When I joined JPMorgan from Google in 2018, the first thing I had to unlearn was how I thought about AI.
At Google, AI was the magical pill that delighted your customer. Everyone cared whether it made the product better and didn’t think twice about how it worked. At JPMorgan, that thinking is a non-starter. There's no concept of a black box. There's no concept of "it works 80% of the time." It either works 99% or 99.99% of the time, or it doesn't work. A point decimal change in Excel destroys careers. AI can't be meaningfully impactful inside a firm like that unless it earns the same level of trust.
This nuance took me years to fully internalize, and it's the single biggest reason most AI programs running inside enterprises today are going to fail.
Walk into almost any large enterprises today, and you'll find an AI strategy spread across 50 use cases, each one promising 10% productivity gains.
I call this the peanut butter approach. You spread AI thinly across every workflow, every desk, every department. It looks responsible. It generates great slides. It does almost nothing to the economics of a private equity deal.
When you have transformational technology at your fingertips, incrementalism is a death knell. You drive a step-function change in a business by picking one or two needle-moving problems and going all in.
This was the lesson I learned at Silver Lake. The first piece of advice Mark Gillett, who ran value creation at Silver Lake, gave me when I joined to lead AI across the portfolio was simple. Take one portfolio company, go all in, prove the model, and only then think about replication. Don't hire a team. Don't try to apply AI to call centers across 15 portcos and report 10% productivity gains. That makes for a great board update, but it doesn't move value creation.
So, that’s what I did, and in about nine months, I had proof where we are able to fundamentally transform the content generation and monetization workflow of a portfolio company using AI generating >$500M in enterprise value. And this approach has shaped everything I've done since.
Two months into my role at JPMorgan, I was asked to present my AI roadmap to Jamie. The meeting was scheduled for 30 minutes. I'm a Googler at heart, an engineer, and a builder. I don't walk into rooms with slides, so I walked in with a demo my team had pulled together.
That meeting went three hours.
What struck me were the questions Jamie was asking. At first, I thought he was testing me, or checking whether the bank had hired the right person. He wasn't. He was asking, with what I can only describe as childlike curiosity. Where does this technology break? How do I use this to better manage my risk? Where is the black box? These were first-principles questions about where the technology would fail, not where it would succeed.
That's the question most institutions get wrong when they evaluate AI. They start with "what can it do?" The leaders who drive transformation start with "where does it break, and what does it cost me when it does?"
Two lessons stayed with me from that period.
The first was AI transformation is top-down, but AI adoption is bottom-up. The CEO has to mandate the direction. The teams have to own the execution. If either side is missing, the transformation stalls.
The second was build capability, not just bespoke use cases. Lori Beer, Global CIO, taught me that. The proof point of an AI program isn't the press release or the feature launch. It's the platform underneath that makes the next 50 use cases possible. We built that platform at JPMorgan. Most of the team I hired is still there. Most of what we built is still in production.
When I started Obin AI, we made a decision that surprised some people. We wouldn't chase the CIO or CTO as our primary buyer.
With due respect to the CIOs and CTOs we work with (they're critical to every engagement), deploying agentic AI inside a regulated workflow isn't an IT decision. It's a strategic decision about how the firm operates. If the CEO isn't driving it, we're not the right partner. And the firm isn't ready.
This is also why I think most AI deployments in enterprises fail. The work only gets done when someone presents a use-case map. This means first, someone has to run a small pilot, then six months later, the firm has learned something interesting but has moved nothing. The CEO is the only person who can authorize the kind of focused, high-conviction bet that changes how the firm operates.
In the next post, we’ll dive into what AI transformation looks like in practice now and in the next 5+ years.
Listen to the full conversation with Marc Andrew on Modern Capital: The Private Markets Podcast here.
