The Capability Gap Nobody Is Talking About
A few months ago, we had to make a staffing decision on one of the most consequential projects of the year.
Major technology implementation. Touches every team. Changes how frontline staff work, how supervisors manage, and what the customer feels on the other end. The kind of project where you feel the pressure before it even starts.
We had a choice. Bring in someone organized and capable. Someone who could build a clean project plan and execute against it. Or put our most seasoned operator on it.
We put our most seasoned operator on it.
Not because they are better at project management in the abstract. Because they have been on the floor. They have lived through implementations before. They know which teams resist change and why. They know the personalities in each room. They know which past decisions will surface as constraints that nobody put in the plan. They know which success measures actually matter and which ones get quietly dropped when things get hard.
No tool | LLMs gave them that. The work did. And no AI tool, no matter how capable, could have told us who that person was.
Mohanbir Sawhney, a professor at Kellogg, (He writes on substack as well linkedin) recently drew a distinction I had not been able to put into words. He separated two gaps that AI is creating. He called them the fluency gap and the capability gap.
The fluency gap is what AI closes. What you could not do before, you can do now. A junior analyst today can produce in hours what used to take days. Drafts, research, synthesis, build models. The tools level the playing field in terms of output. This is most of the conversation right now.
The capability gap is what AI cannot close. It is about what you can do with those tools based on what you already understand. That gap does not shrink with a better prompt or a faster model. It closes slowly, through experience, through pattern recognition, through having been in enough situations to know what a situation is about to do. One of those is most of the conversation. The other is most of the truth.
W. Edwards Deming spent decades making a related argument: put a good person in a bad system and the system wins. What I keep seeing play out is the reverse. Put the wrong person in a good system - give them every tool, every resource, every process and the system still breaks. Because systems do not run themselves. People with judgment run them.
The clearest signal of the capability gap is not the answers someone produces. It is the questions they ask. AI is exceptional at generating answers. Draft me a project plan. Write me a synthesis. Give me an analysis. It delivers. What it cannot do is tell you whether the question was the right one to ask in the first place.
The Innovator’s DNA - Dyer, Gregersen, and Christensen, named questioning as one of the five core skills of innovators. What they could not fully capture is that questioning at depth is not a workshop skill. It is not something you develop by reading about it. It is built through exposure, the ability to walk into a room and notice what is missing from the slide before anyone names it, because you have sat in enough of those rooms to know what should be there.
Judgment earned by being wrong and then right enough times that the instinct becomes reliable.
You do not get there faster by having better tools. You get there by being in it long enough that pattern recognition stops feeling like thinking and starts feeling like sight.
My best operators do not ask more questions. They ask fewer, better ones. And they ask them at the moment that actually changes the outcome, not after. This is why the middle of the org chart is in its most exposed position in decades.
For years, that layer held the space between leadership and execution. They translated strategy into action. They delivered on results, OKRs. They synthesized data into decisions. They absorbed the complexity so that leadership did not have to.
AI is starting to do something like that. And what is less discussed is the second-order effect: leadership, the people who own the why and the what, can now access the output directly. What I am hearing from operators across the industry, and seeing in how the most effective leaders are operating, is that the top layer is getting more hands-on, not less. The tool makes it possible. Which means the layer that used to be the indispensable translator is being compressed from both ends.
If you are in that layer, the answer is not to get better at the tools. Everyone has the tools. The answer is to develop the part of the work that does not get easier when the tools improve.
The judgment. The taste. The questions that nobody else is asking.
Here is one of those questions that I think the industry is not asking loudly enough. If AI is supposed to reduce the cost of serving customers, why are most companies’ AI bills growing faster than their cost savings?
Because AI has done something the original business case did not account for. It has surfaced demand that was not visible before.
Before AI, only a subset of your customers reached out for help. Most either figured it out themselves, gave up, or never tried. The friction was a filter. Now the friction is gone. Asking is easy. Experimenting is easy. The bar to engage has dropped, and the volume has expanded in ways that were not in the model.
What looked like a deflection story is quietly becoming an inflation story. Consumption-based pricing means your bill grows with every interaction the system handles, including the ones that would never have happened before. The volume goes up. The tech spend goes up. And the ROI calculation you used to justify the investment starts to show its assumptions.
I want to be precise here. I am not talking about the ROI of AI broadly. I have spent real time putting numbers behind avoidance costs and value creation, connecting AI outcomes directly to the bottom line. That math is real and important.
What I am talking about is shadow demand. When a subset of your customers becomes most of your customers overnight, and when the model you built to validate the spend was designed around the old subset, you have a structural shift on your hands. Not a tool problem. A thinking problem.
The companies that navigate this well will be the ones with people inside who could see it happening before the quarterly numbers made it obvious.
That is not a fluency skill. That is a capability one. The capability gap rarely announces itself. It does not show up as a wrong answer. It shows up as a question that did not get asked. A pattern that did not get named. A decision that looked clean in the moment and cost something real six months later.
There is so much momentum behind AI right now that the conversation has its own gravity. Predictions, frameworks, adoption rates, deployment stories. The noise is real and some of it is useful. But the questions that matter most, what are you actually displacing, what is being created that you did not intend, what are you not measuring — are quieter. They make worse content. They get less engagement.
You only learn to hear them by having been in enough rooms where missing the quiet question was the thing that cost you. The tools do not teach that. The work does.
I run AI at scale. The value is real. I am not writing this as a skeptic. I am writing it as someone who has been on the floor long enough to know that the technology is rarely the actual variable. And who keeps noticing that the most important questions in this moment - what are we displacing, what is surfacing that we did not plan for, what are we not measuring - are the ones that get the least airtime.
If you are sitting inside an organization right now watching this play out, what is the question you are not allowed to ask out loud?
If you know someone who needs to read this, send it to them.
Disclaimer: Guneet Singh writes The Ins & Outs of Experience, a newsletter for operators, CX leaders, and executives who want real insight from inside the work. Views are his own and do not represent the views of employers or past companies.