AI agent failures offer hidden value for enterprises

We've seen this movie before

The AI agent boom has many similarities to the mobile app explosion. In 2025 alone, there were 4.2 million apps on Google Play and 2.5 million on the Apple App Store, with thousands of new apps shipping daily. But the overwhelming majority never found an audience.

In 2015, Andrew Chen, a partner at Andreessen Horowitz, reported the average app lost 77% of its users within three days, 90% within a month and 95% within 90 days. By any production-grade metric, almost all of them failed.

But that framing missed the point. Most of those apps were never destined to be commercial hits. They were how companies learned mobile-first thinking, including:

• How to design for a small screen.
• How to earn attention.
• How to build for a new type of user.

The "failures" were tuition. The companies that treated the experimental phase as education instead of demanding a hit from every release built the muscles that defined the next decade.

AI agents are on the same journey. But this time the lesson cuts deeper, and it pays off in two ways at once.

The difference that changes everything

One thing separates the two eras: A mobile app lets you defer the data reckoning. An agent forces it.

A mobile app could succeed despite a messy backend. Clever design and a smooth interface could paper over years of data chaos that the user never saw. That's partly why people happily keep 80 apps installed while actively using only nine. The polish was real, even when the substance was thin.

An agent has nowhere to hide. It's only as good as the data it can reach, understand and act on. While building one, every weakness in your data surfaces with uncomfortable precision, including:

• Incompatible systems and inconsistent formats.
• Missing or meaningless metadata.
• A "source of truth" that three departments define differently.
• Institutional knowledge that lives in people's heads and was never written down.
• Integration debt between the CRM, the ticketing system, the product database and the knowledge base, which all suddenly must speak to each other.

What's left behind is worth more than the agent

A shelved agent isn't a wasted one. While the agent might not survive, the data work does.

Every agent experiment forces an organization to complete tasks it kept putting off. By the end of an agent's development, organizations will have documented tribal knowledge, set data quality standards, broken down team silos, defined ownership and governance and built the integration layers that make data usable everywhere. None of that disappears when the pilot ends. It compounds.

In practice, this looks like the customer service agent that never launched leaving behind a unified, cleaned customer record that improves every system downstream. The discarded research assistant leaving a tagged, searchable knowledge base that employees use for years. Or the sales agent that stalled in testing and left standardized pipelines that power better analytics across the business.

People quietly gain AI skills

The part that does not show up on any dashboard and makes this such an exciting moment to be inside an organization is the people involved.

In the mobile era, building was mostly left to specialists. This time, it's everywhere. People who have never written a line of production code -- marketers, analysts, operations leads, support managers, finance teams -- are prototyping agents, testing what AI can and cannot do, and building real intuition about where it helps. They are learning the grain of the technology by using it instead of sitting through a training deck.

That's a staggering amount of capability spreading through a company at once, and you couldn't have mandated it. No budget line buys AI literacy at this speed. The mobile boom taught a generation of teams to think about products and users. The agent boom teaches a far broader group of people to think about data, automation, and what good judgment looks like when a machine is doing the work. Human learning compounds alongside the data, and it may end up being the more valuable of the two.

You are counting the wrong thing

The metric to watch isn't the number of agents in production. This early in the market, that number tells you almost nothing.

The metrics that matter are quieter: how much cleaner and better governed your data is than it was a year ago, and how many people in the building know how to put AI to work. The organizations that will own the AI-native era are using this phase to understand their data landscape, raise their data quality and turn thousands of curious employees into confident builders.

Data-first thinking is key

The mobile revolution taught us to think mobile-first, but when a failed app vanished from the store, it took its lessons with it. However, the agent boom teaches us to think data-first, as the data work behind an agent gets reused, making the next ten agents faster, while the people who built it keep the skills for good.

So, when the next agent pilot gets shelved, log it as tuition paid toward becoming a data-ready, AI-literate company. After all, winners in the mobile era weren't the apps with the most downloads, but those who learned the most from them. The same is true in the agent era -- the winners will be those whose data and people are ready when it counts.

Stephen Catanzano is a senior analyst at Omdia, where he covers data management and analytics.

Omdia is a division of Informa TechTarget. Its analysts have business relationships with technology vendors.