AI Doesn't Eliminate Operations. It Changes Where Human Judgment Belongs.

AI is getting good at doing work.

As we see in the news too often, it creates a temptation for leaders: if automation can handle 50% of a workflow, why not 80%? And if it can handle 80%, why not eventually remove the human layer altogether? Layoffs here we come!

This is taking us in the wrong direction - and yet the ship is taking a long time to right itself.

Companies are still making the mistake of treating automation as the operating model, when it's just not there yet!

High-stakes AI systems still need, and I'd argue will always need, deliberately designed points for human judgment, escalation, quality control, and accountability. AI should make workflows faster and easier. Eliminating people without redesigning where judgment belongs eliminates the guardrails, institutional experience, and contextual understanding required to maintain service quality and protect customers and users.

The objective shouldn't be maximum automation.

It should be the "right" allocation of work between machines and people.

Automation is a capability. Operations is a system.

I've spent much of my career building and running large-scale operations where technology, people, policy, vendors, and customer impact intersect.

One lesson has remained consistent: improving one component of a system does not automatically improve the system - it's a series of trial and error and slow periodic improvements to get things right.

AI doesn't change that.

A model can become substantially better at classifying cases, answering questions, identifying patterns, generating content, or taking curated actions which can eliminate enormous amounts of repetitive work.

It should.

But an operating model has responsibilities that extend beyond completing individual tasks.

Someone—or something—must recognize exceptions. Someone must determine when a policy doesn't adequately cover a new situation. Quality has to be measured, and sometimes it's subjective. Escalations have to go somewhere. Failures have to produce learning. Customers need recourse when the system gets something wrong.

And when consequences are significant, accountability still has to live somewhere.

That is operations.

Human-in-the-loop isn't the answer either

There's an easy response to these concerns: put a human in the loop.

I don't think that's sufficient - it's just the start - but we need to know where to insert them.

Requiring a person to approve every AI decision is excessive, and can eliminate much of the speed and scale that made automation valuable in the first place. Worse, people who spend their day approving hundreds of machine recommendations can gradually become confirmation mechanisms rather than meaningful decision-makers.

A human checkpoint only creates value when the person at that checkpoint has a reason to be there which adds tangible value.

So the question is:

Where does human judgment materially improve the outcome?

That changes how we design the workflow.

A practical model: Automate → Observe → Escalate → Judge → Learn

I think about an AI-enabled operating system as five connected functions.

Automate

Let AI perform work where it is reliable, repeatable, and economically sensible.

There is little value in preserving manual work simply because people have historically performed it.

If a system can safely resolve routine cases, summarize information, classify inputs, identify patterns, or execute predictable processes, automate them.

But automation should have defined boundaries.

Observe

Automation without visibility creates operational blindness.

Leaders need to know how the system is performing—not merely how much work it is completing.

That means monitoring quality, exceptions, customer outcomes, unusual patterns, model confidence, policy violations, and changes in behavior over time.

The question isn't just:

Did the AI complete the task?

It's:

Did it produce the outcome we intended?

Those are very different measurements.

Escalate

Not every decision deserves human attention.

The operating system should identify the ones that do.

Escalation might be triggered by low confidence, unusual behavior, conflicting signals, policy ambiguity, customer impact, financial exposure, safety concerns, or simply a situation the system hasn't encountered before.

A mature AI operation doesn't put humans everywhere.

It puts them where uncertainty and consequence intersect.

Judge

This is where experienced people become more valuable, not less.

Humans are particularly useful when a decision requires context the system doesn't have, competing priorities have to be reconciled, policy is ambiguous, consequences are difficult to reverse, or the organization must make a judgment rather than retrieve an answer.

This is also why eliminating experienced operators too aggressively can create an unexpected problem.

You aren't only removing labor.

You may be removing the people who recognize when something doesn't look right. The checkpoints and gut decisions that made your product or process good might be going away - and you might be making things worse.

Institutional knowledge often lives in those judgments long before it appears in a process document, policy, or training dataset.

Learn

The human layer shouldn't simply catch AI mistakes forever.

Every intervention should have the potential to make the system better.

Why was the case escalated?

What did the system miss?

Was the problem the model, the data, the policy, the workflow, or the operating assumption?

Should a similar case be automated next time? If not, how can we make sure we kick these cases out to the best person to handle them?

Over time, the boundary between automated and human work should move.

That's how AI creates increasing leverage without sacrificing control.

The amount of human involvement should depend on consequence and uncertainty

This suggests a fairly simple operating principle.

When consequence is low and confidence is high, automate aggressively.

When consequence is higher but confidence remains high, automate and monitor.

When confidence drops or the situation becomes unusual, introduce escalation.

And when consequence is high, ambiguity is substantial, or a decision is difficult to reverse, human judgment becomes much more valuable.

The important point is that these boundaries shouldn't be permanent.

As models improve, policies mature, and organizations learn from exceptions, more work can move toward automation.

Likewise, new failure modes may require temporarily moving decisions back toward human review.

A good operating model can move in both directions.

Efficiency isn't the same as removing people

This distinction matters because organizations understandably want measurable returns from AI.

Headcount reduction is easy to measure.

But it isn't the only form of leverage.

AI can allow the same organization to process dramatically more work. It can reduce response times. It can let experienced employees focus on difficult cases instead of repetitive ones. It can expose patterns humans wouldn't easily detect. It can make smaller teams capable of operating systems that previously required much larger organizations.

Those can be better outcomes than simply asking how many positions can be eliminated.

The strongest AI operating model may therefore have fewer people doing routine work while concentrating more experienced people around exceptions, quality, policy, improvement, and judgment.

That's a fundamentally different workforce design.

The operator's job is changing

As AI becomes more capable, operations leadership becomes less about coordinating people performing individual tasks and more about designing the system through which work gets done.

That system increasingly includes people, AI models, agents, vendors, policies, controls, data, and escalation paths.

The leadership questions become:

What should be automated?

What needs to be observed?

What should trigger intervention?

Who has authority to override the system?

How do exceptions become learning?

How do we know whether customers are actually receiving a better service?

And who is accountable when the system fails?

Those aren't primarily AI questions.

They're operating-model questions.

AI will continue eliminating work that doesn't require people. That's a good thing.

But the organizations that get the most from it won't necessarily be the ones that remove humans fastest.

They'll be the ones that become exceptionally good at deciding where machines should act, where people should judge, and how each makes the other better.