The AI Capability Overhang

New capabilities can arrive faster than the habits needed to use them well. That gap is what I mean by the AI capability overhang.

The Overhang

There's a term floating around AI circles that leaders need to understand: the capability overhang.

The gap is between what a system can do on an appropriate task and what an organization has learned to use reliably. Closing it takes more than access to a model.

I've been talking about AI arbitrage for the past three years. The idea that you can secure massive gains by applying AI quickly in undervalued or overlooked areas of your business. Like financial markets, the window doesn't stay open forever. If you spot a valuable use case and act decisively, you reap outsized benefits before others catch up. If you drag your feet, the advantage evaporates.

The capability overhang is AI arbitrage with the stakes turned up to 100.

In its December 11, 2025 GPT-5.2 announcement, OpenAI reported GDPval scores of 70.9% for GPT-5.2 Thinking and 74.1% for GPT-5.2 Pro. These are wins or ties against human experts on the evaluation’s specified knowledge-work tasks, not the share of all work on which AI is preferred.

And yet most organizations are still using AI the way they used GPT-4 two years ago: ask a question, get an answer, move on. Summarize this document. Draft this email.

They're not running agent loops overnight. Not assigning hour-long tasks to AI co-workers. Not managing fleets of parallel workers across their backlog.

The people living in the future see a fundamental break. Everyone else sees incremental improvement. They're using the same tools.

Availability is only part of adoption

Here's the insight that changes everything once you see it.

Every major technology revolution in history required building physical infrastructure before it could go mainstream. Steam power needed factories. Electricity needed power grids. The internet needed fiber optic cables and cell towers. That physical buildout took decades. It was the constraint that slowed adoption.

Existing infrastructure can make tools easier to distribute.

Cloud services, broadband and familiar devices let many teams try AI without building a new delivery network. That lowers one barrier. It does not remove the need for compute, energy, integration or reliable access.

For a business, the next question is which available capability can be useful in its actual operating environment.

So what's left? Us.

People and operating practices are part of the constraint, alongside the technology and infrastructure required for the task.

That helps explain why an impressive demonstration can coexist with slow adoption. A useful tool still needs to fit the work.

The practical response is to keep learning, while making decisions at a pace appropriate to the work and its consequences.

The Real Skill Shift

In my own use of these tools, the change is from asking for a suggestion to assigning a bounded piece of work and reviewing what comes back.

Move from questions alone toward well-defined tasks and projects.

The chatbot mental model is a trap. When you treat AI as an oracle that answers queries, you're using a jet engine to power a bicycle.

The new approach is declarative. Describe the end state you want. Provide success criteria. Let the system figure out how to get there.

Instead of "how do I fix this bug," try "fix this bug, run the tests, and keep iterating until they pass."

Instead of "what's the best way to approach this market," try "analyze these three market entry scenarios against our criteria and rank them with supporting evidence."

The bottleneck moves from implementation to specification. "What should this be?" becomes harder than "How do I build it?"

That's a fundamental shift. Most leaders have spent careers developing intuitions about implementation. The new skill is specification: describing what you want precisely enough that a system can build it.

Build a learning rhythm

Choose a workflow where you can compare the result with a clear standard. Give someone responsibility for the inputs, the review and the decision about whether to use it again.

That is how a capability becomes operating experience.

The Window

AI arbitrage was always about windows closing. The capability overhang is the same dynamic on a compressed timeline.

The leaders who figure out how to actually use these systems, who assign tasks and projects instead of asking questions, will have compounding advantages over those who wait.

Specification becomes more valuable when a system can carry out more of a task. Clear instructions help; testing and judgment still determine whether the result is usable.

Make Sure You're Using What's Actually Available

One more thing, and it sounds obvious but isn't: many leaders are forming opinions about AI based on tools that are generations behind what created this overhang.

Late-2025 releases expanded the work teams could explore. The relevant test now is the current tool available to you, used on a task you understand, with clear requirements and appropriate review.

Do not base your whole view on an old demonstration. Revisit a real task, compare the results and update your expectations from what you observe.

The leaders building the future haven't fully moved into it yet. If they can't shake old habits instantly, you shouldn't feel bad that you haven't either.

But you should probably start.