AI Found 400 Problems. Congratulations! Now You Have 400 Problems.

We've spent the better part of the last decade convincing companies that they need more data, more analytics, more dashboards, and better visibility into absolutely everything happening inside their organizations, and now that artificial intelligence can analyze all of it in a fraction of the time, we've discovered a rather inconvenient problem.

We might have too much information about what's wrong and not nearly enough understanding of what to do about it.

Think about a typical product organization, customer support collects complaints, sales records objections, product managers conduct research, engineers track bugs, and somewhere in the middle of all this, a rather exhausted group of people tries to figure out what deserves attention.

Traditionally, analyzing that information took time, sometimes an unreasonable amount of time, which meant that important patterns could sit unnoticed for months, buried under thousands of support tickets and customer conversations.

AI has made a remarkable difference here, we can now process enormous amounts of feedback, identify patterns, generate insights, and produce recommendations at a speed that would have seemed ridiculous just a few years ago.

Wonderful, except there's a slight complication... What happens when your AI identifies 400 problems? Not four, not fourteen, but four hundred, all neatly categorized, summarized, and accompanied by recommendations that sound perfectly reasonable.

Someone opens the report, scrolls through the first twenty, realizes there's another 380 waiting, and quietly closes the tab.

Congratulations, you've just automated the creation of a new backlog nobody has time to deal with.

I find this particularly interesting because the technology is doing exactly what we asked it to do, it's discovering problems, analyzing information, and producing recommendations, all of which are genuinely valuable capabilities.

But discovering that something is wrong and understanding whether fixing it matters are two completely different things.

A customer might complain that a button is difficult to find, another might report that the application takes too long to load, and a third might be frustrated because a critical workflow requires six unnecessary steps.

An AI system can identify all three issues, maybe even explain why they're happening, but which one should the company address first? Well, that depends...

How many customers are affected, which customers are affected, how frequently does the problem occur, what business objectives does it influence, and what would it actually take to resolve it?

A minor inconvenience experienced by ten thousand customers might be more important than a severe issue experienced by five, unless those five represent half your annual revenue, or the issue prevents them from completing something essential.

And now we've moved from identifying problems to understanding their significance.... That's where things get a lot more interesting.

Atlassian published its State of Teams 2024 research, and one finding caught my attention, 70% of the knowledge workers surveyed agreed that they would make better progress if they had fewer, more specific goals, while 64% said their teams were constantly being pulled in too many directions. You can explore the research here: https://www.atlassian.com/blog/state-of-teams-2024

Now consider what happens when we introduce AI systems that can produce recommendations faster than the organization can evaluate them, we might accidentally make an existing problem significantly worse.

Instead of a product manager struggling with twenty competing priorities, we give them an AI-generated list of two hundred, complete with thoughtful explanations, suggested solutions, and perhaps a confidence score that makes everything look impressively scientific.

And because every recommendation comes wrapped in a professional-looking summary, it's tempting to assume that all of them deserve serious consideration.

This reminds me of something else that has been bothering me about the way we're using AI.

BetterUp Labs and Stanford Social Media Lab published research in 2025 examining what they called workslop, AI-generated work that looks polished but leaves the difficult thinking to whoever receives it, their survey of 1,150 US desk workers found that 40% had received this kind of work in the preceding month, with respondents reporting an average of about two hours to resolve each incident. The findings are here: https://www.betterup.com/workslop

Now, that research wasn't specifically about AI-generated product recommendations, but the underlying problem feels familiar.

It's remarkably easy to confuse a beautifully formatted output with a useful one.

I can generate a report containing fifty recommendations, attach charts, confidence indicators, and descriptions of potential benefits, but if nobody can explain which three recommendations will have the greatest impact on the business, have we actually made a decision easier?

Or have we simply given someone fifty things to read?

This is where I think enterprise AI needs to mature beyond its current obsession with producing more... More insights, more recommendations, more tasks completed, more agents running, more dashboards showing how much activity is happening, and at some point, somebody needs to ask whether any of it is making the organization better.

The challenge isn't to create an AI system that magically knows what every company should prioritize, that would be a rather ambitious claim, especially considering how often executives disagree about priorities themselves.

The challenge is to give the system enough understanding of the organization's goals, customers, constraints, and available evidence that its recommendations can be evaluated in context.

If a company is trying to improve retention, issues affecting customers at risk of leaving deserve particular attention, if it's trying to improve conversion, friction in the purchasing journey becomes more significant, and if engineering capacity is limited, implementation effort has to be part of the equation.

Of course, priorities change, sometimes rather dramatically, particularly when the board decides that the strategy everyone spent six months developing needs to be completely reimagined by next Thursday, so the understanding behind those recommendations needs to change as well.

And importantly, the system should be able to explain why something deserves attention, what evidence supports that assessment, and where uncertainty remains, because a recommendation without a defensible explanation is really just an opinion delivered by a machine.

McKinsey's 2026 State of AI research offers an interesting perspective here, finding that organizations generating the greatest value from AI are much more likely to redesign workflows and focus on broader business outcomes rather than treating efficiency as the only objective... Their research is available here: https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai

That makes sense, because the real benefit of intelligence isn't the ability to produce an endless collection of observations, it's the ability to use those observations to make better decisions, and perhaps we've been approaching this from the wrong direction.

We've been measuring how much information AI can process, how many patterns it can discover, and how quickly it can generate recommendations, all while assuming that more output naturally translates into more value, but organizations have limited resources, limited time, and a surprisingly limited capacity to agree on what they should do next.

Giving them another four hundred recommendations isn't necessarily progress, but helping them understand which four are worth doing, and why, might be considerably more useful.

Because the next competitive advantage in enterprise AI may have very little to do with how many problems your agents can find.

It may come down to whether they can help your company stop wasting time on the wrong ones.