Most conversations about AI begin with productivity.
How many hours can it save?
How many people can it replace?
How much faster can software be written?
Those are interesting questions.
They just aren't the most important ones. The more important question is this:
How do you adopt AI in a way that increases human capability without allowing your organization to lose understanding, judgment, or control?
That question is much harder to answer. And it isn't something you learn from product demos or benchmark reports. You learn it by doing the work.
That's why I recently took what I call a "codecation".
What is a Codecation?
A codecation is something I've done for years. Once or twice a year I intentionally step away from executive work for several days and return to engineering. Not because the company needs another developer. Because I need to remain connected to the work we're asking our engineers, and increasingly our AI, to perform.
Every codecation has one concrete engineering objective. This year's goal was improving Unicode performance inside VAST. But the engineering objective was only part of the reason for the trip.
I wanted to experience AI exactly the way an engineer experiences it: not during a polished demonstration or through carefully prepared examples, but while solving a difficult problem inside a mature production system with decades of engineering history behind it.
If we are going to help organizations adopt AI responsibly, leadership needs to understand where it creates leverage, where it creates risk, and where human judgment still matters.
My biggest takeaway from this codecation was not that AI can produce more work. It was that AI can dramatically expand the amount of complex work one experienced person can investigate, while also creating a new management challenge: the technology can move faster than human understanding.
That creates both extraordinary leverage and real risk.
The AI could analyze a large amount of information, identify problems, propose improvements, generate tests, and suggest new directions at a pace that was difficult to absorb. The limiting factor was no longer the number of ideas available. It was my ability to understand them, judge them, and decide which ones were safe and worthwhile.
That is why I believe this was such a strong use case.
I chose a difficult performance problem in a mature part of our product. It involved many interconnected layers, years of accumulated engineering decisions, and multiple paths optimized for different situations. Improving one area could easily create a problem somewhere else.
The VAST AI Assistant made a significant difference because it was connected to the actual working environment, not operating as a separate chat tool.
It could work across the live application, the debugger, the current runtime state, the tests, the underlying platform code, and the Git repository. I could begin with a real problem, attach the debugger to the existing conversation, inspect what was happening, follow the issue through the system, make a controlled change, and measure the result.
That continuity was remarkable.
The work produced meaningful results:
The benefit was not simply that the AI wrote code faster.
It helped me examine more possibilities, challenge more assumptions, and move across a much broader part of the system than I could reasonably have covered alone in the same period.
The experience was intense and tiring, but the payoff was substantial.
The AI never ran out of ideas.
It could identify a possible defect, suggest an optimization, propose a broader redesign, recommend another test, and surface a new capability almost simultaneously.
I still had to determine:
My own understanding had to keep pace with the work.
That was the source of much of the cognitive strain. The technology could continue expanding the investigation long after my ability to hold the full picture in mind began to decline.
This is an important leadership issue.
AI can increase the rate of work faster than it increases the human capacity to evaluate that work.
I learned quickly that I could not base confidence on how persuasive the AI sounded.
When it said it found a defect, I required evidence that the defect existed in the current environment.
When it proposed a performance improvement, I required measurements before and after the change.
When it suggested a broader change, I broke the work into small steps so that I could understand the impact and reverse course safely.
My standard became straightforward:
The AI could help create the tests and measurements, but the evidence had to stand on its own.
This experience raises a larger question:
Does it matter for the human mental model to remain synchronized with the work the AI is producing?
One view is that people should stop reading or deeply understanding AI-generated work. Instead, organizations should surround AI agents with enough controls, tests, metrics, policies, and approval gates that the output can be trusted from the outside.
That model may work well when the rules are already known.
My experience was different because much of the work involved discovering the rules themselves.
The challenge was not only producing an answer. It was learning:
In that kind of work, controls are essential, but they cannot fully replace understanding.
Tests can enforce a known rule. They cannot identify a rule no one has discovered yet.
Metrics can detect a known failure mode. They cannot measure a risk no one thought to examine.
When the constraints are stable and well understood, greater autonomy may be appropriate. When the constraints are still emerging, the human mental model remains critical.
Large organizations will naturally view AI through the lens of productivity, margins, speed, and quarterly results.
The argument for maintaining human understanding cannot simply be that people should remain involved because work is meaningful. The stronger business argument is control.
When human understanding falls too far behind AI-produced work, the organization accumulates hidden risk.
That risk may appear later as:
A company can increase output while weakening its ability to understand how the work is being done.
That is a form of organizational debt.
It may remain invisible while productivity measures improve. It becomes visible when conditions change, a model behaves unexpectedly, a vendor changes direction, a critical employee leaves, or the organization encounters a situation its controls were never designed to handle.
Human understanding is not a sentimental constraint on productivity.
It is part of the company’s control system.
The goal should be maximum sustainable, risk-adjusted leverage.
Routine, reversible, well-understood work may be suitable for high levels of AI autonomy.
Work that is strategic, consequential, difficult to reverse, or central to the company’s capabilities should require stronger human understanding and ownership.
The more important the decision, the more important it is that a person can:
That does not mean every leader must inspect every detail.
It means the organization must consciously decide where human understanding remains essential and where external controls are sufficient.
Every company will have access to increasingly capable AI.
The durable advantage will not come from access alone.
It will come from how well the organization can:
An organization that allows its people to lose understanding may gain short-term efficiency while weakening its long-term ability to learn, innovate, and remain in control.
This codecation began as an engineering exercise and it ended as a leadership exercise. The AI did not replace expertise during this codecation, it multiplied the reach of expertise.
But that leverage became trustworthy only when it was paired with proof, controlled change, human understanding, and clear accountability.
That's the leadership challenge. Not how quickly we can automate, not how many people we can replace with AI, but how intentionally we can augment. Because the organizations that win won't be the ones that remove humans from the process. They'll be the ones that combine human judgment with AI capability in ways that make both stronger.
That, to me, is responsible AI.
