Thomas Hobson ran a livery stable in Cambridge in the early 1600s. He kept a large stable and had one rule: you take the horse nearest the door, or you take none at all. Customers thought they were browsing. They were being handed a horse. The phrase outlived him by four centuries, and it now means the illusion of a choice that isn’t one.
I watched a version of this play out during a sprint.
The ask was simple: Get an app stood up, fast. Someone on the team was moving quickly, and I asked what I thought was a throwaway question, “Aren’t you using the company’s agent?”
They weren’t. Neither, it turned out, were several others.
At first, I thought training was the issue. Maybe we hadn’t onboarded people well enough and they didn’t know what the platform could do, so they reached for what was familiar. That explanation held for about a week.
As it turns out, they didn’t know that if GPT wasn’t working, they could go to Claude or Gemini, natively, inside the sanctioned platform. The choice was already there. Going outside was actually more work, because they couldn’t give an unsanctioned agent the context that made the sanctioned one useful in the first place. They were paying a tax to escape a constraint that no longer existed.
That isn’t a training gap. Training tells you what a tool does. It doesn’t tell you the tool is yours.
The Three Months That Cost a Year
Here is the part that should worry every organization rolling out an internal AI platform.
When the agent first launched, there was no choice. One model. GPT, or nothing. That was the moment teams left. Not gradually, not after a fair trial. Right then.
The ability to choose between models shipped about three months later. A real fix, shipped fast by most enterprise standards. It landed in silence.
By then everyone had already tuned out. They weren’t evaluating the new release and rejecting it. They never saw it. They had stopped reading the announcements, stopped checking the changelog, stopped treating the internal platform as a live thing that might be different this month than last. The org kept improving a product for an audience that had stopped listening.
This is the second-order failure, and it’s more expensive than the first. A bad launch costs you a launch. A bad launch that goes uncorrected in the minds of your users costs you every improvement that follows it. The platform team sees flat adoption on a genuinely good feature and concludes the feature missed. It didn’t miss. It was never seen.
Three months of silence undid a year of roadmap.
Technical Teams Left First, and the Rest Followed
The sequence matters.
Technical teams went first, for code. They had the least patience for a model that wasn’t fitting the task and the most awareness that alternatives existed. Non-technical teams followed, for presentations and documentation. Different work, same conclusion, arrived at a few months apart.
Where they went splits into two patterns. Some bought their own subscriptions outright. Others simply picked a different platform, or a different model on a different platform, and worked there.
Neither group was being reckless. Both were solving for the same thing: get the work done with a tool that responds the way I need it to respond. Okta’s 2026 research found roughly half of employees are using unapproved AI tools. That number gets reported as a security problem. It’s also a product feedback signal, and it’s the most honest one your organization will ever receive. Nobody fills out a survey to tell you your internal tool doesn’t fit. They just stop opening it.
Autonomy Isn’t a Feature Flag
So why did shipping the choice not bring anyone back?
Because a choice you’re handed is not a choice you made. People want options they feel they selected, not options that appeared in a dropdown after they’d already decided the tool wasn’t for them. The autonomy has to arrive before the disengagement, not after it.
I’m not exempt from this. I read Gemini’s output more easily than Claude’s for certain tasks. That’s a preference, not a technical judgment, and I’d want it respected too. So would you. PwC’s 2026 AI Jobs Barometer makes the point at scale: as AI absorbs routine work, judgment and human factors matter more, not less. That applies to how people choose their tools, not just how they use them.
The mistake wasn’t launching with one model. Constraints at launch are defensible. The mistake was launching a constrained tool without telling anyone it was the first version of something, and then assuming the fix would find its own audience.
What Actually Reconnected People
Not a memo. Not a mandate. Not another enablement deck.
Leadership, IT, and managers went and talked to their teams. Then they ran a survey that asked people how they were actually using AI, not how they were supposed to be. The results showed the real picture: which tools, which tasks, which teams, and why.
That insight changed the cadence. Training moved from monthly to weekly, because monthly was never going to catch a workforce whose habits were forming in days. And the communication went both directions, with concerns raised and answered rather than logged. Technical and non-technical teams got different sessions, because they had left for different reasons and needed different things to come back.
None of that is sophisticated. It is exactly what evolv’s mission describes, an authentic focus on relationships, applied to a platform rollout instead of a client engagement. The survey worked because people believed someone would read it. The trainings worked because they answered questions people had actually asked.
Take the Horse Nearest the Door
Hobson’s rule worked for him because his customers had nowhere else to rent a horse. Your employees have somewhere else. They have a credit card and a browser tab, and they will use both the moment your platform feels like a stable with one horse in it.
The instinct behind a governed internal AI platform is right. Context, security, and cost control are real, and the people who went outside paid for leaving in extra work they didn’t need to do. But governance that arrives as constraint gets routed around. Governance that arrives as capability gets adopted.
If you are rolling out an internal AI platform, ship the choice on day one, even if it’s a smaller choice than you’d like. Tell people what’s coming next. And when adoption goes flat, ask before you assume, because flat numbers can mean your feature missed or they can mean nobody was left to see it, and those two problems have opposite fixes.
Bottom line: your people didn’t reject your AI platform. They stopped watching it. Win the attention back before you ship anything else.
Need help with your AI strategy? Connect with evolv today.
Nicolás Castex Giménez is data and analytics leader with 8+ years of experience driving enterprise digital transformation through cloud modernization, data engineering, and business intelligence. He specializes in building scalable data ecosystems, modernizing analytics platforms, and translating complex technical solutions into measurable business outcomes. His experience includes leading cross-functional teams, developing data governance strategies, and enabling data-driven decision-making using GCP, Snowflake, Power BI, and Tableau. Nicolás is fluent in English, Spanish, French, and Portuguese.
