Ethan Mollick published “The Overhang” on September 18, making an argument that is easy to state and harder to act on: AI capability is still growing exponentially, and most organizations are using it at a fraction of what it can actually do. The gap between those two curves — what AI can do and what we ask of it — is not closing. It is widening.
stateDiagram-v2 direction TB state "AI tool acquired" as A state "Generic prompts" as G state "Shallow output" as S state "Context invested in workflow" as C state "Domain-specific prompts" as D state "Compounding advantage" as P [*] --> A A --> G : default path G --> S : no domain framing S --> [*] : underuse A --> C : deliberate investment C --> D D --> P P --> [*] : overhang closing S --> C : restructure approach
The rundown
In “The Overhang,” Mollick draws on two recent developments — an AI’s apparent progress on a famously difficult mathematics problem, and rising public discourse about AI risk — and offers an interpretation that differs from most coverage. The concern is not that AI is developing too quickly for human institutions to absorb. It is that human institutions are moving too slowly to engage with AI as it already exists. The limit, in his framing, is not compute or capability. It is the user.
Mollick’s core observation: even if AI development stopped today, most people and organizations would have years of productive catch-up to do before hitting any real capability ceiling. The bottleneck is organizational, not technological. What makes the difference between shallow and deep AI output is not access to a better model. It is the context, domain knowledge, and clear task framing that the user brings to any capable model.
This is not a comfortable observation for leaders who have made AI tool investments and consider the problem addressed. It is a useful one.
For the working software engineer
The practical implication is about what actually compounds. Engineers who are ahead on AI usage right now are not generally the ones with access to the newest models or the most tools. They are the ones who have built scaffolding that provides AI with context: project documentation that feeds into code generation, architecture specs that constrain what AI produces to what the system actually requires, feedback loops that catch hallucinations and incorrect assumptions early rather than late.
The gap between a prompt that produces a generic answer and a prompt informed by three years of system context and team conventions is not small. That gap is the overhang in practice. Generic prompts against general AI models produce generic output. The investment that pays off is the institutional scaffolding around the model — not the model itself.
The productive work for engineering teams right now is not learning the next model’s quirks; those change quarterly. It is building internal documentation, context libraries, and process structures that make any capable model more effective in your specific environment than it is for a team that hasn’t done that work.
For business owners and operators
The overhang reframes the competitive question entirely. Organizations that will have durable AI advantage in 2028 and beyond are not necessarily the ones spending more on AI tools. They are the ones building deeper AI context: documented processes with clear decision criteria, institutional knowledge structured to be AI-readable, and specific task framing a capable model can act on accurately.
That is a leadership and culture problem before it is a technology problem. If your teams are treating AI as a general-knowledge query tool rather than a system that can engage with your specific processes, customers, and constraints, you are operating at a fraction of what is available. The overhang grows wider the longer that pattern holds.
The business case is direct: organizations that close the context gap early build a compounding advantage. The same model produces better output in a high-context environment than a low-context one — and that advantage widens on every subsequent model generation, because your context infrastructure carries forward while a generic user starts fresh each time.
My take
I ran into an earlier version of this problem at LERETA, the second-largest property tax payment processor in the U.S., where we were processing roughly $18 billion annually across a legacy mainframe stack that needed modernization. The technology capable of replacing the legacy system existed. The architecture team understood the path forward. The constraint was organizational: leadership could not see the full scope of what they were managing, which meant they could not commit to the investment required to change it.
What moved the needle was not a new piece of technology. It was a set of wall-sized enterprise architecture diagrams that made the full system visible at once — every developer team, every data dependency, every integration point, every risk on a single surface. When the board could look at the complete picture, they had context they could reason from. That context, not any specific technology, was what unlocked a multi-year, multi-million-dollar modernization program. My personal research became known as the Livermore Report — and it mattered not because it introduced new technology but because it provided the organizational clarity to act on technology the company already had.
Mollick’s overhang describes the same dynamic, now compressed by a factor of years. The technology is ready to do work most organizations have not asked of it. The organizational scaffolding to actually capture that capacity — the context, the process specificity, the institutional knowledge fed back into the tools — is what most organizations are still building. That is the gap worth closing.
