AI agents in operations: where they start to pay off

The conversation about AI agents usually stops at the demo: a clever question, a fluent answer, applause. But an agent doesn't earn its place by sounding smart. It earns it when it closes a ticket at three in the morning without waking anyone up.

The right question isn't "what can it do?"

It's "which task, repeated hundreds of times a month, currently eats a person you'd rather have resolving exceptions?". Agents pay off on high-volume, low-ambiguity work: classify, route, summarize, check against a policy, draft the first version.

A useful agent doesn't replace the expert. It clears the first ninety easy decisions so they arrive fresh at number ninety-one — the hard one.

Where we've seen them pay off

• First-line support. Triage, answering the known, and a clean handoff with context when it's time to escalate.
• Internal operations. Reconciling documents, extracting fields from PDFs, flagging the record that doesn't add up.
• Sales. Preparing account research before the call, so the human shows up with hypotheses, not a blank page.

Design matters more than the model

Almost nobody fails by picking the wrong model. They fail by releasing an agent with no clear limits: no sense of when to say "I don't know," no human at the right point in the flow, no way to audit why it did what it did.

We design an agent the way we design any system: with explicit states. What it does alone, what it proposes and waits for approval on, and what it escalates immediately. That boundary is the product. The model is just the engine.

What a good first project looks like

Narrow and measurable. One task, one team, one number you want to move: first-response time, cost per ticket, hours recovered. Small enough to ship in weeks; real enough that, if it works, nobody wants to go back to how it was.

That's the point where AI stops being a demo and starts being infrastructure.