AI, Automation and the Discipline of Judgment
Capability is becoming abundant. Authority still has to be earned.
6 min read · Executive Insight
I have spent enough of my career around technology transformation to become suspicious of sentences that begin, “Now that we have this technology…”
Now that we have cloud. Now that we have automation. Now that we have better data. Now that we have AI.
The technology changes. The organizational temptation does not.
A new capability arrives and the conversation quietly reverses itself. Instead of asking what problem deserves intervention, we start asking where the new technology can be used.
AI makes that reversal more consequential because capability is arriving faster than most organizations can develop judgment about its use.
The interesting question is no longer simply whether AI can perform a task. Increasingly, it can. The harder question is what an organization should permit intelligence to decide, recommend and do — and what must remain deliberately human.
The capability question is becoming the easy question
AI can interpret documents, summarize information, generate recommendations, interrogate data, communicate, invoke tools and initiate actions. Those capabilities will continue improving.
That makes “Can it do this?” a progressively weaker executive question.
I want to know something different.
What problem are we actually solving? What kind of decision does the problem require? How much uncertainty can we tolerate? What happens if the answer is plausible but wrong? Who remains accountable when a machine participates in the decision? What evidence would tell us that the intervention made the organization better rather than merely more automated?
Those questions are less impressive in a demonstration. They are much more important in an operating model.
Not every nail needs the newest hammer
I do not see deterministic automation, data science, large language models and agents as generations replacing one another. I see different instruments.
Stable rules and low variance may call for deterministic automation. Prediction and optimization may call for data science. Language and unstructured information create a natural domain for LLMs. Reasoning combined with context, tools and controlled action makes agents interesting.
But interesting is not the same as appropriate.
One sign of a mature technology organization is its ability to look directly at a fashionable capability and say: we do not need it here.
That is not resistance to innovation. Sometimes it is evidence that the organization understands innovation well enough not to confuse novelty with value.
Technology has always been willing to institutionalize our mistakes
One lesson from earlier generations of transformation keeps returning: automating a bad process gives you a faster bad process.
If the process is unnecessarily complicated, automation can make the complication more efficient. If accountability is unclear, workflow can move the ambiguity faster. If the underlying information is poor, analytics can give bad information greater authority.
AI does not repeal those lessons. It amplifies them.
Before automating work, I want to understand why the work exists, where the friction really occurs, what decisions are being made, who owns them and whether every step still deserves to exist.
Sometimes the highest-value automation decision is eliminating the work.
That is why I have never viewed transformation primarily as a technology exercise. Technology creates capability. Transformation changes how an organization operates.
AI is introducing a new form of delegation
Traditional software largely executes instructions we explicitly encode. AI increasingly participates in interpretation. Agents begin to participate in action.
That changes the governance problem.
We have spent decades designing delegation among humans: responsibility, authority, escalation, segregation of duties, approval limits and accountability. We learned, often painfully, that giving someone responsibility without clear authority creates dysfunction.
We may now be preparing to repeat the same management mistake with machines.
The architecture of an AI-enabled organization will eventually have to describe more than systems, applications and people. It will also have to describe machine actors: what they may know, infer, recommend, change, execute and escalate.
A system recommending how to categorize an internal document presents one level of consequence. A system communicating with a customer, altering financial information or initiating a transaction presents another.
The architecture should know the difference. So should the leadership.
Recommendation is not authorization
This distinction has become increasingly important to me.
A system may become extraordinarily good at producing a recommendation. That does not automatically mean it should have authority to act on it.
Learning from an outcome is not permission to change policy. Confidence is not authorization. Autonomy is not accountability.
If those boundaries are not explicit, an organization can slowly delegate authority without ever consciously deciding that it has done so.
That is not primarily an AI failure. It is a leadership failure.
Governance therefore cannot be the group that arrives after the prototype carrying a checklist. Evidence, authorization, traceability, security, quality and accountability have to influence the capability while it is being designed.
The prototype is where the difficult conversation begins
I like prototypes. They make abstract ideas tangible and expose assumptions quickly.
But I have learned not to confuse enthusiasm at a demonstration with readiness for an enterprise.
The distance between “look what this can do” and “we can depend on this” contains architecture, security, testing, integration, data quality, exception handling, observability, support, cost, controls, ownership and operating accountability.
So after an impressive prototype, I prefer a different question from “How quickly can we deploy it?”
What would have to become true for us to trust it?
That question changes the room. It moves the conversation from technological possibility to operational responsibility.
The executive does not have to replace the expert to challenge the evidence
I do not believe a senior technology leader needs to be the strongest production coder in the room. I do believe that leader has to know enough to challenge the room.
When engineering tells me something is ready, I want to understand what evidence supports the conclusion. What did we test? What did we not test? What assumptions are embedded in the architecture? Where is the technical debt? What are the security findings? What happens when the system encounters something we did not anticipate? Can failure be detected? Can we recover? Who owns the remaining risk?
The purpose is not to substitute executive opinion for engineering expertise. It is to make consequential assumptions visible before the organization inherits them.
AI increases the number of assumptions we can operationalize at machine speed. That makes disciplined challenge more important, not less.
Innovation should have to earn its place
I have become less interested in counting pilots, use cases and technologies introduced.
I want to know what changed afterward.
Did we make a better decision? Remove meaningful friction? Create capacity? Reduce risk? Improve a customer experience? Stop asking people to perform work technology should be doing? Build something the organization can actually sustain?
And did the value remain after the demonstration team left the room?
If not, we may have produced innovation theatre rather than transformation.
Organizations have finite money, attention and change capacity. Choosing what not to pursue is therefore part of innovation leadership.
When intelligence becomes abundant, judgment becomes scarce
This is the part of the AI transition I find most interesting.
We are building machines because they can perform increasingly sophisticated forms of reasoning. It is tempting to conclude that human judgment therefore becomes less important.
I think the opposite may happen.
When technological capability was scarce, much of technology leadership concerned itself with what systems could technically accomplish. Capability is becoming abundant.
The scarce capability moves upward: deciding what matters, what deserves automation, what should remain human, what evidence deserves trust, what uncertainty is acceptable, where authority should reside and when an organization should deliberately refuse a capability simply because it exists.
Judgment cannot be installed like a platform.
It accumulates through successes and failures; through sitting in rooms where every option has consequences; through discovering that the technically elegant answer can be operationally wrong; through learning when governance protects the organization and when governance itself becomes the constraint; through knowing enough to challenge experts without pretending to replace them.
AI may eventually make intelligence inexpensive.
It will not make consequence inexpensive.
That is why I suspect the organizations that extract the greatest value from AI will not ultimately distinguish themselves by having access to better intelligence. That advantage will narrow.
They will distinguish themselves by knowing where intelligence belongs, how much authority it should have, and when the responsible answer is still: not here, not yet, or not this way.
That is the discipline behind intelligent transformation.