TL;DR: Inspired by Microsoft Ignite this week – where Microsoft flagged "agent sprawl" as a growing source of shadow IT – we look at what happens when relaxed governance lets employees build thousands of agents faster than IT can track them. Now it’s time to ask what these agents are actually doing, who owns them once teams depend on them, and what happens if you turn them off. The Org Age of AI series is coming back!
I remember a moment at a huge conference hosted by one of the biggest enterprise technology companies. The organizers asked journalists and analysts to use their agent. It was terrible. It couldn’t find information about their own conference schedule. It was so bad that I felt embarrassed for them and wanted to apologize.
Since then, companies have continued building agents at astonishing speed. Some help employees find information; others draft answers, update records, or coordinate work across systems. And that brings us back to a question we asked in the first episode of The Org Age of AI. Why does AI feel so powerful while the ROI remains so difficult to find?
It’s no longer a question of whether people can build agents or whether they work. They clearly can, and they clearly do, and employees are discovering uses for them faster than governance can keep up. The harder question now is: how much of that experimentation should continue to be encouraged, when does it need to be brought under control, how do governance processes need to adapt, and is any of it making a meaningful difference to the bottom line?
In today’s episode we discuss:
When the same question has two different answers
Agent sprawl before AI agents
The people who owned the wires
From “this helps me” to “we depend on it”
How much sprawl is too much?
What can the agent actually do?
Take one invoice through the business
When the same question has two different answers
This spring, FICO CIO Mike Trkay told The Wall Street Journal that employees at the company were creating dozens of agents a day. One concern was that different agents could produce conflicting answers to the same business question. Meanwhile, GitLab CIO Manu Narayan told the Journal that he was comfortable with some short-term agent sprawl because of what employees might discover.
Let’s take both attitudes seriously. Employees should be empowered to do their jobs with these new tools, and they are in a position to experiment – and incentivized to – in a way that will never be captured by jockeying to get a slot on an engineering roadmap. Local efforts are narrowly optimized and suffer from not seeing the whole picture. Some of those experiments can become part of shared work, while some are redundant and others may be at odds with each other. Easy example: a sales team starts using an agent to prepare a report, another team builds its own version, and someone eventually has to reconcile the results.
Another example: one team counts a customer as renewed when a contract is signed, while another waits for the first payment. An agent built around either definition could produce a reasonable answer. If leadership needs one number, the company has to decide which definition to use and make sure both teams can access it. It’s a headache.
And this is the central tension of agent sprawl, where it becomes an organizational question. Companies want employees experimenting with tools that will improve output. But once the organization starts relying upon them, the outputs need to be reliable and consistent – which is a question of governance. The challenge becomes deciding when and how to apply governance in a way that doesn’t squash the experimentation that produced the useful ideas in the first place.
A lot of enterprise leaders are talking about "agent sprawl" right now. The irony? The term is over fifteen years old – and its history reveals exactly how this ends.
We unfold how agent sprawl became an enterprise problem and what companies from BBVA to Ryanair are doing about it. Then we walk through a real bookkeeping example and share four ways to assess an agent’s work, from human review to its effect on the bottom line.
Join Premium members from top companies like Microsoft, NVIDIA, Google, HF, OpenAI, a16z, plus AI labs such as Ai2, MIT, Berkeley, .gov, and thousands of others to really understand what’s going on in AI.






