Welcome to the Logan Bartlett Show where today's episode features a conversation with Aaron Levy, co-founder and CEO of Box. This discussion covers the current state of enterprise technology and artificial intelligence, exploring Aaron's insights on what his customers are thinking about AI.

The conversation spans multiple topics related to AI including perspectives on closed source versus open source models, the future opportunities in the enterprise space, and the business model shifts occurring with usage-based pricing.

"We are probably overindexed on the flashiest parts of AI in the consumer world and then underindexed on the things that will just make hundreds of billions of dollars in B2B."

Logan introduces this as a "really fun conversation" that delivers valuable insights on the enterprise AI landscape.

Timestamp: [00:00-01:07]

Aaron Levy provides a technical definition of what constitutes a "model company" in today's AI landscape. He argues that there aren't many pure "model only" companies, but rather companies that have models along with other offerings.

Making a clear distinction, Aaron explains that most organizations in the AI space are not purely model companies: "If I'm being extremely technical about the definition, there's probably not that many like model companies."

He identifies Mistral as perhaps the only true model company, while clarifying that organizations like OpenAI and Anthropic would better be classified as "AI foundation frontier lab companies" that have a broader portfolio including "products and models and services and just like all the things."

This distinction helps frame the conversation about where value is created in the AI ecosystem and how different companies position themselves within it.

Timestamp: [01:10-02:19]

Aaron Levy addresses Logan's question about whether companies can build long-term value around model offerings for enterprises, or if that value will ultimately be consolidated within cloud platforms like AWS, Microsoft Azure, or GCP.

Aaron expresses skepticism about building a massive business solely around AI models, citing a fundamental challenge: "You basically are dealing with this issue where all of the hyperscalers at a minimum... need to make sure their models get used at an unbelievable scale."

He explains the economic pressure these large players exert on the market:

"It is so existential to them to have their models get used at massive scale that they will constantly drive down the price of AI and the quality up in AI that you would just get squeezed... into such a difficult position."

Aaron believes that independent AI companies would struggle to compete in this environment dominated by tech giants that have both the resources and strategic necessity to win the AI infrastructure war. This creates a challenging competitive landscape for standalone AI model providers seeking to serve enterprise customers.

Timestamp: [02:19-03:44]

Aaron Levy shares revealing insights about the rapidly shrinking gap between open source and commercial AI models, based on Box's internal testing with Google's Gemma model.

"We ran these tests on Gemma from two weeks ago... the new open source model from Google. Not to be confused with the Severance character Gemma."

Aaron reveals that in Box's evaluations, Gemma performed "only like a couple points worse than Gemini" for accurate data extraction from enterprise documents and data. This finding has significant implications for enterprise AI strategy.

The rapid advancement of open weights models reaching near-parity with commercial offerings means enterprises will soon face a choice between using open weights models they can self-host or commercial models with managed infrastructure. The decision will increasingly be less about model capability and more about operational preferences.

Aaron questions how a standalone company could succeed in this environment, competing against tech giants like Alphabet, Meta, Microsoft, OpenAI, and Anthropic. He suggests that to be viable, companies would need to build comprehensive software around their AI models—technology that enables practical application of the models rather than just "selling tokens back and forth."

This convergence of performance between open and closed models raises existential questions about the sustainable business model for independent AI providers in the enterprise market.

Timestamp: [03:44-05:40]

Logan asks Aaron whether OpenAI should continue pushing the boundaries of frontier models or redirect energy to consumer products like ChatGPT. Aaron endorses OpenAI's current approach, noting he doesn't have "much critical feedback for OpenAI or any of the major players at the moment."

He identifies interesting strategic questions facing other tech giants: "What does Google do about core search?" and "What does Apple do about the Siri problem?" These questions revolve around whether companies build proprietary AI capabilities or leverage external models.

Aaron sees OpenAI's strategy as creating a productive flywheel rather than separate workstreams: developing the best models in the world, incorporating them into consumer and business products, and now exploring applied use cases through agents. He emphasizes that maintaining frontier model capabilities is crucial to OpenAI's business position:

"It's critical that they are always at the forefront of the state-of-the-art models because that's the calling card into the business conversation."

Aaron shares an interesting thought experiment about whether users would notice if ChatGPT's underlying model were swapped with Gemini or DeepSeek. He suggests we've reached "convergence on model quality for most consumer use cases" but follows with an important caveat about the 80/20 rule:

"We probably have solved like 80% of 'hey what's the score to this sports game, tell me about World War II'... but the 20% is going to still be orders of magnitude more quality that we need, which is like 'go out and book a flight'... and those we're nowhere close to solving."

This insight highlights how quickly AI has solved obvious use cases while the remaining challenges will "take years and years to keep cranking through."

Timestamp: [05:40-08:46]

When asked if any AI agents have achieved product-market fit today, Aaron identifies several areas where agents are already delivering value:

1. Outbound sales: Aaron mentions 11x (a company he invests in) as an example of AI agents successfully augmenting sales capacity and generating demand. He notes that "lead gen" appears to be a clear area where agent technology has found traction.

2. Coding agents: Aaron cites examples like Cursor, Windsurf, Replit, and Boltit, noting that while 80% of AI coding might be "more fun and interesting to watch and not going into production," there's significant product-market fit for internal IT tools and applications that previously might have gone to low-code platforms or contract development.

3. Document data extraction: Box's own agentic use case extracts data from documents like contracts and invoices, putting information into databases and automating workflows. Aaron notes this has clear product-market fit: "Mostly our problem is we can't keep up with all the range of use cases customers want."

Aaron contrasts this business success with consumer agents, where he sees limited product-market fit so far. He then shares a perspective that has been consistent in his thinking:

"We are probably overindexed on the flashiest parts of AI in the consumer world and then under indexed on the things that will just make hundreds of billions of dollars in B2B."

This insight suggests that while consumer AI applications may capture public attention, the most significant economic value creation may happen in enterprise contexts through more practical, less flashy applications focused on specific business problems.

Timestamp: [08:46-11:32]

• Few "pure" model companies exist today - most organizations like OpenAI and Anthropic are "AI foundation frontier lab companies" with broader offerings beyond just models

• Hyperscalers (AWS, Azure, GCP) have an existential need to drive AI model adoption at massive scale, creating pricing pressure that makes it difficult for independent AI model companies to build sustainable businesses

• Box's testing shows open source models (like Google's Gemma) performing nearly at parity with commercial models (like Gemini) for enterprise use cases, suggesting rapid commoditization of base model capabilities

• The choice for enterprises will increasingly be about operational preferences (self-hosted vs. managed) rather than fundamental model capabilities

• We've reached "convergence on model quality" for 80% of common consumer use cases, but the remaining 20% of more complex tasks will take years to solve

• AI agents have achieved product-market fit in specific B2B contexts like sales outreach, coding assistance, and document data extraction

• The market may be "overindexed on the flashiest parts of AI in the consumer world and underindexed on the things that will just make hundreds of billions of dollars in B2B"

• For OpenAI, maintaining frontier model capabilities is strategically crucial as "the calling card into the business conversation" even as they build consumer products

Timestamp: [00:00-11:32]

Companies & Organizations:

• Box - Aaron Levy's company, enterprise content management platform
• OpenAI - Mentioned as an AI foundation frontier lab company with products, models and services
• Anthropic - Mentioned alongside OpenAI as not a pure model company but more comprehensive AI company
• Mistral - Identified as perhaps the only "true model company"
• Google/Alphabet - Referenced regarding their AI models (Gemma, Gemini) and search strategy
• Meta - Mentioned as one of the major AI players
• Microsoft - Referenced as hyperscaler with AI capabilities
• AWS - Mentioned as a hyperscaler cloud platform
• Azure - Microsoft's cloud platform referenced as a hyperscaler
• GCP - Google Cloud Platform mentioned as a hyperscaler

AI Models:

• Gemma - Google's open source model that performed nearly as well as commercial offerings
• Gemini - Google's commercial AI model
• DeepSeek - AI model mentioned in comparison to other models
• ChatGPT - OpenAI's consumer product mentioned in strategic context

AI Companies & Products:

• 11x - AI company in outbound sales space (Aaron is an investor)
• Cursor - Coding agent mentioned as having product market fit
• Windsurf - Coding agent mentioned as showing promise
• Replit - Coding platform with AI capabilities
• Boltit - Coding agent mentioned as showing promise

Media & Entertainment:

• Severance - TV show referenced when clarifying Gemma is not the character from this show

Timestamp: [00:00-11:32]

Logan asks Aaron whether usage-based pricing with AI has the potential to be as disruptive as the transition from on-premise to SaaS was historically. He notes that while people appreciated the delivery mechanism of SaaS, the pricing model's disruptive power may have been underappreciated.

Aaron agrees that usage-based pricing is highly relevant in the AI era. He characterizes the SaaS and cloud business model disruption as fundamentally transforming the economics of enterprise software:

"It took something that you used to spend upfront like maybe millions of dollars on to now you spent thousands of dollars on and the capex to opex transition."

He provides a vivid example of deploying a CRM system in the 1990s, which required substantial upfront investment:

"If you were to deploy a CRM system in the 90s, just think about what had to go into that - data centers, system integration... you bought software for the CRM system but you still had to buy hardware, you had to buy all the servers, you had to have a network."

Aaron contrasts this with Salesforce's revolutionary approach:

"Salesforce shows up and they're like literally like tomorrow you can just have a CRM system and oh by the way you could use it for 20 people in your company... Just like start using it."

This transformation democratized access to enterprise software, expanding the total addressable market by making technologies accessible to businesses of all sizes. The combination of seat-based pricing, subscription models, and dramatically lower upfront costs created the conditions for SaaS to revolutionize enterprise software.

Timestamp: [11:45-15:21]

Aaron identifies two major areas of upside that AI agents bring to business models and market economics:

First, AI effectively automates forms of labor, dramatically expanding the total addressable market (TAM) for software:

"It expands the TAM of software from only being able to sell the software to people on the other end. So you were kind of capped by the number of seats to now you have kind of like... categories where the category of AI agent spend will be an order of magnitude more than the spend of the software in a non-AI world."

Aaron illustrates this with contract management software as an example:

"Contract management software as a category is like a couple billion dollars globally in the world. Legal services is a couple hundred billion dollar category. So can software take 10% of that spend through AI and pull from the legal services spend or expand the legal services spend? And then all of a sudden this category that was relatively niche... could that now be a $20 billion category as opposed to a couple billion dollar category?"

He suggests this pattern could repeat "20-50 times across software," underwriting massive TAM expansion by automating tasks previously done by human labor.

The second transformative aspect is that AI enables product capabilities that were previously impossible, unlocking new budget and value creation:

"AI has a way of just completing the things that people have always wanted to do in the products themselves but just were never possible, which then also unlocks a whole bunch of new budget."

This creates a new form of disruption where software can "complete the outcome that the customer was looking for" without requiring customers to deliver it themselves with their people, allowing more value to accrue to the software provider through usage-based or outcome pricing.

Timestamp: [15:21-17:34]

Logan raises concerns about the potentially deflationary impact of competition among AI companies, wondering if the opportunity to capture service budget might be limited not by customer willingness to pay but by competitive pressure from the many venture-backed companies entering the space. He asks Aaron about durable value creation, using Zoom as an example of a service with high intrinsic value but price constraints due to competition.

Aaron responds pragmatically, emphasizing that market dynamics remain unchanged in the AI era:

"I don't think that any law of market dynamics changes because of AI."

He points out that competitive pressure has always been a reality in technology markets, referencing the CRM market of the 1990s:

"I bet there was a CRM market map from Gartner in '96 that was like 47 companies that did CRM and somehow only Siebel and like one or two others made it to the other end."

While acknowledging that competition does suppress what companies can "maximally charge," Aaron outlines how successful companies eventually differentiate themselves:

"Eventually you get some degree of network effects, some degree of reference customers that tell each other that you can all use X thing."

To illustrate how companies break out despite crowded markets, Aaron shares an anecdote:

"My dad was asking how did Wiz get sold for $30 billion, and we're Wiz customers and I think the technology is fantastic. How does that happen in a world where obviously there's like 500 security startups?"

His answer reveals the compounding advantage that creates durable winners:

"There's a little flywheel where like your product's just 20% better, then you get a few more reference customers, and then you learn faster, and then your product gets another 20% better faster... and it kind of spins out and then all of a sudden you have one with escape velocity outpacing the rest of the category."

To maximize moats in AI, Aaron advises following timeless principles from enterprise software:

"The things you probably want to do to maximize your moats are the same lessons probably from 35 years ago in enterprise software, which is get data, do workflows, make sure you've got enough happening within your ecosystem."

This perspective suggests that while AI creates new opportunities, the fundamentals of building defensible business advantages remain consistent with historical patterns in enterprise software.

Timestamp: [17:34-20:58]

• The SaaS revolution transformed enterprise software by converting large upfront capital expenditures into affordable operational expenses, dramatically expanding market access

• AI creates two major areas of business model disruption: automating forms of labor (expanding TAM) and enabling previously impossible product capabilities

• Contract management software illustrates the potential TAM expansion: the category is currently worth a few billion dollars, while legal services is a hundred-billion-dollar category - AI could help software capture 10% of that larger market

• Aaron believes this pattern could repeat "20-50 times across software" - creating massive new market opportunities as AI takes over tasks previously done by human labor

• Usage-based or outcome-based pricing models align with AI's value proposition as they let customers pay for completed outcomes rather than just software access

• Despite competitive pressures, successful AI companies will likely follow similar patterns to previous enterprise software winners: creating slight advantages that compound over time

• The fundamentals of building defensible businesses remain unchanged in the AI era: "get data, do workflows, make sure you've got enough happening within your ecosystem"

• Companies that build products just "20% better" can create virtuous cycles: better products → more customers → faster learning → even better products → escape velocity

Timestamp: [11:45-20:58]

Companies & Products:

• Salesforce - Referenced as revolutionary in disrupting the CRM market with SaaS pricing
• Slack - Mentioned as an early innovator in "fair use pricing"
• Siebel - Referenced as one of the few survivors from the early CRM market
• Oracle - Mentioned as acquiring Siebel
• Zoom - Referenced as a product with high intrinsic value
• Wiz - Security company mentioned as selling for $30 billion despite competitive market

Business Concepts:

• SaaS (Software as a Service) - Central to the discussion about business model disruption
• Cloud - Referenced alongside SaaS as transforming business models
• Capex to Opex transition - Key concept in how SaaS changed enterprise software economics
• Usage-based pricing - Discussed as potentially disruptive pricing model for AI
• Outcome pricing - Referenced as an emerging model for AI products
• Total Addressable Market (TAM) - Discussed in context of AI expanding markets
• Network effects - Mentioned as factor in creating durable businesses
• Reference customers - Noted as important for building momentum in enterprise sales

Industry Categories:

• CRM (Customer Relationship Management) - Used as example of software category transformed by SaaS
• Contract management software - Used as example of category that could expand dramatically with AI
• Legal services - Referenced as large market that AI could partially disrupt
• Security startups - Mentioned as a crowded field where companies like Wiz still managed to stand out

Organizations:

• Gartner - Referenced regarding historical market maps of the CRM industry

Timestamp: [11:45-20:58]

Aaron continues his discussion about creating durable value in AI by emphasizing how network effects and compounding value will drive competitive advantages—similar to earlier enterprise software generations.

"There's a compounding value more data that comes into the system, the next run of the AI agent gets just a little bit smarter. There's going to be a lot of network effects on the embeddings of your data."

He illustrates this with a powerful example contrasting individual versus organizational adoption of AI coding tools:

"As a single programmer, you could probably switch between Cursor, Windsurf, and Replit and a few others, but as a network of 500 engineers in a company and I've done embeddings on the whole codebase and I'm talking to all the different systems it's wired up to... change management on that becomes a lot harder."

Aaron believes the fundamental rules of building successful enterprise software businesses will apply to AI:

"I think it's going to look a lot like SaaS and the lessons are going to look a lot like SaaS. You want data, you want workflows, you want strong customer references, and you want to build a network effect around your customers."

He adds one important caveat about the current pricing dynamics in AI:

"There might be some pricing dynamics right now in play in AI that won't necessarily last forever, which is right now you can kind of comp it to labor in some areas. That probably feels like a little bit of a temporary advantage that you might have, when ultimately it'll probably be more comped to typical software margin structures."

This transition from labor-comparable pricing to software-comparable pricing represents a maturation journey the industry will undergo as it discovers the "terminal business model" for AI.

Timestamp: [21:04-22:34]

Logan raises an interesting question about whether AI solutions might face lower switching costs than traditional SaaS products, particularly those that are less workflow and UI-centric.

He compares AI applications to security solutions, where "the work is getting done oftentimes outside of those screens" by ingesting and processing data. Logan suggests this might result in lower friction when switching between providers:

"The things that are just tapping into work automation, be it customer support or SDR or whatever it is, ultimately because there's less workflow embedded into it, the change from one to the next I think there's less friction associated with that change."

Logan contrasts this with traditional project management tools:

"What's the difference between Asana, Monday, and Trello? Well, the difference is you got to switch your entire team onto something different in a meaningful way, versus when the work is actually being done outside of the UI, it might just be one-click change."

Aaron offers a fascinating counterpoint, suggesting that AI systems might actually create stronger lock-in through learning and data accumulation:

"The question would be like is there going to be an analogy where it's inverted... it's the APIs into your other systems, it's the APIs into your e-commerce thing, it's then the logic around when to make the decision about the e-commerce thing."

He argues that the self-reinforcing nature of AI systems could create significant switching barriers:

"Sure, you could switch but then it's a cold start to start over because... I could underwrite the exact opposite of this point, which is that when you have a learning system, that learning system will probably perform better than your people did."

Aaron explains that unlike static software where people's knowledge is the primary asset, with AI "it's going to track every event it's done in history and just get a little bit better... and then that becomes some opaque data source to you the customer that is inside this AI product that just makes the thing always perform five or 10% better."

This creates a novel form of customer lock-in that doesn't depend on user interface familiarity but rather on accumulated learning and optimization.

Timestamp: [22:34-26:29]

Logan references an anecdote about Aaron giving a talk on cloud technology to a disinterested audience, asking him to elaborate on that experience and contrast it with current AI adoption trends in enterprise.

Aaron recounts speaking at a financial services event about cloud technology:

"In the peak of cloud adoption, I did this keynote on the future of running your bank or financial services, and it was mostly just the pitch cloud... I find it very exciting, but the audience didn't because it was just like 'yep, cool, we get it man, the servers are not in our data center anymore, they're now in a cloud data center.'"

He explains that this reaction captured the essence of the cloud transition—it made IT operations faster, enabled more data storage, and improved employee efficiency through better software, but "it didn't radically transform entire industries or sectors."

In stark contrast, Aaron observes that AI "has the opportunity to actually radically transform how you run your business, how you make decisions, how you serve your customers, how you build products."

While implementation challenges remain similar—both cloud and AI require "the same change management... the same meetings... the same privacy and policy and governance and compliance review"—Aaron identifies a critical difference in initial reception:

"In the very early days of cloud you would go into a meeting with a bank, a pharma, a government agency and you'd say 'I'd love to talk to you about moving your data to the cloud.' And it was sort of like shock, like definitely not gonna happen here... maybe there'll be some small little workload that's like a dev test workload."

The contrast with AI adoption is dramatic:

"We're only about two and a half years into ChatGPT... If I look at AI right now, that exact same meeting with a customer or set of customers is like 'we know this is going to change everything. Everybody's clamoring for it. Every employee needs it. The new generation coming into the workforce is asking for it. My boss is asking for it. The board is asking for it.'"

With AI, enterprise conversations have shifted from "if" to "when and how":

"It's not if, it's when, it's how. It's like what could we do? I need to do more yesterday."

Aaron concludes that enterprise has "bypassed all of the normal resistance" that typically accompanies technological transformation, accelerating the timeline "five to seven years faster than we did with cloud." However, he cautions that actual implementation will still follow similar trajectories due to the realities of organizational change.

Timestamp: [26:29-31:04]

• Creating durable AI businesses will follow many of the same rules as traditional SaaS: accumulate data, build workflows, establish customer references, and create network effects

• Current AI pricing that compares to labor costs represents "a temporary advantage" that will likely evolve toward traditional software margin structures as the market matures

• While individual users may easily switch between AI tools, organizational adoption creates stronger lock-in through system integrations and accumulated learning

• AI creates a paradoxical switching cost dynamic: less UI-centric products might seem easier to replace, but accumulated learning and data integration can create stronger lock-in than traditional software

• Unlike static software where user familiarity creates stickiness, AI systems create lock-in through continual improvement based on historical usage patterns and data

• Cloud adoption faced significant initial resistance from enterprises, with acceptance coming gradually over many years

• AI adoption has "bypassed all of the normal resistance" with enterprises immediately asking "when and how" rather than "if"

• Despite faster initial acceptance of AI compared to cloud (by "five to seven years"), actual implementation will still face similar organizational hurdles

• AI has the potential to "radically transform" business operations in ways cloud technology did not, which explains the dramatically different reception

• The key question for enterprises has shifted from whether to adopt AI to more nuanced challenges like determining appropriate human oversight and implementation strategies

Timestamp: [21:04-31:04]

AI Tools & Companies:

• Cursor - AI coding tool mentioned when discussing switching costs
• Windsurf - AI coding tool referenced alongside Cursor
• Replit - Coding platform mentioned in discussion of AI tools
• Kodium - AI coding tool referenced in discussion of switching behavior

Traditional Software:

• Asana - Project management tool mentioned when discussing UI-driven switching costs
• Monday - Project management tool compared to Asana and Trello
• Trello - Project management tool used as example of traditional software with high switching costs

Business Concepts:

• Network effects - Central concept discussed as creating defensibility in AI businesses
• Embeddings - Technical concept mentioned related to AI's data processing and organization
• Cold start - Problem referenced when switching between AI systems that have accumulated learning
• Change management - Process discussed as similarly challenging for both cloud and AI adoption
• APIs - Application Programming Interfaces mentioned as creating integration points for AI
• Human-in-the-loop - Concept mentioned regarding AI implementation strategies

Industry Categories:

• Customer support - Mentioned as area where AI automation is being applied
• SDR (Sales Development Representative) - Role mentioned as being augmented by AI
• E-commerce - Referenced as domain where AI integration creates switching barriers

Organizations & Sectors:

• Financial services/banks - Referenced regarding cloud and AI adoption patterns
• Pharmaceutical companies - Mentioned in context of enterprise technology adoption
• Government agencies - Referenced regarding traditional resistance to cloud technology

Technologies:

• ChatGPT - Referenced as marking the beginning of the current AI wave (2.5 years ago)
• Cloud - Extensively compared to AI in terms of adoption patterns
• SaaS - Software as a Service referenced as paralleling AI business models

Timestamp: [21:04-31:04]

Aaron continues his comparison between cloud and AI adoption, revealing an interesting reversal in customer behavior patterns:

"If I compare to cloud, cloud could do more than what customers were asking for because they were resisting it. So you could have always implemented cloud 50% more than you actually were at the time."

With AI, he observes the opposite dynamic:

"Most conversations now, the customer is actually jumping like three steps ahead once they see the demo and they're like 'Oh that's great. We'll just totally automate that entire thing.'"

This enthusiasm often outpaces current capabilities, with Aaron noting they sometimes have to temper expectations: "Let's slow down... your data is not yet in the environment that it's going to need to be to do that use case" or "we're still waiting on some reasoning model breakthroughs from OpenAI or Claude to get that extra 3% of accuracy."

Aaron acknowledges his perspective may be skewed by the forward-thinking customers who engage with Box: "I'm talking to the people that are talking to us which means that by definition they're leaning in." He estimates there's still "80% of customers that are like 'whoa, my head's exploding, where should I start?'"

However, for many organizations, especially those at the forefront of adoption, "they're now ahead of the technology based on the set of use cases that they have." This represents a significant shift from cloud adoption, where the technology was typically ahead of customer readiness.

Timestamp: [31:10-33:03]

Aaron explains that the primary barriers to AI adoption aren't resistance but practical considerations that enterprises must navigate:

"I can't just turn on an AI thing that looks at all my data and then an employee can ask a question because guess what, it's all of a sudden going to reveal secrets in the organization that they shouldn't have had access to but they were overprovisioned for access like five years ago."

This data governance challenge is just one of several real-world implementation hurdles enterprises face. Another significant challenge involves determining acceptable AI autonomy in critical systems:

"AI agents are really good at doing code auto-completion, but am I ready to put an AI agent in production for generating a full multi-file output? And then how do I review that when the human actually wasn't along the ride for the actual editing process?"

Aaron highlights the contrast between the experimental excitement in the AI community and the careful consideration required in enterprise settings:

"Imagine how unhinged we are online where we're like 'Oh man we just made a video game in three hours'... and then you're an engineer in a banking system. Are you supposed to go implement that right now? Are we ready for your bank account to be driven by vibe coding? We don't know."

These questions represent the core challenges enterprises are grappling with: determining appropriate implementation pace, assessing data and system readiness, and making future-proof decisions in a rapidly evolving landscape.

"The industry is moving so fast. How do I make decisions right now that I'm going to be stuck with in a fast-changing space?"

This leads many organizations to focus on creating adaptable architectures: "Can you get your architecture set up for optionality in the future? Because maybe the vendor you work with today is not going to be the one that's the best in the future."

Timestamp: [33:03-34:58]

Logan references a previous caution from Aaron that being early to AI can be a liability without the right architectural abstractions. He asks for Aaron's perspective on building flexible AI systems that can adapt to the rapid pace of change.

Aaron breaks down his approach to AI architecture at Box into several categories:

1. Stable components: Parts of the stack that will remain consistent regardless of model breakthroughs:

   "There'll be parts of the stack that just like will never change. A model breakthrough will happen and it doesn't matter... we're not going to change that particular interface with our system."

2. Evolving technologies: Areas where current solutions may be temporary:

   "We think there's a best-in-class technology now... we'd probably prefer there to be an open source version. That's a spot where you probably need to be really thoughtful about how much you lean into a particular part of the architecture."

   He cites vector search and RAG (Retrieval-Augmented Generation) as examples where the landscape is still evolving: "There's a bunch of first movers... it's not exactly clear where it all ends up in five years from now."

3. Maximum flexibility components: Parts requiring frequent updates:

   "This space is changing every three minutes... Do you want maximum flexibility? These are the model providers. Do you want to be able to have a way of swapping in different parts?"

Aaron reveals that Box uses multiple model providers for the same workflows in some cases, requiring "a high degree of optionality, a high degree of flexibility in that part of the stack" with different abstraction layers between interfaces.

He observes this pattern has become standard practice across the industry:

"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."

This approach recognizes that users increasingly expect model choice: "We're all realizing you probably want to let the user change their model, or at least you as the developer have some degree of optionality there because we know how fast that space moves."

The key architectural challenge becomes determining which parts of the stack to make flexible and which to standardize and "lock in on."

Timestamp: [34:58-38:00]

Logan brings up a notable example of AI adoption causing industry buzz: Clear CEO Sebastian Siemiatkowski tweeted about consolidating SaaS vendors and moving to AI-centric workflows, which "set off a little bit of a firestorm" raising questions about whether agentic workflows might eventually replace traditional software interfaces. Logan asks if Aaron sees this as broadly applicable or as an exceptional edge case.

Aaron finds the Clear example "fascinating" and values that "we do need people testing the boundaries," but he doesn't see it as signaling the end of SaaS. He shares an analogous story from his past:

"I was meeting with an IT leader at Tesla and they were building their own ERP system, and I was like 'Holy crap, I cannot believe that you're doing this'... This is sort of an Elon special where he needed to control end-to-end the whole thing... The amount of fortitude and just sheer entrepreneurialness you need to have to even want to attempt that is kind of insane."

He places Clear's experiment in a similar category—interesting but not representative of most organizations. Aaron emphasizes a critical responsibility gap in the agentic approach:

"The AI provider that made the agent that wrote your HR system is not taking responsibility for the HR system code that was generated. So that means by definition you, the customer, are now responsible for whatever the thing is that it produced."

He elaborates on the accountability implications:

"It's not going to be the AI models. Claude is not—Anthropic is not taking any responsibility for how your HR system works, and if it leaks data and you get sued by the EU because there's some kind of data privacy thing... that's Workday. Workday will do that, not Anthropic."

Aaron identifies specific scenarios where agentic workflows might be viable:

1. Areas where users are deeply dissatisfied with existing software
2. Systems requiring extensive customization
3. Non-strategic applications where temporary "micro apps" are sufficient:

   "Which parts of software are so unstrategic that you don't mind whipping up these micro apps for one-off things... that you can just deploy quickly and turn off... it needs to live in the world for only three months."

For the next two years, Aaron believes only these niche cases could be disrupted: "I don't think your HR system is going away. I don't think your CRM system is going away."

Timestamp: [38:00-41:56]

• Unlike cloud adoption where technology capabilities exceeded customer readiness, with AI many customers are "jumping three steps ahead" of what the technology can currently deliver reliably

• Most adoption barriers are practical rather than philosophical: data governance issues, determining appropriate AI autonomy, and creating responsible oversight mechanisms

• Organizations face the challenge of making architectural decisions in a rapidly evolving landscape, requiring thoughtful consideration about which components need flexibility and which can be standardized

• Box recommends a three-tiered architectural approach: stable components that won't change, evolving technologies requiring careful consideration, and maximum flexibility components (like model providers) that need built-in optionality

• Using multiple model providers for the same workflow has become a standard practice across the industry, reflecting the need for adaptability and choice

• Clear's experiment with replacing SaaS applications with agentic workflows represents an edge case rather than a mainstream trend, similar to Tesla building its own ERP system

• A critical accountability gap exists with agentic workflows: AI providers don't take responsibility for applications generated by their models, leaving organizations fully accountable for any issues

• Agentic workflows may be viable primarily for three scenarios: areas with high user dissatisfaction, systems requiring extensive customization, and non-strategic "micro apps" with short lifespans

• For the next two years, Aaron believes core enterprise systems like HR and CRM will not be replaced by agentic approaches due to complexity and accountability requirements

• The industry-wide focus has shifted from "if" AI should be adopted to managing practical implementation challenges like data readiness, system integration, and governance frameworks

Timestamp: [31:10-41:56]

AI Companies & Models:

• OpenAI - Referenced regarding model breakthroughs needed for certain accuracy levels
• Claude/Anthropic - Mentioned regarding model capabilities and responsibility boundaries
• Perplexity - Cited as example of application allowing model switching
• Cursor - Mentioned as platform allowing users to change models

Companies & Organizations:

• Box - Aaron's company, referenced throughout regarding their AI architecture approach
• Clear/Klarna - Company that experimented with replacing SaaS applications with AI-generated solutions
• Tesla - Referenced in anecdote about building custom ERP system
• Workday - Mentioned as example of SaaS vendor that takes responsibility for their product
• European Union (EU) - Referenced regarding data privacy regulations and accountability

People:

• Sebastian Siemiatkowski - CEO of Clear/Klarna who tweeted about AI-centric workflows
• Elon Musk - Indirectly referenced regarding Tesla's approach to building custom systems

Technical Concepts:

• Vector search - Mentioned as evolving technology in AI stack
• RAG (Retrieval-Augmented Generation) - Identified as area still developing in AI architecture
• Abstraction layers - Discussed as architectural approach for managing model flexibility
• Model providers - Referenced regarding the need for optionality in AI systems

Business Systems:

• ERP (Enterprise Resource Planning) - Referenced in Tesla anecdote
• HR systems - Discussed regarding viability of replacement with agentic workflows
• CRM (Customer Relationship Management) - Mentioned as system unlikely to be replaced by AI agents

Business Concepts:

• Micro apps - Term used for temporary, single-purpose applications that could be AI-generated
• Agentic workflows - Central concept discussed regarding AI systems operating with some autonomy
• SaaS (Software as a Service) - Discussed throughout regarding potential disruption by AI

Timestamp: [31:10-41:56]

Aaron continues his discussion of agentic workflows in enterprises by emphasizing that even if someone builds "Oracle with AI agents," the result would still be a software company using AI for engineering augmentation. He maintains his perspective that enterprises don't want to be responsible for maintaining such software:

"I don't think the customer wants to be responsible for that software at the end of the day."

He references the concept of "core versus context" (which he attributes to Geoffrey Moore):

"Context is just like that... you just want done for you. You don't want to worry about it. Core obviously being like this is existential to my business. And for most people, their core business is not their HR system."

This distinction explains why enterprises outsource non-core functions rather than building them in-house, even with AI assistance:

"That's why you don't really want your internal IT team being responsible for rebuilding something that you can get off the shelf and just it's done for you."

Logan agrees with this perspective, adding: "That's the way I feel whenever technical founders want to reinvent sales comp. It's just like... it's hard enough to differentiate on one vector. This isn't going to be the make or break."

This exchange reinforces Aaron's view that while AI will transform many aspects of enterprise software, it won't fundamentally change the outsourcing dynamic for non-core business functions. Companies will continue to prefer purpose-built solutions for these areas rather than building their own, even with the assistance of AI agents.

Timestamp: [42:03-43:41]

Logan shifts to a rapid-fire format, asking Aaron several quick questions about AI's impact on work and organizations:

Will AI decrease knowledge worker hours in 5-10 years?

When asked if AI will reduce the number of hours worked by the average knowledge worker in the next 5-10 years, Aaron answers: "I'm gonna stick with no for now." Logan agrees with this assessment.

Will AI flatten organizational structures?

On whether AI will flatten organizational structures by reducing middle management or enabling more self-service work, Aaron is more optimistic:

"I think yes. And part of it is because I think that AI kind of lets functions sometimes do their adjacent function. So it lets the designer write the front-end code. It lets the backend developer write the front-end code. It lets the copywriter generate the full white paper."

Aaron believes this will create more "full stack workers" who can handle a broader range of responsibilities:

"I think we've gotten maybe overly disciplined in job functions in some parts of the economy. And I think AI kind of lets you have a little bit more expansive responsibility. So then that probably means that's a little bit of a flattening when you don't have to hop through and around as many orgs to get things done."

Logan connects this to David Epstein's book "Range," which explores how generalists often drive breakthroughs by applying knowledge across domains. Aaron agrees that work has become "a little bit too Adam Smith's" with excessive specialization:

"I think work has probably gotten a little bit too Adam Smith's. And we've like all just doing our little part of the paperclip factory... I think maybe we get a little bit of a chance of a reset, which is like no, I can just throw a little bit more of a general problem at somebody and they've got the ability to go and fully solve it, which frankly I think is a hugely exciting thing and much more fulfilling probably for the knowledge worker class."

This perspective suggests AI may reverse some of the hyper-specialization trends in knowledge work, potentially creating more satisfying and holistic roles.

Timestamp: [43:41-47:00]

Logan asks Aaron about the most mind-blowing AI demo he's seen in the last three to six months. Aaron's response is candid and reflective, acknowledging the challenge of maintaining a sense of wonder in the face of rapidly advancing AI:

"I'm having a funny journey because you do start to now unfortunately build such high tolerance and desensitization. It's like the hedonic treadmill in some ways."

He references a Louis CK comedy bit about people complaining about airplane Wi-Fi despite the miracle of flying through the sky with internet access, drawing a parallel to how quickly we adapt to AI breakthroughs:

"That's the way I feel about a lot of the AI things where Deep Research hallucinates a little bit and I'm like... it only saved me seven hours."

Aaron expresses nostalgia for his initial reactions to AI:

"I wish I could have my pre-ChatGPT wiring for every individual breakthrough because I think I would just be like 'Holy—this thing is witchcraft' like every day."

He recounts his early amazement with large language models:

"I remember yelling at a friend who was deep in LLMs like a week into ChatGPT and I was like 'I'm sorry, I literally don't understand how did it write the business strategy on a thing it's never scanned from the internet?' And like I was just obviously an idiot at that point on all this stuff... but I want that energy again where I'm just surprised by everything."

Despite his "desensitization," Aaron highlights two AI applications that still impress him:

1. Research assistance: "From a pure utilitarian standpoint, I still love Deep Research... I had a thing 48 hours ago. It saved me an hour and a half... I checked a couple of the answers. Fantastic... I'm now doing that probably five to 10 times a week on something, which is just an awesome productivity boost."

2. Front-end design creation: "I'm still incrementally surprised by and excited by front-end design creation with AI. I'll pop in a screenshot into Cursor and it codes it, or I'll pop in a prompt into v0 (Replit) and it produces something that I would have had to ping an engineer or a designer about... My browser history kind of looks a little unhinged because it's like 11:30 p.m. on Saturday I'm on v0 doing designs."

Aaron's response reveals how even as we become accustomed to AI capabilities, certain applications continue to deliver value and occasionally reignite that sense of wonder.

Timestamp: [47:00-50:21]

Logan asks Aaron whether multimodal capabilities (image, text, code, voice) or agentic workflows will have a greater impact on enterprise AI. Aaron's answer is definitive:

"If you have perfectly executing agentic workflows, that's the holy grail."

He elaborates using a data science analogy to explain the transformative potential of agentic AI:

"In data science as an example, and anything obviously that you put into a structured database, we've always had this ability to be like 'okay, I'm going to go compute some large set of things... we would talk about like 'I have a job running and I'm gonna check back in in an hour and figure out what the answer was.'"

Aaron then extends this concept to unstructured knowledge work:

"Imagine in all forms of unstructured knowledge work a world where you just have jobs running where you're just like 'I just sent these 10 agents to go and figure out and synthesize all of this clinical drug trial research to give me insights.'"

He emphasizes the scale of this potential transformation:

"We were only wired to think about sending off jobs and compute tasks for like 3% of all corporate work. Now imagine if you do that for the other 97% or at least 80% of the 90%."

Aaron provides compelling examples of how this might work in practice:

"I sent out a job to go research this new region I want to enter as a business. I sent out a job to go and figure out what all of my product feedback was and then build me a road map that I should go look at and decide if it's validated or right."

He concludes that the ability to execute "full agentic workflows" — whether for research, task execution, or customer experience automation — represents "the biggest area of upside" over the next two, five, or ten years.

Logan then asks whether Aaron believes the best models will be closed source/proprietary or open source in those time periods. Aaron suggests the distinction will become less relevant:

"For lack of having any imagination on this, I think I would mostly be in the camp of... it almost doesn't matter because within 3 months of the breakthrough you'll have an open source version."

This rapid knowledge transfer between closed and open source means that "by definition you can almost still expect then a large amount of the traffic and the use will be open source because you're not trading off that much because you'll still get the open weight version of what just emerged."

Timestamp: [50:21-53:22]

Logan asks Aaron about the most exciting AI features that Box has rolled out which might be underappreciated by their user base. Aaron highlights two features that he's particularly excited about:

The first innovation addresses a fundamental file organization challenge:

"If you go to Box and you go to any file system, you create a folder, you put files in the folder, you share that with people... you have this problem which is... folders are kind of unintuitive for the recipient because they're like 'I don't know which third subfolder down did you put the sales material that I'm looking for to use for my customer pitch.'"

Box Hubs solves this by creating an intelligent overlay on the folder structure:

"We created Hubs, which is an overlay onto folders. So you can put as many of your files and folders into a Hub, and then you can go and just ask questions of the Hub."

This approach creates contextual knowledge centers that reduce ambiguity and improve search precision:

"If you go to a general purpose just AI chat window, you have to do a lot of work to guess what stuff is on the other end of this chat window. If you go to a Hub in our case, you kind of know by Hub what it has access to. So we have a sales Hub which means by definition that has all the sales information. We have an HR Hub which means it has all your HR information."

This contextual knowledge approach dramatically improves the quality of results:

"You go to the HR Hub and you ask the HR question, and then we're basically doing RAG on your documents for all of your HR documents... That basically gives you these micro knowledge portals for every topic that you want."

Aaron emphasizes the core advantages of this approach:

"The user instantly knows what topic is going to be covered in this thing. So you dramatically reduce the hallucination. You dramatically increase the authoritative content that will be in the Hub."

Aaron's second highlight is Box's document data extraction capabilities:

"The final one that is just working out kind of way better than we had hoped initially... is just AI data extraction from documents. It sounds way too straightforward for anybody listening, but it's insanely powerful if all your job is like review contracts, review invoices, pull out data from resumes, standardize my financial reporting documentation."

The impact is significant: "We can just now automate basically most of that."

This practical application of AI to everyday document processing tasks represents the kind of "unsexy" but highly valuable AI implementation that creates immediate business value by automating routine work.

Timestamp: [53:22-55:55]

• The "core versus context" principle remains relevant in the AI era: enterprises will continue to outsource non-core functions rather than building them in-house, even with AI assistance

• Aaron doesn't expect AI to significantly reduce knowledge worker hours in the next 5-10 years, but does believe it will flatten organizational structures

• AI enables workers to handle "adjacent functions" (designers writing front-end code, backend developers handling front-end tasks, etc.), potentially reversing decades of increasing job specialization

• Work has become "overly disciplined in job functions" and AI can enable more "full stack workers" who can handle broader responsibilities, making work more fulfilling and reducing organizational silos

• AI capabilities quickly become normalized, creating a "hedonic treadmill" effect where revolutionary technologies rapidly come to seem ordinary

• For Aaron, research assistance (saving hours weekly) and front-end design generation remain particularly impressive AI applications despite growing desensitization

• "Perfectly executing agentic workflows" represent the "holy grail" of enterprise AI, extending the concept of background computing jobs from data science to all knowledge work

• Aaron envisions sending "jobs" to research business opportunities, synthesize feedback, or build roadmaps—applying the compute job model to the "97%" of work that hasn't previously been automatable

• The distinction between closed and open source models is diminishing, with open source versions appearing within months of major breakthroughs

• Box's AI strategy focuses on context-specific applications like "Hubs" that create micro-knowledge portals for specific domains, reducing hallucinations by narrowing the context

• AI data extraction from documents is "insanely powerful" despite seeming straightforward, enabling automation of contract reviews, invoice processing, and financial reporting tasks

Timestamp: [42:03-55:55]

People:

• Geoffrey Moore - Referenced regarding the "core versus context" concept
• David Epstein - Author of "Range" book discussed regarding generalists vs. specialists
• Adam Smith - Economist referenced regarding specialization of labor
• Louis CK - Comedian referenced regarding bit about airplane Wi-Fi and taking technology for granted
• Andrew Ross - Briefly mentioned regarding "Predictable Revenue" concept

Companies & Products:

• Box - Aaron's company, with specific mention of their AI features
• Box Hubs - AI-powered knowledge portal overlay for Box folders
• Oracle - Referenced hypothetically in discussion about rebuilding enterprise software with AI
• Workday - Mentioned as example of HR system vendor
• Deep Research - AI research tool mentioned as saving significant time
• Cursor - AI coding tool that can generate front-end designs from screenshots
• v0/Replit - Front-end design generation platform from Replit
• ChatGPT - Referenced regarding early AI impressions and timeline

Concepts & Terms:

• Core versus context - Business concept (attributed to Geoffrey Moore) regarding which functions are essential to a business
• RAG (Retrieval-Augmented Generation) - Technology mentioned in context of Box Hubs
• Hedonic treadmill - Psychological concept referenced regarding desensitization to AI advances
• Multimodal AI - Capabilities spanning image, text, code, and voice processing
• Agentic workflows - AI systems that can autonomously execute complex tasks
• Hallucination - AI phenomenon of generating incorrect information, mentioned regarding Box Hubs reducing this issue

Books & Media:

• "Range" - Book by David Epstein about how generalists triumph in a specialized world
• "Predictable Revenue" - Business book indirectly referenced

Business Functions:

• HR system - Mentioned as example of non-core function companies outsource
• Sales compensation - Mentioned as area technical founders sometimes try to reinvent
• Front-end design - Area where Aaron finds AI particularly useful
• Data extraction - Key AI feature at Box for processing documents like contracts and invoices

AI Model Types:

• Open source models - Discussed regarding their rapid ability to replicate closed source capabilities
• Closed source/proprietary models - Discussed in comparison to open source alternatives

Timestamp: [42:03-55:55]

Congratulations on making it through all the knoweldge insights of Aaron Levy's AI wisdom! That's like completing a mental marathon without the knee pain.

You've just consumed more practical AI insights than 99% of LinkedIn thought leaders will post this year. Your brain is now officially Box-certified in enterprise AI strategy.

"The best time to learn about AI was when ChatGPT launched. The second best time is right now. - Ancient tech proverb we just made up"

We hope these insights help you navigate the AI landscape with the confidence of someone who knows the difference between RAG and a regular old rag.

Remember: In a world of AI hallucinations, you're now grounded in reality.