The Next Bottleneck Is Judgment: Why We're Building Quint
The Next Bottleneck Is Judgment
Why We're Building Quint#
For decades, software has been automating repetitive work. I've spent enough time around engineers to know that if a task has to be performed twice, someone is already spending a week automating what could have been done manually in ten minutes.
Until recently, most automation targeted routine tasks; software processed invoices, generated reports, moved information between systems, and eliminated manual workflows. Humans still spent most of their time interpreting information, making decisions, and determining what mattered, but AI changes that.
Today we're not just automating repetitive labor, we're increasingly automating portions of knowledge work itself. AI can write code, summarize documents, analyze data, generate presentations, and answer questions, meaning the cost of execution is falling rapidly.
As execution becomes cheaper and more abundant, a different resource starts to matter more: judgment.
There's been a growing body of work arguing that as machine prediction improves, human judgment becomes more valuable. AI can help almost anyone move faster, but it can't always determine which opportunities are worth pursuing, which risks are worth taking, or which tradeoffs matter most. The bottleneck increasingly shifts from doing the work to deciding what work should be done.
Historically, organizations were designed around this distinction → .
Junior employees gathered information, prepared analyses, executed tasks, and learned through repetition.
Senior employees made decisions.
Organizations scaled by adding layers of people.
Over time, junior employees accumulated the experience needed to develop judgment themselves. Much of what we call expertise was really apprenticeship. You reviewed spreadsheets, built decks, debugged systems, and investigated incidents. You learned which details mattered and which didn't and eventually, after enough repetition, you developed intuition for making decisions under uncertainty.
The challenge is that AI is beginning to automate many of the activities that historically created that intuition.
If AI removes a meaningful portion of the "junior work," how do people develop senior judgment? Some argue that AI itself will eventually solve this problem, and maybe it will, but even in a world where models continue improving, organizations will still need to understand consequences, evaluate tradeoffs, and decide what outcomes they care about. Those are fundamentally judgment problems.
My guess is that most existing industries won't radically reinvent how expertise is developed. They'll automate larger portions of the work, reduce the number of junior roles, and rely on a smaller apprenticeship pipeline to replenish senior talent over time. Newer companies may look very different: fewer people, more automation, and dramatically higher leverage per employee.
That creates an interesting challenge. If judgment becomes more important while opportunities to develop it become less common, how do we help people build better judgment in the first place?
That's one of the reasons we're building Quint.
Much of the industry is focused on creating better AI developers, AI agents, and AI reviewers to automate more of the implementation process. Those tools are important, but we're interested in a different question: how do we help people understand the consequences of their decisions?
Generating code is becoming cheap but understanding what that code actually does is not.
A recent paper showed that AI agents lead to a 741% increase in lines of code and a 65% increase in pull requests, but releases increase by only 20%. The bottleneck is no longer generating software, it's deciding whether or not that software should make it into production. Will this introduce risk? What behavior changes? What business outcomes are affected? What assumptions are being made? What tradeoffs are being accepted? These are all judgment questions, and they are exactly where Quint comes in.
Quint uses formal verification and behavioral analysis to help teams reason about the systems they build. Rather than only accelerating execution, it helps surface consequences, expose assumptions, and make the effects of change easier to understand. It’s the decision layer between generated code and production.
In many ways, that's what makes judgment valuable in the first place. Judgment isn't just choosing between options. It's understanding the landscape well enough to know which options matter.
The part I keep coming back to isn't whether AI will get better at execution, it clearly will. It's whether the organizations that win will be the ones that automate work fastest, or the ones that figure out how to scale judgment. We're betting on the latter.
Or maybe the future is just a handful of humans, a small army of agents, and Quint sitting in the middle asking uncomfortable questions about the consequences of everyone's decisions.