Apollo GraphQL's CEO on Building for AI: It's Migration Velocity, Not Code Speed
I sat down with Matt DeBergalis, CEO of Apollo GraphQL, expecting to hear the usual AI hype. Instead, he said something that stopped me cold:
“If I had to bet on one thing that’s not going to change with AI, it’s that enterprises will still have silos.”
This from someone whose company powers APIs at Netflix, Zillow, and hundreds of other enterprises. Someone who’s spent the last decade helping companies integrate their systems. And he’s telling me that the legacy systems, the organizational boundaries, the decades-old APIs nobody wants to touch … simply aren’t going away.
He’s right. And the implications are more interesting than you think.
The hotel reservation system
Matt gave me the perfect example:
“If you go look inside a hotel, what you find is the room reservation system goes back to the Mesozoic era. It was never built for the web. It’s the thing the person at the desk is using when you show up. And so you have to adapt that system. It isn’t going anywhere. They’re not going to rip that out.”
This isn’t technical debt. This is business capability. The ability to assign hotel rooms is core to running a hotel. The fact that it lives in a system from 1987 makes it MORE valuable. It’s successful.
Here’s what makes this interesting: these systems aren’t just old, they’re stuck. “APIs sit around for decades,” Matt explained. “Just by their nature, because they’re a rendezvous point between software systems, they’re really hard to change.”
The more things that depend on your hotel reservation system, the riskier it is to touch. Every integration is another reason not to change it. Every team that built on top of it is another stakeholder who’ll resist the upgrade. The system is load-bearing.
And now AI is about to demand that these unchangeable systems connect to everything new you want to build.
AI Isn’t Breaking Down Silos. It’s Creating More of Them.
Everyone thinks AI is going to magically solve integration problems. Matt sees it differently.
“If you just think about what it means to arm every developer with a copilot or an autonomous agent, we’re going to get a bunch of microservices. We’re going to get a bunch of new consumers of APIs. There’s a giant integration problem because of that.”
Every team can now ship faster. Which means more services. More touchpoints. More things connecting to that unchangeable hotel reservation system. The first-order effect of AI is to enable creation, which means more mess to wire together.
And here’s the thing: the companies winning with AI are leaning into this. They’re exposing more surface area, not less. What matters is the new capabilities they can build, like chatbots that can answer questions in 15 seconds at 2am. That’s better than human support, and it requires a herculean effort to line up the pieces.
The winners aren’t the ones with the cleanest greenfield architectures. They’re the ones that can navigate complexity at velocity.
The job changed, you’re an architect now
Matt’s been using AI to write code since the early days. His take on what makes it work surprised me:
“It works to the degree that I have a systems-level understanding of what I want the code to do and how I want it to work. If I can be very prescriptive about where state lives in different parts of the app or what the API interface looks like… the rest of the code between those interface layers I don’t have to worry about so much.”
Read that again. The bottleneck isn’t writing code anymore. It’s designing systems.
He continued: “When you ask an agent to write code for you, the more that code can sit on top of strong abstractions, the better it works. It’s better to write concise code… if your options are writing a bunch of procedural code to call APIs in a certain order, but the alternative is just writing a three-line GraphQL query, you get better results with the query.”
This is the actual shift: developers are moving up the stack. You’re not going to be writing the same kind of code. You’re going to be making architectural decisions, defining interfaces, choosing abstractions. The agent handles the implementation between those layers.
“AI really rewards that kind of skill set in developers,” Matt said. “It’s a powerful tool in the hands of a developer that sees it that way.”
The question is: do you understand your system well enough to tell the AI what to build?
The migration window is now
MongoDB won the NoSQL wars despite being “really obviously the worst” database at the time. But they understood adoption better than anyone else. They knew how to sell to developers, both individually and their VPs.
Fast forward 15 years. If you picked one of Mongo’s technically superior alternatives back then, congratulations: you’re now the proud owner of a migration project.
AI is about to do this again, but faster. The models are trained on what’s popular on the internet. React. Python. The median tech stack. They’re getting absurdly good at these frameworks, which creates a compounding advantage.
“The models are trained on what’s on the internet, and they’ve gotten really good at the things that are most popular,” Matt explained. “There’s this interesting effect where it’s a virtuous cycle for those technologies. The models are really, really good at writing React code. And I think that has interesting implications for the stuff that isn’t as popular.”
What that means is: every quarter you stay on that niche or outdated framework, you’re falling further behind in AI leverage. The gap compounds quickly.
Migrations are the strategy
Migrations are how you get to the stack where AI works best. The companies that win aren’t the ones avoiding migrations. They’re the ones doing 10 a year instead of 2.
Because you don’t just need to upgrade React once. You need to connect that hotel reservation system to the chatbot, then the mobile app, then the partner API, then whatever comes next quarter. In a world where AI proliferates touchpoints and compounds advantages toward the median stack, migration velocity is the bottleneck.
So how do you get good at migrations?
You apply systems thinking to the migration itself. Matt’s insight—“AI works to the degree that I have a systems-level understanding of what I want”—applies here. Decide what “done” looks like. Encode the pattern. Validate it on 10 files. Scale to 500.
You’re not reviewing every change individually. You’re verifying the system that makes changes. Test the approach, trust the pattern, flag the genuinely hard 5% that needs human judgment. That’s how you go from a 3-month variable project to a 2-week predictable one.
You have maybe two years. Your enterprise silos aren’t going anywhere. Can you move fast enough around them?
You can find Matt at ApolloGraphQL or on LinkedIn.