More From Green Room
Twenty five years ago, I was a founder of a logistics software company that was one of the very first SaaS businesses in our industry. We very much benefited from the technology architecture shift from on-premise / client server to cloud hosted / multi-tenant. We leveraged internet-based architecture to create a better application for our customers as compared to tools deployed on premise.
AI is the most impactful architecture shift since that .com transition. Our investment firm CreativeCo is thrilled by the opportunity to invest in new AI applications. We do not believe that software is dead. Rather, AI introduces a meaningful opportunity for all software companies to deliver more value to their customers and continue the long-term growth trajectory of the software market.
I’m an investor in all three public companies above — NVIDIA, Samsara and Toast. This image sums up the current software vs AI sentiment, with IOT down 50%, TOST down 30% and NVDA up 50%+ this year. The declines in software are driven by the narrative that AI LLMs will render software obsolete or a low cost commodity. The generalization is impacting all public software companies without discretion regardless of their positioning relative to AI.
I’m going to take a stab at sharing our point of view on this narrative. First, let’s look at the big picture:
Shift from on-prem to cloud. SaaS took hold in the 2000–2010 decade, winning over the largest application spaces with the advantages of cloud deployment and lower cost of ownership. Between 2010–2023, SaaS penetrated all niches of software applications to the point that some people talked about “micro-SaaS businesses” and we see system of record SaaS companies targeting incredibly small TAMs. After two decades of SaaS, attractive markets have been vended.
Shift from human users to AI agency. Neural networks / LLMs introduced in the 2022–2023 period represent the biggest application architecture shift since the beginnings of SaaS, changing the way applications can be built to deliver new levels of value proposition. It’s clearer by the month that AI will enable software applications that not only facilitate work, but actually do work with agency. As a result, the next generation of applications will create meaningful labor productivity, as they perform more of the roles of our traditional software users.
As a result of the AI innovation and the resulting software application driven productivity, we can expect the total global software market to continue to at least double every ten years — growing from about $800B in 2025 to $2 trillion in 2035. AI pundits would argue this forecast is conservative.
AI is part of the software stack. Software applications are a “stack” of enabling technologies. AI LLMs enable a new application architecture, incorporating AI into the stack. Below is a very simple representation of how these generalized layers work:
Software applications have been top of the stack, positioned as products that assemble underlying technologies into a solution for end users. With AI in the mix, software applications gain impressive new capabilities. These applications are technically still software as a service. Since they are enabled by AI, people could call them AI SaaS or Agent SaaS. At our fund CreativeCo, we’ve been seeing demos from AI software companies for the last two years, and one way to describe these products is simply next generation, new and improved SaaS.
The architecture change. As illustrated above, traditional SaaS applications could be depicted as layers: 1) user interface, 2) business logic, 3) database, and 4) integrations to other systems and data. With LLMs as part of the Agent SaaS stack, the application architecture shifts to a new set of layers. AI LLMs replace the traditional in-code business logic with intelligence. The LLM layer is guided/managed by a layer of workflow, prompts and evaluations to achieve the target outcome. And then the user interface design pattern changes from enabling humans to move the process forward to enabling humans to oversee and interact with agents similar to a co-worker/collaboration role.
Following this new architecture, OpenAI and Anthropic have released Agent SaaS capabilities to complement their LLMs with the goal of providing a generalized framework for putting AI agents to work. Anthropic CoWork and Plugins and OpenAI Frontier are clearly competing with traditional SaaS, but in a broad horizontal approach that would enable a business with an in-house technical team to use their platforms to build out internally-developed Agent SaaS.
It’s important to note that creating a robust Agent SaaS solution that delivers quality and accuracy across complex or broad tasks is not simple nor fast work. Therefore, we believe there is a tremendous opportunity to apply this new architecture and solve market-specific problems deeply and elegantly. These are the application opportunities we are seeking out for our CreativeCo fund.
Agent SaaS. There is a race to service obvious AI-enabled use cases: AI software coding, AI legal document creation, AI customer service, AI health care diagnosis, AI construction management, AI accounting for example.
Startups in these obvious Agent SaaS categories are growing at incredible rates, as the new level of value proposition unlocks expanded TAM, tapping into the value of labor. Software coding application startups like Cursor are displacing the value of labor in software development, and as a result growing into an enormous TAM — not just the value of the software application but more the value of the labor on top. As a result, the pricing of AI tools reflects this expanded value proposition.
We recently met with a startup with AI agents for paralegal work in a law firm, and in our diligence spoke to a lawyer that stated “I am AI all day long, and I’m spending $4,500 per month with Eve. It’s insane.” That $50k lawyer spend is 50%+/- the salary of a paralegal, so Agent SaaS startups are demonstrating that they are expanding the TAM by demonstrating labor replacement. Eve Legal has raised $164M and is valued at $1B and has 4x team growth in the last year to 230 employees.
Our fund CreativeCo is an investor in Aiwyn, an Agent SaaS company for accounting that has been working over the last couple of years to build an AI-first solution for tax preparation. The TAM for tax preparation is not only the traditional tooling from companies such as Thompson Reuters, but also the labor that has traditionally performed the tax prep work. It’s a huge opportunity for Aiwyn, but also a direct threat to non AI-first companies in the tax prep tooling space. Look at the Thompson Reuters stock chart above — they are traditional SaaS and information services in both accounting and legal — both labor intensive use cases perfect for AI agency. Aiwyn and Eve are startups demonstrating that the TAM is growing — but the market share will migrate to companies that leverage AI, and now Thompson Reuters is likely scrambling to innovate in this area.
The chart above was circulating X this week. It’s clear that the shift to AI agency will grow the total TAM pie. The opportunity for software application companies today is to embrace the shift to Agent SaaS and take share from slower more traditional SaaS movers. The shift and the competitive scrum is definitely underway, as all meaningful SaaS businesses are pushing into AI one way or another.
We do agree that if you are old school SaaS and choose not to embrace AI, then your product will find itself less competitive and potentially obsolete at some point, with time dependent upon your use cases and end markets.
The shift to Agent SaaS is use case dependent. It’s obvious that LLMs can write code and generate legal documents from templates. But can LLMs enable a better version of Toast, the leading application for restaurant operations management and payments? Half of Toast’s revenue is payment processing. The other half is SaaS for things such as reservations, payroll, websites, reporting etc. One could probably come up with some LLM-driven features that are slick for a restaurant, but how much incremental value can AI drive into a restaurant operation. Unlike coding, legal and accounting, the labor component of a restaurant operation is less replaceable by AI agency.
Yes, if you re-invisioned Toast in AI form, you could build Agent SaaS for this category, but we would argue that Toast is in the best position to make this shift happen and is certainly already heading in this direction. There are less obvious labor replacement use cases in restaurant operations and associated value proposition to drive Toast customers to switch to a new startup. Toast has over 100,000 restaurant locations as customers, is the system of record for running a restaurant, has a very broad product footprint, and has the resources to innovate AI into their solution mix.
The type of SaaS matters: Vertically-focused versus horizontal / Enterprise versus SMB. The production of code by AI is changing the dynamics of the classic build versus buy decision for software customers. We’re still in a place in time where you need experienced people to produce a high quality software application — experienced developers and designers coupled with people that truly understand the problem and solution. The amount of time required for this team to produce their initial application and then keep iterating is compressing quickly. We believe domain expertise and an understanding of quality output will keep skilled teams in the mix of product creation going forward, but certainly teams will do more with less resources.
Given this dynamic, we can assume that enterprises with large software teams will do more in-house versus buy externally. But enterprises will still make a buy versus build evaluation, and this equation comes down to the costs of the two options. Given this decisioning, simple horizontal SaaS tools and unduly expensive enterprise software are areas under pressure.
It’s also reasonable to assume that non-enterprise businesses without engineering resources would be unlikely to vibe code their own solution versus partner with a software vendor that focuses solely on that application as their core mission and provides it at a reasonable price. Therefore, vertically-focused solutions continue to outperform horizontal software in the public markets.
But these are all generalized assumptions — and the shift to Agent SaaS is happening at a pace that will continue to make the buy versus build equation challenging. New startups will emerge with dedicated and skilled product teams that dream up products better than what internal teams can do. This week, it was announced that Sequoia invested in an Agent SaaS business in the CRM space — one of the largest horizontal application categories. The threat to Salesforce is not vibe coders.. it’s talented new startups shifting to AI agency architecture.
Public and private multiples are back to normal. Given the AI architecture shift in software, public multiples have contracted dramatically from what was clearly a valuation aberration during Covid. For those that have been in the software business for decades, we’ve returned back to normal valuations, where the range has been somewhat consistent over time for good, better and best. We would argue that the current sell-off is overdone, but we’re not going back to the 2020–2022 period of time.
While public multiples have been back to normal, the growth rates of public cloud / software companies are declining as they have grown in scale over the last decade.
However, for startups tapping into the AI agency shift, the growth is unmatched in history. The chart below from Bessemer Venture Partners demonstrates that the time to reach $100M ARR for the fastest growing software companies has declined by 25% in the last decade, and more for AI centric startups that are tapping into the AI agency TAM opportunity.
This growth rate is rewarded in private company valuations. But even the fastest growing companies command more reasonable multiples than during the Covid peak.
There will not be 1–2 companies devouring the entire software market. OpenAI and Anthropic replacing all software applications is not logical. The total software market is expected to reach $2 trillion in size in the next decade. Today, the largest software company in the world is Microsoft with $240B in software revenue in 2025. This implies that Microsoft has a 30% share of the global TAM (which sounds high if you live in the fragmented application space every day). Meanwhile, Toast with about $2B in revenue has a quarter of a percent of the software market share. Perhaps the leading foundational AI models will have large market shares like Microsoft. But just like Microsoft and other big platforms, there is a tradition of distributing platform capabilities to smaller end markets through specialized software companies (like Amazon, Google and Microsoft did with cloud). We expect this platform distribution model to continue, with giant TAMs served directly where appropriate and more niche markets served indirectly through market-specific distribution as part of the stack.
The scale of AI growth is unprecedented. Given LLMs key role as part of the new software architecture, and the ability to displace labor value, AI foundational model platforms are growing at incredible pace.
The growth of Anthropic and OpenAI is unmatched. This growth does not happen without disruption to established market participants. In just three years, Anthropic is now larger than ServiceNow — an enterprise software company founded over 20 years ago. Anthropic raised $30B on a $380B valuation — 27x run-rate, which seems quite reasonable given their opportunity ahead. It is estimated that $2.5B of Anthropic’s run rate is generated by their Claud Code product, which was accomplished in one year. The balance of revenue is from paid subscriptions across the global business and consumer market. Unlike the .com shift that took years to monetize, the AI shift and value generation is moving at a rate that is challenging to grasp. This pace of change and monetization is both a risk and an opportunity for all participants in the software application market.
Meaningful architecture shifts always lead to better applications. The shift from human users working on applications to AI agency enabling labor replacement is moving at a faster pace than the move from on-prem to the cloud back in 2000–2010. But there will still be a transition period to leverage the shift, and as an investor and operator in software, I’m incredibly optimistic about the potential for software applications in the years ahead.
Twenty years from now, we will likely have AI that is intelligent enough to run a business — like a restaurant. But at the end of the day, businesses are built by people and relationships. We predict there will be teams of people that fine tune and package all of this future intelligence for solving problems they understand in their markets. These teams will work with their customers — who are also people with taste and a POV to create their own special business. We are moving forward with the optimism that there will always be a role for humans to build with creativity, create value, and remain in (and in control of) the loop.
.svg)
