AI Agents, One Year Later — What Remained After the Hype Passed

A year ago we called it "the age of agents." Now it's time to talk about what that age was actually like.

Declaring Is Easy, Operating Is Hard

In How AI Is Changing the Way We Work, we described the shift of agents from an "AI that answers" to an "AI that works" — an AI that, given a goal, plans on its own, calls tools to execute, and checks the result.

A year has passed. The market moved firmly in that direction. But the direction it moved and the places that actually delivered results were not the same.

Two Truths the Numbers Tell

Start with the bright side. According to the research firm Gartner, 80% of the enterprise software shipped or updated in 2026 came to include at least one AI agent — a sharp jump from 33% in 2024. And one survey found that companies which pushed all the way to real operation saw an average return on investment (ROI) of 171%, roughly three times that of conventional automation.

Now the dark side. That same Gartner warned that about 40% of agent projects will be canceled before reaching operation by 2027. In reality, only about 31% of companies are running agents in production today.

Same technology, opposite outcomes. This gap is the heart of this article.

Why the Pilot Becomes a Graveyard

One small distribution company we met successfully completed a pilot for an AI agent that handled customer inquiries. The demo was excellent. Yet six months later, that agent had quietly been switched off.

The reason wasn't technical.

• Real inquiry data was far messier than the demo questions.
• There was no agreement on who was responsible for fixing the agent when it got things wrong.
• From the start there was no baseline number to judge whether it was "working well."

The trap we covered in The Pilot Succeeded — Now What? simply repeated itself, more sharply, with agents.

Four Traits the Surviving Few Shared

The few companies that reached operation had clear traits in common. What differed wasn't their level of technology but their way of preparing.

1. They laid the groundwork before adopting. Data, permissions, connections. For an agent to actually do work, it has to reach into the company's systems. The companies that laid this plumbing first were the ones that survived.

2. They set the rules in writing before starting. What the agent is allowed to do, and what must be handed off to a person. The companies that wrote this down before an incident met a different fate from those that wrote it down after.

3. They measured the "current numbers" before the pilot. How many hours, how many people, how much money this task takes right now. Without measuring this, you can't measure results later. And if you can't prove improvement, the budget gets cut off.

4. They assigned an owner. "Everyone's job" is no one's job. Only the places with someone accountable for the agent's performance made it to operation.

A Phase That Actually Favors Small Businesses

The large enterprises that ran many experiments at once on huge budgets are now winding those projects down. The froth of the hype is draining away.

For a small company, this isn't bad news. We were never able to do anything but focus on exactly one problem. Start narrow, confirm with numbers, and scale only what works — the very approach the surviving companies took is the approach small businesses were always good at.

The Difference Between "Bought It" and "Made It Work"

An agent isn't something you buy. It's a process of making it work.

The lesson a year of experiments left us is simple. Don't be won over by a flashy demo. Pick one narrow problem, measure the current numbers, assign an owner, and write the rules first.

The hype has passed. What remains now is only the companies that have agents quietly doing the work.