Service is the new SaaS
All Things Venture #100
There are two moments in technology history that every investor should have seared into their brains. Those moments are: January 23rd, 1993 (the beta release of the Mosaic browser) and November 30th, 2022 (the commercial release of ChatGPT). These two dates mark the rise of two successive technological mega trends, the internet and AI, and from 1993 to 2022 the world saw profound, consistent, technology driven change. Culturally, politically, and economically technology amassed power and significance with relentless velocity. Nation states emerged from HTML and CSS, entertainment platforms became matters of national security, and the weight of the global economy is now concentrated within 10,000 square miles.
As we close the door on the era of the internet and step into the age of AI,the next 30 years will represent an even more significant change than ever before. I think that AI is my generation's (shout out millennials) version of the internet and the closest technological trend we will experience where there is a clear before and after.
My operating thesis is simple: as the technology industry continues to improve the reasoning capabilities of large language models and drive toward AGI, we will enter a new paradigm wherein services businesses enjoy a renaissance. Software will no longer operate as a service, software is the service.
I believe that large language models and the software workflows they are packaged in are increasingly developing a competitive advantage over existing service providers.
In other words, LLMs and software workflows are developing a competitive advantage over humans.
They are developing competitive advantages over you and me, and all of our peers (I’ll talk about the early evidence of this a bit later on). I know that it’s easy to view this ambition as purely the realm of science fiction, as the fantasy of software engineers and tech billionaires the world over, but we are barreling into a world where the constructs of intelligence: memory, reasoning, planning, and the ability to understand the physical world are being encoded into ones and zeros. If the engines of capitalism work as designed, at some point my job as a VC won't rely on code, it will be executed in code. And knowing this, I believe it’s in all of our interests to act and adjust accordingly.
More and more companies are not only willing to experiment with AI, but they are willing to push AI native applications into production. For example, earlier this year Klarna shared some incredible statistics about the performance of their AI chatbot, which is, “doing the equivalent work of 700 full-time agents,” and is “estimated to drive a $40 million USD in profit improvement to Klarna in 2024.” If we assume that the $40M profit improvement is driven by the replacement savings associated with either removing 700 full time agents, or foregoing their costs, then what Klarna just publicly stated to the world is that a roughly ~$60,000 per year service job can be done more accurately, more quickly, and more precisely by software.
Similarly, you could turn to DLA Piper, the third largest law firm in the world. It’s been reported that DLA is working with C3 AI to automate aspects of due diligence on limited partner agreements and that they’ve been able to, “reduce the time it takes to create over 200-point due diligence analyses of limited partner agreements, and [they] reduced the effort by 80%” Klarna and DLA Piper’s experiments aren’t just isolated incidents either, Morgan Stanley, Bain, Bloomberg, and more have all publicly announced initiatives that are leveraging AI. The interest that these companies are showing in adopting AI is indicative of the potential of the technology, and while it’s probably good to apply a healthy dose of skepticism to self reported statistics, I am on the side of believing that they are directionally correct. This directionality tells us two things 1) We are living in an unevenly distributed and uncertain future. 2) Given this uncertainty, and the magnitude of change ahead, we should be prepared for what’s next. The promise of AI is to enable service providers to do their jobs better, faster, and cheaper than ever before, and there is a growing mountain of evidence that AI’s promise will become reality.
In regards to the timing of this reality, it may come sooner than we think. For example, you could look to the field of radiology where an LLM based application operated certain tasks faster than humans and at 99% of the cost. You could look to the software industry itself where software developers that leverage LLM based copilots report higher task completion rates, greater productivity, and were described as being able to “complete [tasks] significantly faster - 55% faster than the developers who didn’t use GitHub Copilot” You could look to the consulting industry where consultants using AI have been characterized as “significantly more productive” and able to, “complete 12.2% more tasks on average, and completed tasks 25.1% more quickly [with] more than 40% higher quality compared to a control group.” You could look to the freelance marketplaces where in the 8 months after ChatGPT was released professional services related jobs (i.e writing, software development, website & app development) saw between a 10% - 30% decrease in demand relative to less AI sensitive opportunities. You could look to the legal profession where GPT- 4 was shown to enable law students to complete entry level legal taks between 12 - 32% more quickly than a control group and the “lowest-skilled participants saw the largest improvements.” You could look to the dark web where, “New York University researchers have tested out how well GPT4 can perform in hacking competitions and discovered it is better than 88.5% of human players” demonstrating “evidence that today’s frontier language models are capable of augmenting and accelerating hackers”
So given this evolving body of evidence, how should we be thinking about what’s next? How should we be thinking about a future where software’s domain of being better, faster, and cheaper is applied to white collar work?
Personally, I believe the best way to have a clearer view into the future is to look toward the past. Which brings us to the man who helped kickoff the first of our two watersheds, Marc Andresseen. In the seminal essay, “Why Software is Eating the World” Marc Andresseen addresses the turning point that was the internet and focuses on industries. He writes, “Over the next 10 years, I expect many more industries to be disrupted by software, with new world-beating Silicon Valley companies doing the disruption in more cases than not… Companies in every industry need to assume that a software revolution is coming.” It’s obvious today that Andressen was correct.. 28% of US equity value is concentrated in just seven technology companies. This is the incredible feat of the internet, disruptive power and wealth creation on an unprecedented scale.
However, the fundamental difference between the internet revolution and the current AI revolution is the atomic unit of value that will be disrupted. Counter to Andreeseen’s observation in 2011, industries are not at stake. Internet search, whether it is dominated by Google, Microsoft, Perplexity, or some other upstart, will continue to be a going concern. E-commerce will be some multiple higher than its estimated $6T market size today. We will still rely on taxis to get from Laguardia to Manhattan, whether that taxi is an Uber, a Lyft, or a yellow cab.
Today, industries are safe.
In contrast,
Services are not. I’ll explain.
At its core, a professional services job is a series of manual data translation tasks that 1) have different degrees of complexity, and; 2) are dependent on different amounts of unstructured data. These jobs combine the canonical traits of human intelligence: memory, reasoning, ability to plan, and understanding of the physical world. Let’s bring this to life.
What does a banker do all day? She likely begins the morning reading the Wall Street Journal, the New York Times, or some other media publication. Some information is valuable (committed to memory), some is discarded. Later, she may conclude based on prior information, that one of her clients has a window of opportunity to purchase a competitor (reasoning). Upon this realization, she may decide to discreetly staff junior bankers on creating a presentation (ability to plan), and from there she makes the decision to call her client (understand the physical world) to set up some time to meet. She may be in the same social circles as her client and be attending a wedding or similar social event that weekend, so the client suggests they convene there. When the gathering occurs, the banker realizes a competitor of hers is also at the event (and similarly wants her client’s business), so she must adapt in due fashion to preserve her business (all of the aforementioned skills).
Whether it is a banker's pitch deck or a lawyer's memo the core function of the professional servant is to translate the myriad amounts of unstructured data they receive into structured outputs and to communicate in a manner that their client (or their boss or their colleague, or whomever the powers that be) will understand, all while maintaining and building interpersonal relationships. This is the bare metal of professional services, this is the atomic unit of value at risk.
Professional services are omnipresent in the American economy, and in many ways the lifeblood of the modern world. We rely on professional services every day, and they range from the specialized white collar careers such as lawyers, bankers, and consultants that are industries unto themselves, to the process oriented teams of customer service, finance, and HR that are spread horizontally across the Fortune 500. Professional services exist across the healthcare value chain like radiologists, medical affairs teams, or medical coders, and are deeply embedded into seemingly niche verticals like renewable energy development or my recent favorite, global compensation analysts (aka HR for HR). In any event, these jobs represent a significant portion of the global workforce and in the US alone I estimate that there are roughly 27M professional services jobs worth $2.3T in aggregate wages.
As the evidence of frontier language models capability grows (as noted earlier), entrepreneurs are increasingly taking notice of the opportunity at play as well. The proverbial game is afoot.
For example, entrepreneurs are building in financial services where startups like Portrait Analytics, Lumosity, Quilt Labs, and Brightwave are looking to bring step functions in efficiency to the investment research process. Entrepreneurs have crowded the legal field with the likes of Harvey, EvenUp, Casetext, Noetica AI, Leya, Solomon, Draftwise, Atticus AI and Casehopper all competing for supremacy in various portions of a lawyer’s workflow. Entrepreneurs are building for blue collar service industries like automotive repair and the trades where companies like Revv and Dashlar, are opening up new revenue opportunities for local entrepreneurs, or supercharging their ability to find and serve profitable customers. Life sciences entrepreneurs have set up shop and are bringing novel data sets and automation to different parts of the value chain, and have started startups like Atrix AI, Century Health, Octozi, and Convoke.
You could look to Durable which is automating website creation. You could look to Tome which is automating presentations. You could look to Fortune 500 companies and see that they are using companies like Synthesia to leverage generative AI based videos for their global learning and development modules. You could look to main street businesses like restaurants and notice that they are using Slang AI or Momos to automate their customer service. You could look to the insurance industry and encounter EvolutionIQ handling claims automation, you could look to the banking industry and encounter Interfold AI or Tidalwave automating commercial loan and mortgage origination, you could look to healthcare and see Abridge streamlining the clinical documentation process or Abstractive summarizing clinician notes.
These companies are the canary in the coal mine of the transition from software as a service, to software as the service. If we assume that the complexity of service industries like software engineering, radiology, and legal services are representative of the service industry overall, it is clear that we are at an inflection point. These three fields alone employ nearly 2.5M people, account for $300bn+ in aggregate wages, and are being affected by AI today. Knowing this and believing that we are at an inflection point, I think the questions to ask are will this technology continue to improve? At what rate? And why?
These are all difficult questions, and I wish I had the answers but I don’t. What I do have is an informed perspective and the benefit of history. For example, in an effort to answer the question of will this technology improve we can look back to the summer of 2020. During this time, my colleague Matt Turck penned a perspective titled “When is AI, not AI.” In it, he discussed his viewpoints on the failure of ScaleFactor, a company that sought to use AI to replace accountants, and had raised more than $100M in venture capital funding from firms like Bessemer, Canaan, and Coatue.
Matt notes that a core issue on why ScaleFactor failed is that they likely never were able to graduate from having a human in the loop to deliver the core service. Matt’s prescient, and still applicable observation was, “With a bunch of humans involved in the early days, you have a terribly negative gross margin business. The hope is that over time you graduate to something that looks much more like a SaaS business in terms of overall economics (with the added benefit of some defensibility as your core AI is hard to replicate). But if you don’t dramatically improve automation and get rid of the humans in the loop over time, you don’t have a business.”
Could ScaleFactor be to the AI age what Webvan was to the internet? Potentially. But it’s important to note that from 2020 to where we are today, the technology landscape has dramatically changed. We now have a healthy set of open and closed source models (i.e GPT-4, Claude, Llama, Mixtral, Gemini, etc) and model libraries (i.e Hugging Face) that are available and consistently improving, but most importantly they are in active competition with one another. This competition is paramount. This competition breeds the incremental innovations that lead to a 1 million token context window, or the breathtaking uniqueness of a product like Sora.
Similar to personal computing, mobile phones, or the internet, over time we will look back on the current state of the art and ask, “Can you believe we thought that was impressive?”
This brings me to the second question, at what rate will we continue to improve? Unfortunately, to my knowledge there is no version of Moore’s Law charting the potential path forward. But there has been steady, compounding progression. Case in point:
2009: ImageNet is created
2012: AlexNet dominates & spurs interest in deep learning
2014: Sequence to sequence learning and attention mechanisms are developed
2015: TensorFlow is released
2016: AlphaGo victory occurs
2016:OpenAI release initial research on generative models
2017:Attention is All You Need Paper is released
2018: GPT-1 is released and leverages the transformer architecture
2019: GPT-2 is released
2022: GPT-3 is commercially released
2023: GPT-4 is commercially released
In less than two decades, we’ve gone from having no notion of what a generative pre-trained transformer is, to startups like Perplexity threatening the dominance of someone like Google. Since the release of the transformer architecture, a new GPT has been released every year with the exception of Covid and the odds are that GPT-5 will be released this summer. The emergence of these foundational large language models and the velocity through which they have improved gives me general comfort that the technology writ large will continue to improve.
This brings us to our third (and likely trillion dollar question), why? Why will AI as a fundamental technology improve? My own personal hypothesis is deeply unsatisfying, but it fits Occam’s Razor, and I think the answer is capitalism. The horses are out of the barn, and the industry is shifting into overdrive. Call it reductive, call it ignorant, call it whatever you want, but it makes sense. So long as the early experiments with commercializing AI continue to prove their ability to generate significant amounts of revenue, and over time generate significant amounts of profit, the market will continue to giveth, and giveth it will in the form of improvements to AI. We started with the Model T and now we have Tesla. We began with the ENIAC and now we have the iPhone. Technological progress is relentless, inevitable in my mind because it is directly linked to the incentive system of a market economy. Make something better, faster, and cheaper for people and you can take a little cut of it for yourself. Make something better, faster, and cheaper for people that is wanted or needed at scale? Do that and you’re in a whole different ball game.
As we transition toward a world where software is the service, I think it’s important to take stock of what we do and do not know today, in addition to the world we are likely headed toward if the industry continues to improve. First off, what we do not know.
As mentioned earlier, we do not know the rate of improvement the foundational LLMs will achieve. Nor do we truly have 100% understanding of how the models work. We do not know if the foundational LLMs over time will operate as an infrastructure focused oligopoly, or if they will extend up the value chain toward end distribution and workflows (i.e traditional SaaS). We do not know if the “scaling laws” will continue to be valuable, and we do not know if AGI - the stated purpose of many foundational LLMs - will ever actually be achieved.
That being said, what we do know is that OpenAI is the fastest growing company in history, reaching $1bn ARR in roughly a year. We do know that, within that same vein, technology outcomes have been successively larger in shorter time frames over the past 30 years. We do know that the historical SaaS playbook of digitizing data and manual workflows has created trillions of dollars of economic value driven by companies like Microsoft, Salesforce, Adobe, and more, but represent just one component of the jobs to be done in any organization. We do know that the engines of capitalism have every incentive to adopt AI once it is proven to be an adequate substitute to equivalent labor, and we do know that we’re sitting against the backdrop of a global macro environment that is yearning for growth.
Taken together we can view the unknowns as variables of industry velocity and value capture. The desired direction is clear, but the speed at which industry advances and where value accrues is less certain. For the knowns we have the historical examples that paint a picture of the opportunity at hand. Multi-trillion dollar outcomes are within reach. Once a company, or a set of companies demonstrates their ability to repeatedly automate the service professional at scale, the market will quickly demand that a new standard of innovation is adopted no different than previous periods of automation or technology adoption. First movers will take a lion’s share of the gains, laggards will react accordingly, and we will have a clearer view of what the world we are headed to will look like.
Since getting into venture capital I loved to ask myself the rhetorical question of, “If I were an investor in the early 2000s, would I have bet on the internet?” I love the question for two reasons: 1) it screens out intellectual honesty as millions of people did not make that bet and; 2) it forces you to be reflective on why you wouldn’t have. On face value, we are now faced with a similar question, “If you are an investor in the early 2020s, are you intending to bet on AI?” For the majority of the venture capital community, the response is a resounding yes, but I would like to take it one step further. I would like to pose the question of if you are willing to bet on the global realignment of services and the service economy? Are you willing to view the unit economics and growth trajectory of a services firm in a new light? Are you willing to think critically about what the edges of capability are and are not during this paradigm shift? Are you willing to accept that a new set of rules are being rewritten with respect to what a venture capital backed company looks like and how it will scale?
Are you willing to see that Service is the new SaaS?
I know I am, and I know I am willing to do the work alongside others that do as well. The next thirty years of venture capital is unlikely to look anything like the last. Industries are safe. Services are not. Place your bets and act accordingly.
A big thanks to my friends Ben Kany & Vardan Gattani for being early readers & thought partners in helping refine this article. Y’all the homies