Enterprise AI shouldn't automate the way we work today. It should redesign the work with the agent at the center.

Every week I see enterprise AI providers announce features that show an agent logging in to a system, opening an app, filling a form, clicking Submit and confirming completion. The demos are impressive. The more I think about them, the more convinced I am that they point in the wrong direction.

The autonomous aircraft problem

Imagine you were building an autonomous aircraft. You wouldn't put a robot in the pilot's seat to work the steering, the brakes and every other control. You'd design an aircraft without a cockpit. The cockpit was only ever there because a human had to fly the plane.

Elon Musk made the same point to a room full of US Air Force leaders. At the Air Force Association's Air Warfare Symposium in February 2020, the conversation turned to what comes next in air combat. Most people in that room had spent their careers thinking about faster, more powerful aircraft, with a pilot still in the seat. Musk's answer was blunt: “The fighter jet era has passed." The future, in his view, belonged to locally autonomous drones.

You can debate the timeline. But the logic holds. When the machine can make the decisions, you stop building better cockpits and start asking whether you need one at all.

Enterprise IT is no different. Legacy systems and enterprise tools were built with interfaces for humans to operate. Every console, dashboard and carefully designed UI exists because an admin had to click through it. Over the years, our processes and SOPs grew up around those screens: log in here, check that, escalate if this.

An AI agent doesn't need to click a button on a fancy UI. When we hand it the same SOP we wrote for a human admin, we aren't transforming anything. We're dumbing down a capable system and asking it to pretend to be a tired operator.

The question isn’t “which steps can AI automate?” It’s “if an agent ran this from day one, what would the process look like?”

2,000 tickets a day

Last week I met the IT head of a large enterprise. They run one of the best-known network monitoring (NMS) tools on the market, set up the way most enterprises set it up: configure thresholds, and the NMS logs a ticket for every alert.

After an expensive, six-month deployment, their network team now receives around 2,000 tickets every day. Their next step is to bring in agentic AI to triage those tickets and separate signal from noise.

It sounds sensible. But it's worth asking why the problem exists in the first place.

Until recently, our tools had no ability to make a decision. So the safe design was to create as many alerts as possible, route all of them through the ticketing system, and have a service desk engineer look at each one and choose what to act on. The ticket became the place where decisions happened, because only a human could make them.

Bringing in an AI agent after the ticket is logged keeps that entire design in place. The agent ends up cleaning a queue that shouldn't exist.

Redesigning with the agent at the center

Now flip it around. Instead of the ticketing system being the front door, the AI agent is:

• Every alert goes straight to the AI agent, not to the ticketing system.

• The agent triages and investigates. It correlates related events, checks topology, recent changes and history, and works out the likely root cause.

• The agent decides what is noise, what is a symptom of something else, and what actually needs to be resolved.

• Resolution is split by capability. The agent fixes what it can on its own. A ticket reaches the service desk only for work that needs a human, with the investigation already done and the context attached.

The result: the service desk sees the 20 issues that matter, not 2,000 alerts. The ticket is no longer the entry point. It becomes the hand-off.

Don’t automate the queue. Remove the reason it exists.

What changes for people

Putting the agent at the center doesn't take humans out of the loop. It changes where they add value:

• Agents do the work. They talk to systems directly through APIs, events and data, not screens.

• Humans set intent and apply judgment. They define outcomes, policies and approval limits, and step in on the exceptions that need them.

• Processes run at machine speed. No queues waiting for someone to pick up a ticket, and no hand-offs that exist only because of how teams are organized.

• Success is measured by outcomes, not by tickets closed or steps followed.

A note on legacy systems

Not every tool in the enterprise has an API, and screen-driving agents will sometimes be the only way in. That's fine as a bridge for brownfield systems. It just shouldn't be mistaken for the destination.

The real shift

The biggest opportunity in agentic AI for IT operations isn't making today's processes faster. It's admitting that many of those processes were designed around a limitation, the absence of anything that could make a decision, that no longer exists.

So stop writing SOPs for agents. Start redesigning processes around them. Build the aircraft without the cockpit.