# Fast-moving AI layers should stay replaceable

Markdown copy of https://realalpha.app/s/fast-moving-ai-layers-should-stay-replaceable/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

- Speaker: [Aaron Levie](https://realalpha.app/people/aaron-levie/) · Aaron Levie is cofounder and CEO of Box.
- Host: [Logan Bartlett](https://realalpha.app/people/logan-bartlett/) · Managing director, Redpoint Ventures · Backs enterprise software companies like Ramp, Cribl, Cyera and Legora.
- Show: The Logan Bartlett Show
- Episode: [Box CEO on Enterprise AI Trends No One is Talking About Yet](https://realalpha.app/v/g9E7iCnYOoI/), Mar 28, 2025
- Moment: 34:45 to 38:01 of the episode (3 min), playing on https://realalpha.app/s/fast-moving-ai-layers-should-stay-replaceable/; the episode from there: https://realalpha.app/v/g9E7iCnYOoI/?t=2085

## Quote

> “There's multiple parts of box where we use **multiple model providers** for the same workflow.”
> — [Aaron Levie](https://realalpha.app/people/aaron-levie/) at [36:52](https://realalpha.app/v/g9E7iCnYOoI/?t=2212) · The Logan Bartlett Show

## Takeaways

1. Lock in interfaces unlikely to change.
2. Keep uncertain layers like vector search and RAG replaceable.
3. Preserve a way to switch model providers.

## About this segment

Aaron Levie separates stable interfaces from unsettled parts of the AI stack, such as vector search and RAG. He says Box uses multiple model providers in some workflows so it can adapt as the field changes.

From “Box CEO on Enterprise AI Trends No One is Talking About Yet”, Mar 28, 2025. [Watch the full episode](https://realalpha.app/v/g9E7iCnYOoI/)

## Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

[34:57](https://realalpha.app/v/g9E7iCnYOoI/?t=2097) a tweet or something, but that being early to AI in some ways can be a liability if you don't set up the architecture with the right abstractions in some ways. I guess as you think about the applicability of this for Box as an organization and therefore the advice you might have for others. How do you think about being flexible with this pace of change? And is it more about being model agnostic or modular in elements of the design? What would you share on that?

[35:30](https://realalpha.app/v/g9E7iCnYOoI/?t=2130) Yeah, all of the above. So there's how we build and then how we even implement AI internally. how we build is, is basically we try and figure out the parts of the stack that there's probably a few different categories and I'm making up some of the, maybe the nomenclature on the fly, but like there'll be parts of the stack that just like, we'll never change, you know, a model breakthrough will happen as it doesn't matter. Like we're not going to change that particular interface, you know, with our system so that you can just like lock and load and start executing.

[36:01](https://realalpha.app/v/g9E7iCnYOoI/?t=2161) There's other parts where you're like, you know, we think there's a best in class technology now. You know, maybe we're kind of like, we're shaky on it. You know, we probably prefer there to be an open source version. So that's like a spot where you're, you know, you probably need to be really thoughtful about, you know, how much do you lean into a particular... particular part of the architecture. I'm thinking like your vector search, your rag kind of execution. These are spaces where there's a bunch of first movers, people started adopting them. It's not exactly clear where it all ends up in five years from now. So you kind of want to be thoughtful about that. And then there's the final part on your modular point, which is like, you know, this space is changing every three minutes. You, you, you need to decide, do you

[36:57](https://realalpha.app/v/g9E7iCnYOoI/?t=2217) want, do you want like maximum flexibility? So these are the model providers. Do you want to be able to have a way of, of, you know, swapping in different parts? There's multiple parts of box where we use multiple model providers for the same workflow. Um, uh, and so you just need to then have like a high degree of optionality, a high degree of flexibility in that part of the stack. And, um, uh, and so that, that, you know, so you, and then you have different like abstraction layers between those different interfaces. But I think basically anybody in AI at this point has been building with that pattern. You go to perplexity, you change your model, you go to cursor, you change your model. I think we're all realizing, okay, you probably want to let the user change their model, or at least you as the developer have some degree of optionality

[37:42](https://realalpha.app/v/g9E7iCnYOoI/?t=2262) there because we know how fast that space moves. And then you have to decide which other parts of the stack to just kind of, you know, lock in on and don't worry about again.