems to share and reconcile what they're learning, the support agent's understanding of customer frustration may never meaningfully influence the product agent's recommendations.

The information exists, but the intelligence to connect it does't seem to.

And here's the part that worries me… As we give AI agents more autonomy, we may also be giving them more opportunities to act on incomplete information, than can lead to some serious distasters. A human employee might misunderstand something and make a bad recommendation, an AI agent might misunderstand the same thing, create a series of tasks, trigger other agents, and initiate work before anyone notices the original assumption was wrong.

I'm not suggesting this happens in every deployment, well-designed systems have controls, permissions, reviews, and safeguards. But those controls don't automatically give an agent a complete understanding of the organization, they help manage what happens when that understanding is incomplete.

There's a difference.

McKinsey explored a related problem in its June 2026 report, AI Data Readiness: The Key to Scaling Impact. According to the report, only 7% of companies had fully scaled AI across their organizations, the researchers pointed to a familiar obstacle, companies possess enormous amounts of information, but much of it isn't organized, governed, or connected in ways that make it reliably usable by AI.

That finding didn't surprise me nearly as much as it probably should have.

Anyone who has worked inside a reasonably large organization knows that information is rarely as tidy as the software demonstrations suggest… Important decisions are buried in Slack conversations, product requirements are scattered across documents, customer feedback lives in support platforms, someone updates a policy but forgets to update the knowledge base, and a critical decision gets made in a meeting, and three months later, nobody remembers exactly why. We've learned to live with this, humans are remarkably good at compensating for imperfect systems, we ask colleagues for clarification, question information that looks suspicious, and use our experience to fill in the gaps… AI can do some of those things too, but only when its architecture and tools give it the means to do so.

Take something as ordinary as a refund policy, your company has three documents describing it, one says customers have 30 days to request a refund, another says 14 days, and a third says refunds aren't available for a particular product category.

An AI assistant can retrieve all three documents in seconds, it can summarize them beautifully and explain the differences with impressive confidence.

But which policy is actually in effect?

Maybe one document is outdated, maybe another applies only to customers in Europe, maybe the third reflects a decision made last week that hasn't been formally approved.

The correct answer depends on details that may not be obvious from the documents themselves. It depends on their history, their authority, their relationship to other policies, and sometimes on a human decision that hasn't been recorded anywhere useful.

This is one reason I find the current obsession with giving AI access to more data somewhat incomplete, access is necessary, of course, but access alone doesn't guarantee understanding, it's a bit like giving someone the keys to the world's largest library and assuming they'll automatically know which books are correct, which are outdated, and which contradict one another.

The library is valuable, so is the ability to search it, but neither replaces the work of establishing what can actually be trusted… And as it turns out, that work has financial consequences too.

Consider what happens when that fictional company begins deploying hundreds of AI agents, one is analyzing customer complaints, another is preparing a quarterly product strategy, and a third is reviewing accounts that might be at risk of cancellation, each needs to understand something about the same customers, the same products, and the same business, but depending on how the systems are built, each might independently search through documents, retrieve conversations, and reconstruct information that another agent has already processed.

It's not difficult to see how this could become catastrophic and incredibly expensive.

The more information an AI system needs to process, and the more often it needs to process it, the more computational resources it generally consumes, the exact costs depend on the model, the provider, the type of processing, and whether techniques such as caching are used.

Now consider an organization running hundreds of agents, each performing dozens or hundreds of tasks a day, some of those tasks require completely different information, but others overlap considerably, a customer retention agent might need to understand why customers are leaving. a product management agent might need to understand which product issues are driving dissatisfaction, a customer support agent might need to understand which problems have already been identified and what the company intends to do about them.

They're approaching the same information from different directions, and yet without a way to share and maintain that understanding, they may end up doing much of the same investigative work repeatedly.

It reminds me of something we've all experienced at work, you join a meeting, spend twenty minutes explaining a problem, agree on a course of action, and leave feeling that everyone finally understands what's happening, and a week later, you're invited to another meeting with a slightly different group of people, and you spend another twenty minutes explaining the exact same problem.

Nobody is necessarily doing anything wrong, the people in the second meeting genuinely need that information, but the organization has failed to preserve and distribute what it already learned, now imagine paying for every minute of that repeated explanation, that's not a perfect analogy for how AI systems operate, but it illustrates why the ability to reuse organizational knowledge could become increasingly important.

Interestingly, the financial implications of reusable infrastructure are already being explored, in its 2026 data readiness report, McKinsey presents an illustrative financial-services scenario in which reusable data foundations could help avoid approximately $10 to $20 million in costs as additional AI use cases are deployed, those numbers are estimates for a particular scenario, not a promise of what every organization can save, still, the underlying idea is worth considering… If you're going to build something that hundreds of applications need, perhaps it makes more sense to build and maintain it as shared infrastructure rather than reconstructing it independently for every application. Of course, there's a catch, maintaining shared knowledge has costs of its own, information changes, policies evolve, and yesterday's correct answer might be wrong tomorrow, and some information is confidential, and not every agent should have access to it, certain conclusions require human judgment, and sometimes two sources genuinely disagree, a system that simply stores a supposedly correct answer and distributes it everywhere could become just as dangerous as one that repeatedly retrieves unreliable information.

The challenge, then, isn't merely to remember things. It's to maintain an understanding of what the organization knows, where that knowledge came from, whether it's still valid, and who is allowed to use it.

I find that distinction particularly interesting because it changes how we think about enterprise AI infrastructure, instead of treating every AI application as a separate system that needs to discover the organization for itself, we can begin thinking about organizational knowledge as something that is continuously maintained and made available to different applications when they need it.

That doesn't mean replacing databases, search engines, knowledge graphs, or the many other technologies companies already depend on, those systems remain essential, nor does it mean every company needs to adopt an entirely new architecture tomorrow, but it does mean that as AI becomes more autonomous, the infrastructure connecting information to decisions will become just as important if not more important then the models making those decisions.

And perhaps more importantly, it suggests that we're measuring the wrong things when we evaluate some of our AI investments.

Let's return to our CEO one last time.

Thirty days have passed since the board meeting, the leadership team has collected its reports, calculated infrastructure expenses, and documented productivity improvements, they've identified departments where AI is saving time and others where implementation costs are higher than expected… It's a useful exercise, and the company now has a much better understanding of what it's spending.

But imagine the CEO asks a slightly different question this time.

"Can someone show me a business problem we solved because our AI systems understood something we previously missed?" That question changes the conversation.

Suddenly, it isn't enough to demonstrate that customer support answers questions faster or that engineers produce code more quickly, the company needs to show how those improvements connect to something meaningful, did the organization identify a recurring customer problem before it became a major source of churn? Did it discover that two departments were unknowingly working toward conflicting objectives? Did it recognize an emerging business opportunity because information from several systems was finally brought together? And when the company discovered something important, did that understanding become available to the people and systems that needed it? Or did it disappear into another report, another dashboard, or another conversation that nobody outside the room would ever see?

Those are much harder questions to answer, they're also, I would argue, considerably more interesting.

I don't believe AI's greatest contribution to business will be the ability to generate documents in seconds or write code faster than a human developers, those capabilities are remarkable, and they'll continue improving, but they're still largely improvements to how individual tasks get completed, the bigger opportunity might be helping organizations overcome a problem they've struggled with for as long as organizations have existed… Knowing what they know, understanding what that knowledge means, and making sure it reaches the right place before an important decision is made, think about the implications for a moment, a company where customer frustrations don't remain isolated inside support tickets, where product decisions can be connected to the evidence that informed them, where engineering understands not only what it's building but why, and where AI agents don't need to independently rediscover the same organizational history every time they're asked to do something useful.

None of this guarantees that the company will make perfect decisions, people will still disagree, markets will change, information will remain incomplete, and sometimes the evidence will point in several directions at once… But at least those decisions could be made with a more consistent understanding of what's happening across the business.

That's a different ambition from simply making employees more productive. And it's one I think deserves considerably more attention.

We're still early in the development of enterprise AI, the models will improve, infrastructure costs will change, and many of today's limitations will eventually become less significant, but I suspect the problem of fragmented organizational knowledge isn't going to disappear simply because the next generation of models becomes more capable.

A brilliant AI system can still make a terrible recommendation when it doesn't understand the business it's working for.

So perhaps the next time an executive walks into a board meeting to explain the company's AI strategy, the conversation should begin somewhere other than the number of agents deployed, the amount of money invested, or the hours employees have saved.

Perhaps it should begin with a simpler question.

Are we actually helping our organization understand itself better?

Because if the answer is no, I'm not sure another thousand AI agents are going to change much… They might just help us become extraordinarily efficient at being confused.