Most people are saying that AI will be an extinction-level event for software companies. This is only sort-of true. Companies can survive this trend, but it will require a challenging evolution. Many will not make it, while some will emerge more powerful and more profitable than ever. Today we will propose our thesis on what effect AI will have. This is informed by countless company discussions, and two separate meetups we hosted with 25+ late-stage software companies we respect. Some of this is almost certainly wrong, but putting this into the public eye is meant to be a cringe test. We are looking to create the pressure and clear thinking that public writing forces.
We structure our thesis around three types of companies:
- Data and content businesses (worst)
- Workflow providers
- Vertical SaaS (best)
Let’s dive in.
Data and content businesses
The Narrative: All these companies will be sucked into the base models. Since you can’t always put a fence around your data, you end up being a commodity supplier to these model providers.
The Nuance, the Reality, and What to Do About It: Not all data and content companies are created equal. We think of these companies competing in this 2 x 2:
For publicly available data, pure data vendors (the bottom left quadrant) may be in trouble. These vendors historically provide value by aggregating, cleansing, and structuring publicly available data. Generative AI is likely a zero-cost, good-enough substitute over time.
Even if you can put your data behind a paywall, we continue to see use cases where Generative AI can create synthetic data for model training. There are startups doing synthetic data generation for autonomous vehicles, so it doesn’t feel all that far-fetched to imagine a world where this happens to most forms of data. Even in the case where synthetic data doesn’t cut it, we think it’ll be incredibly challenging to put up walls around your data—LLMs will find their way in.
“When it comes to maximizing language model performance, it could be the case that there is little value in proprietary data going forward. This is because we have learned that models trained on adjacent and synthetic data increasingly perform similarly to those trained on proprietary data. As this continues, what will become more important than the uniqueness of your dataset is the system built around your models that enables them to improve over time. The acceleration of AI quality will be met with the need for greater AI control. We see this as the frontier of applied AI." - Mike Murchison Co-Founder and CEO of Ada
What about data rights and licensing paradigms? Outside of specialized cases like music, that have monopoly or oligopoly control over the most important assets, we still feel that true data custody and ownership are 3-5 years away (maybe crypto finally has a use case?). To paraphrase John Meynard Keyes—in the long run, we are all dead. Data and content startups that rely on emerging data custody infrastructure may not live to see it come to pass.
AI will be the meteor for many of these dinosaurs. To survive/evolve/win, startups are going to have to either build out proprietary data sets that are not replicable by LLMs, or seek protection by generating or hiding it in an application or workflow they own. You need to realize the situation and get ahead of it.
Most LLMs won’t have access to real-time data, and even more likely won’t have access reliably enough for enterprise use cases. If real-time data is critical in your space, figure out how you can build on that. Know that all of the older data will, in time, be absorbed into the LLMs and you can’t count on that for ongoing defensible value. You’ll need to find a way to use your data and expertise to expand into adjacent workflows and services, either horizontal (okay) or vertical (better).
In particular, we think workflow software has a lot to consider on the right path of evolution.
The Narrative: Workflows will be decimated by AI agents auto-generating the tooling that took you years to build. The cost of code creation may go to zero, with every individual being able to create a custom app easily.
Our Take: As a workflow, you can’t sleep on AI, but if you are mission-critical, you have a lot of moats that work in your favor. In fact, AI may act as a compounding advantage.
Simply stated, workflow software gets shit done. It takes inputs from people or other systems, processes different steps, and triggers an action. While a consumer audience might find a mildly hallucinated answer cute, even a small error will break an enterprise. Coca-Cola guards the process to create their high-margin, Buffet-admired sugar water under lock and key. They will, in no way, trust AI to figure out the workflow for Diet Coke production. Benedict Evans recently argued a similar point (emphasis added):
“There’s a huge difference between an amazing demo of a transformative technology and something that a big complicated company holding other people’s business can use. You can rarely go to a law firm and sell them an API key to GCP’s translation or sentiment analysis: you need to wrap it in control, security, versioning, management, client privilege, and a whole bunch of other things that only legal software companies know about (there’s a graveyard of machine learning companies that learned this in the last decade). Companies generally can’t buy ‘technology’. Everlaw doesn’t sell translation, and People.ai doesn’t sell sentiment analysis - they sell tools and products, and often the AI is only one part of that. I don’t think a text prompt, a ‘go’ button, and a black-box, general-purpose text generation engine make up a product, and product takes time." - “AI and the Automation of Work”
Workflow providers will be fine because generative AI doesn’t pass the trust test. Of course, we also didn’t trust the horseless carriage for a while, so eventually, this may not be the case. The reality is that you, the cloud SaaS provider, are in a race with native AI challengers. You have the advantage of existing time tested and trusted workflows: just pay the Sam Altman tax to add some OpenAI to your existing software. Our guess is you will incorporate generative AI into your workflows before native AI players build trust.
And it isn't just about trust, it’s about complete solutions. No one wants to stitch together five different partial tools, so founders need to hustle and use this time to integrate AI into your workflows before the AI startups go from bite size to complete solutions.
However, there are some hype areas you need to be concerned about.
What you should be concerned about as an incumbent is the VC hype cycle. There are investors willing to drop $100M into a $3M ARR AI startup that just moved into your neighborhood. You, the experienced pro, may laugh, because you know their approach won’t scale. However, they may make a mess of everything as they burn through all that money. LTV/CAC ratios will go totally wonky, talent will get more expensive, and fiscal discipline may go out the window. You’ll need to be careful.
Additionally, you need to keep an eye on customer behaviors. Sensing the gold rush, customers are starting to worship venture capitalists and celebrate their AI chops on social media. Founders need to get out in front on narrative and product vision. Pretend you are that $3M ARR startup—but rather than pitch for VC funding, pitch for the ambitions and affections of your most forward-leading customers. Find places where you can drive real results for customers: results they can see and point to (see the most impactful areas in the “wedge” discussion below). The ideal outcome is that you can reference them in your marketing. Whether you are a startup or an incumbent, you are a startup again now.
More importantly, there are three major structural evolutions you need to think about.
We think workflow providers may need to seriously rethink their pricing strategy. AI puts per-seat pricing under pressure, and companies will need to evolve their GTM assumptions.
Picture this: AI cuts call center staffing needs by a whopping 75%. The SaaS company serving them suddenly has dramatically less revenue per account while simultaneously providing far more value. Better software makes for a worse business. What happens when this effect is spread across functions throughout an organization? What does the world look like with dramatically fewer staffing needs?
Here’s another way to think about it—is AI a boon or a bust for Adobe? The boon argument is that a content marketer can do significantly more work in Photoshop with AI. The bust case is that far fewer seats are needed.
Internally, we have debated whether the right analogy for AI is the iPhone or Robotic Process Automation (RPA). The iPhone made it mainstream to pay for content, allowing the phone (the workflow in this analogy) to create and monetize distribution. They ended up capturing a large amount of the value created. RPA went in the other direction. It was priced to replace humans and drove massive growth, but there wasn’t enough IP differentiation amongst vendors, and pricing collapsed.
AI will probably force companies to go with usage, transaction, or value- based pricing, or raise their prices 2-3x. For example, Five9s just introduced AI agent pricing at 2x the normal price. They are the first, but there will be many more price increases to come. You may need to expect overall deal sizes to come down before they go up!
On the disruption front, there is one area that founders need to keep a very close eye on.
The Leapfrog Opportunity: Function-Specific-Training of Open Source LLMs
Current conventional wisdom is that it’s too expensive to build or train your own LLM, and while building your own from scratch is cost-prohibitive and constrained by GPU, open-source LLMs are too far behind and talent is too scarce. Most investors will tell you to suck it up and pay your OpenAI tax.
We think that open-source is going to play a far more important role than people realize. We’re hearing that LlamaX is already reaching GPT3.5 power. With this in mind, as you initially experiment with API access LLMs like GPT4, run concurrent tests with open source alternatives. You’ll find that you can get most of the way there with relatively little effort.
You probably should play with open source LLMs. Otherwise, a forward leaning competitor may leave you behind.
Again, we keep hearing a narrative of extinction. Supposedly, UIs will be totally disrupted. Human computing interface as a practice may either be reinvented or totally pushed aside by users. When you can tell a computer what to do in NLP, every user interaction needs to be rethought.
We don’t think UI design will go away—it will just become more strategic. You may remember the VC hype on voice: we were going to talk to Alexa to buy our groceries, plan our trip to Hawaii, and do our math homework. It turned out that voice sucked for that sort of thing, and most people just use Alexa to set a timer in their kitchen. Chat and generative AI are so much better than voice, but are still oddly non-intuitive to use.
UI design with guided or embedded prompts or context are needed to make the most of the new chat based interfaces. This will be an additive form of human computer interaction. There will be multiple interfaces, with chat taking a bigger role than it did 5 years ago but by no means displacing all others. Different tasks will lend themselves to chat versus UI design versus others (video, motion, etc). Different users will have different preferences (think keyboard shortcuts vs. mouse). Our advice is to think about your target customers, the work they accomplish within your app, and the best, most natural way to achieve that, then prioritize that interface.
Another untackled element of hallucinations is trust. What is known versus estimated versus cited is something current systems struggle with. Even once you fix the technology, making this intuitive to a user, front-line employee, or high-stakes position like an airplane pilot really matters. This is a huge opportunity for user design. Think of the early days of Google—ten years ago, it was very obvious what was organic and what was an ad.
The companies that build highly differentiated, complex, and structured workflows will look even better in comparison to their AI peers. The technology is not anywhere near the point where it can replicate mission-critical systems.
Workflow companies doing these things will be better positioned, and probably able to beat native AI challengers. However, there is one type of workflow company that we think is best positioned in an AI world. (To our longtime readers: you will not be surprised).
The Narrative: These are thin, flimsy products. Copilot represents a new type of workflow. Therefore, LLMs will speed up development and reduce the cost of creating a new generation of vertical SaaS products. After all, what use is a vertical-specific workflow, anyways, when you can auto-magically use AI to create or customize a workflow on top of a vertical data set? We’ve also heard that LLMs are growing so fast that the miniscule vertical data sets within a vertical SaaS business will be slurped by base models, anyway.
Our Take: Vertical SaaS will kick AI’s ass. If you’re the Vertical SaaS killer that we describe in the VSKP, you occupy the control point. You have account and workflow gravity, so even if an LLM finds a way to steal your data, it won’t matter. Your customers don’t want multiple solutions; they want one, maybe two at most. You run their business. No restaurant is going to risk their point of sale going down or hallucinating a customer’s order because the LLM got things wrong.
These dynamics are so strong that desktop license software from the 90s still occupies the control point for dentist offices around the world. We’re not saying that the power dynamics don’t change over time, but this fight shows the power of a control point over the latest and greatest technology.
However, even vertical SaaS companies need to be aware. AI is change, and change is opportunity—both for you and for your competitors. You need to watch out for the same thing as normal workflow apps (both with VC and customer hype) but there are also vertical-specific challenges to address:
Watch out for the Wedge
AI entrants will look for opportunities to “integrate and surround” by providing a compelling wedge offering. Watch out for:
- Revenue generation: Companies will happily pay for additional sales.
- Massive labor sub: If you get rid of the accountants, you can offer the accounting software for free.
- High-stakes decisions: In areas like estimates, trading, and credit underwriting, no one is going to worry about changing trading software if the AI is better at making billions-of-dollar trades.
You need to be aware of these add-ons trying to integrate and surround you. Snuff these companies out before they start (call us if you want to talk through the game theory).
Start Moving your UI
You may need to put some time into your UI. This might be the most daunting challenge: merchants already know your product, so adding AI is going to create costs beyond just building out your AI. However, the more you use your UI to build trust with your AI features, the less of a gap that new startup has to elbow into your market! A little defensive sales slowdown today saves you the headache of a competitor tomorrow.
Vertical-Specific LLMs Leapfrog
If the open source LLM approach turns out to be performant and financially viable, we think VSaaS players are best positioned to benefit:
- You are built to serve, built to suit. By focusing on the needs of one customer set, you’ll be able to more effectively and efficiently deliver results for the people you care about.
- You have domain-specific language, rules, and data taxonomy that are hardened by real-world experience. Plus, years of transactional data! While the moat here may not last forever, it is knowledge and data that is not public or easily accessible by LLMs.
- Most importantly, you own the customer relationships and know what matters to them.
A vertical-specific LLM could be game changing. As an incumbent, you have the best and most comprehensive data—just don’t let a native AI startup beat you to the punch!
Native AI businesses
Despite all the incumbency advantages enjoyed by category leading VSaaS players, we do believe there are many AI native startups that are going to thrive.
There’s a massive legacy-installed base of software to integrate and surround. Native AI companies will thrive by targeting the industries where the owners, boards, and management teams are more accustomed to EBITDA than AI. In other words, AI companies win when they target private equity-dominated industries. The legacy base in many verticals represents a large market share and big opportunities, and there is even a chance to go after cloud-native leaders if they are asleep at the wheel.
Every vertical software market is in a battle between legacy software and cloud providers. Most of the time, these cloud providers will be far more fully featured then an AI startup can hope to replicate in their first few years. AI natives will be better at selling against legacy systems than fighting for the same deals that their cloud competitors are bidding for. By using AI to solve problems that cloud couldn’t yet solve, they have a unique edge. Performed correctly, they can integrate and surround both cloud and legacy players, eventually winning the whole market.
To be clear, AI is one of the largest technology disruptions in history, and many of these legacy competitors are margin constrained, underinvested in automation, and can’t find labor. The AI + SaaS value proposition is so compelling that new industries are more willing to bear switching costs. We’ve even seen in our research that the customers “crossing the chasm” and becoming early adopters are industries you would never expect. For example, this is the first time in tech history that sectors like law and health are early adopters of a new platform.
Will AI be an extinction event for all software? We think not. Certainly content and data businesses have and will come under extreme pressure, but workflows have a number of strong moats that allow incumbent companies time to adapt and adopt AI as a complement. Vertical SaaS control point incumbents are even more protected in the short term and, if they are aggressive in pursuing AI features and vigilant on developments in open source LLMs, stand to gain much more from AI. AI will accelerate software eating the world. The disruption doesn’t come to the value of software, but who ends up capturing that value.
Again, this essay is meant for public testing. If you have feedback or want to discuss integrating AI into your SaaS app, you can reach us at firstname.lastname@example.org.