The Breaking Point Between Your First AI Win and Your Twentieth

 By Katy Feldner, VP, AI Transformation & Product

The fifth initiative is usually where it breaks. Or the twentieth. That’s when the person who did the diligence on the first one is gone, and nobody thought to ask what she knew before she left.

I can predict which number it’ll be almost every time. Not because initiative five is worse than initiative one. Because the judgment that caught a bad bet on day one — whether the data could support the decision, whether the workflow had a real decision in it, or whether the cost matched the priority — lived only in her head. It doesn’t hand in a two weeks’ notice. It just stops showing up to the next steering committee.

Andy wrote about the piece of this that compounds, and it isn’t data. As he put it, “the first AI initiative you deploy is the hardest. The second should be faster.” I’d add one caveat to make  that true: the second is only faster if the organization wrote down what the first one taught it before the person who learned it moved on.

He calls what gets written down context. Not the data itself. This is historical knowledge that never lived in a system to begin with. How an exception gets handled. What’s genuinely fixed in a contract and what’s negotiable in practice. Which shortcuts a claims reviewer takes that never made it into a formal policy document.

I ask the same three questions on initiative twenty that I asked on initiative one. Is the data usable? Is there a real decision here? Does the cost match the priority? They shouldn’t get harder to answer over time. In fact, they should get easier, because someone already answered a version of them once and wrote it down. Most organizations don’t write it down, so initiative twenty starts exactly where initiative one did — an expensive mistake. Caused by someone who wasn’t in the room for the lesson the company already paid to learn the first time.

This is where the stakeholder problem lives, and it’s not the trust story people default to. Clinical, ops, and finance each hold a piece of the answer the other two don’t have. A clinician knows which triage calls need a human to override the model. Finance knows which cost assumption was wrong on the last build. Ops knows which “fixed” process was never fixed. None of that is about whether people trust AI. It’s about whether anyone bothered to write down what those three people know before one of them changes jobs.

Get this right, and the returns start looking like Andy’s version of it. Initiatives that used to take six months compress toward six weeks, not because the models improved but because the organization stopped relearning what it already knew. Get it wrong, and initiative twenty will cost almost exactly what initiative one did, no matter how improved the underlying model is.

During the 2026 US Open, I was watching and pulling harder than usual for my fellow Florida Gator, Ben Shelton. He didn’t win — Alexander Zverev took it in four sets — but that wasn’t the part that stuck with me. Getting to the final meant ousting the defending champion, Carlos Alcaraz, in a late-night quarterfinal that ended at 3:33 a.m., the latest finish in US Open history. Shelton is only 23, and he’s the first Black American man to reach a US Open final since Arthur Ashe in 1972. He calls himself “a late bloomer,” the kind built by learning from every win and loss along the way. 

You don’t beat Alcaraz in a match that runs past 3 a.m. because your serve got faster this season. You beat him because everything you’d ever learned about exhausting matches (both mentally and physically) is still there when you lose the first set in a close tiebreaker. The Alcaraz win and the Zverev loss are both what he draws on next in his career. That’s compounding when nothing gets lost.

I keep coming back to the two questions this series started with. What counts as value, and what it takes to keep counting after the first win is proven. 

I’ve come to think the second one is harder, and it was never a measurement problem, or even a technology problem. It’s whether an organization built something that remembers what it learned, or is just hoping the next person figures it out. That’s our case for treating ROAI^TM as an operating model instead of a one-time proof — and it’s the same case we’ve been making before any of this had a name.