AI and Vertical Market Software

Mar 22, 2025

The bear case for software goes something like: the cost and technical ability to develop software is coming down because of generative AI → therefore new entrants can take more shots on goal at incumbents → therefore the chance someone, somewhere will create a better mousetrap that induces customers to switch increases. Barriers to code down → more shots on incumbent → so the incumbent's duration is less certain than before → multiple assigned has to rerate lower because of terminal value concerns. This idea was all the rage in the summer of 2024 with “contrarian” VCs titling google doc essays, “Software is Dead” and making draconian predictions such as,

“Vogue wasn’t replaced by another fashion media company, it was replaced by 10,000 influencers. Salesforce will not be replaced by another monolithic CRM. It will be replaced by a constellation of things that dynamically serve the same intent and pain points. Software companies will be replaced the same way media companies were, giving rise to a new set of platforms that control distribution. SaaS, ARR, magic numbers–these are all shorthand to understand the old model of business building in software, one where the expense associated with creating software was a moat.”

I’ll take the low hanging fruit response that the market quickly caught on to here and then move into something wildly more interesting. The moat in VMS was never “the expense associated with creating software,” if it were, these businesses would trade at multiples far, far below where they do because of terminal value concerns (obligatory play around with a DCF in Excel comment). What multiples? Well…I’ll proffer media multiples…because they have very low switching costs…which is why we saw the outcome we did. The software moat generally and for VMS more specifically is their high switching costs and distribution, not the tech itself or the ability to create. Go to your dentist’s office and look at the UI on that software on the monitor or take a flight on Southwest and ask the lovely crew member behind the desk to see the booking software, better yet go and ask the guy building your house to show you the project management software he is using (if there is one at all) – chances are it’ll be pretty old, and the fonts, colors and general UI layout will be ugly and clunky. There is a reason ERP implementations have been a really interesting short signal, they are really hard and cause deep disruption.

In short, the conversation around AI replacing spend is the same conversation that many enterprises have today on the Build vs Buy (use internal engineer manpower to custom develop apps vs pay for a 3rd party vendor), and I don’t believe that AI has meaningfully changed the Build (which has been losing steam) vs Buy debate. There is inherently a lot of sunk costs in existing software (employees have sunk the cost of learning to use this system and investments have been made in compatible software to customize the system to the clients needs. Changing that comes at an extraordinarily high cost: the time and effort to research competitive offerings, the purchase cost of replacement ERP, new complementary software, transferring data, retraining employees, and risking interruptions of services and data from migrations from one service to another.)

So why bring this up now, a year after the post I just mentioned? Well, it’s happening again: “As a customer, you'll never get 3rd party software perfectly customized to your use case. Unlike Google or Meta, not everyone can afford to build their own totally unique tech stack separate from the rest of the world 6/ But in a world where LLM coding capabilities dramatically improve, then all of these tradeoffs are no longer true! The cost of custom software development becomes low - not high. In that world, why *couldn't* you have 1000 different companies writing 1,000 custom CRMs?”

But we have seen this movie before. Think about applets or the low code and no code movement: every shift has people thinking that every company will build thousands of custom apps, and every time they learn that building and maintaining thousands of custom apps is not something you want to be in the business of doing. No code/low code lowered development cost and enabled non developers to build apps but it isn't like the average Joe account starting building his own custom Bill Pay app. Quality also isn’t evenly distributed: there is a marked difference between professionally produced content on Youtube (much less movies/TV shows) and home video. The cost to develop software and launch a startup has been falling for decades and yet the economics are still healthy.

Vertical market software is such a phenomenal business because of the switching costs they can create when they are at the control point and have data gravity. These companies know their customers so well they effectively become industry evangelists. Yes, software is effectively a UI wrapped around a database. It is easy to point an LLM at a large source of information. The hard part is making it useful, making it do something that solves a pain point so well it is worth leaving an existing solution. The software world is moving away from point solutions to platforms. The ability to consume information and spit back some stuff is a point solution, the platform you build around it is reporting, accuracy, collaboration, compliance, regulations, integrations and everything else that actually makes it into something that a business would adopt. Even after that you have to execute and develop a go to market motion against an entrenched incumbent who is likely heeding the Gingerbread man’s advice "Run, run, as fast as you can! You can't catch me. This is why truly mission critical solutions are so hard to rip out even if they can potentially be developed in house.

Verticalized summed up this viewpoint pretty well in their piece Vertical SaaS is Dead, Long Live Vertical SaaS, “Third, the marginal cost of software may approach zero, and yet it doesn’t matter in vertical software. Many industries have no effective R&D budget to take advantage of it. They will still pay a vendor to manage their technology stack. Think I’m wrong? Explain to me continued prevalence of the MSP industry then! There will always be incredible pure play software opportunities. There may just be far fewer venture-backable ones.”

The second bear case goes something like what Jamin Ball laid out here: The real value of switching costs comes from the workflows and integrations that are built around systems of record (WDAY, CRM etc). These applications are not just repositories; they are tools designed for humans to input and manage data. At their core, they’re interfaces built to help employees take information from one source and enter it into another, often manually. These systems rely on people sitting in front of screens, navigating the user interface (UI) to input, process, and move data through various workflows. This human-centered design has made these systems essential for companies, and replacing not only the data but the workflows tied to them is where the true switching cost lies. With AI, however, this dynamic could radically change. One of AI’s greatest strengths, particularly with large foundation models, is its ability to process unstructured data—data that humans today manually enter into systems of record. So what if AI does this itself and negates the role of the human UI, replaced by an AI agent seamlessly interacting with the system of record.

My pushback here is best summed up by Buck: “even with improved models and tooling, LLMs are unlikely to become fully reliable. Hallucinations will still occur. Humans will need to be in the loop, deterministic scaffolding (data, workflow/process, alerting) will still be very necessary, and specialization will be critical to the performance and efficiency of these new AI systems. And, most of the tasks being addressed by AI are what humans do today, which is complementary to existing systems. The river of AI is likely to flow around the massive existing rock formation that is deterministic software. Gradually eroding it away versus sweeping it away. Software bears are leaping to a fully eroded end state (fully autonomous agents). But that end state is likely well over a decade away.”

I think in the intervening period, most existing software vendors (especially at the application layer) have a right to win and to gradually rebuild themselves from UI wrapped around a database to implementing AI in addition to the workflow that is wrapped around the database. My belief is that A) this will be iterative and B) incumbents have the advantage. This is backed up by a survey Avenir conducted last year:

From Systems of Record to Systems of Action

So if you are a system of record that has come to occupy an existing control point in your vertical market, how should you react to the advent of AI? The race now is to become a system of action as opposed to a mere system of record. These systems of action will enable humans, AI-assisted humans, and fully autonomous AI agents are act on data and trigger/carry out downstream workflows. In doing so, they can subsume the control point and render it irrelevant. Historically, software helped a business run. It automated front-and back-office functions that weren’t the core activity of the firm itself but generative AI is opening a new frontier. It enables the software to do the work itself. At a dentist office VMS historically helped keep patient records, billing, scheduling etc. What they generally don’t do is help dentists actually clean teeth. AI’s new capabilities are being used to blur the line between running the business and doing the work itself. Think of a transcription tool that start by automating the administrative task of documenting and summarizing patient notes. But, thereafter, it’s not hard to imagine triggering administrative tasks like follow-up scheduling, invoicing, and even clinical recommendations. A native AI can offer them tools that makes their lives easier. This is typically digitally-native work that involves lots of manual drudgery such as data entry, reconciliation, summarization, tagging that AI can eliminate. If the experience is “magical” enough, the native AI can quickly acquire these users who live in that software.

If you occupy a control point and embrace AI, you have an opportunity to do much, much more for customers. Instead of VMS merely being the arms dealer an industry, AI enables your software to do the fighting.

On the flip side, doing the work also creates an opening for native AI challengers to gain a foothold and “integrate & surround” an existing system of record. If control point incumbents are slow to adapt, they open the door for a native AI challenger to win large pools of the control point and risk being relegated to the dumb database in the back closet. If these native AI products begin to command workflow gravity, they can force integration into the control point’s data and begin to build even better products because if they become the place where the work gets done they have the right to hoover in all the data (usurping data gravity), and to drive downstream workflows and actions (stealing workflow gravity).

Native AI players may be more effective at pursuing full substitution strategies—eliminating the need for a human by automating tasks completely and, in doing so, bypassing the control point software. AI allows software companies to expand scope from running the business to doing the work itself.

If control points aren’t asleep at the wheel, they can leverage the $100Bs of venture capital poured into the infrastructure players and provide AI products that rival native AI players.

Native AI entrants should target (and existing incumbents should protect!) two types of work:

Make the beer taste better: Make the customer/practitioner better at their job. Make them better at making the good or delivering the service that is the reason why they exist. For an accountant or lawyers, it’s advising clients. For a medical professional, it’s delivering care. ‍
Administrative work: This is the repetitive, annoying work you have to do. Scheduling, summarizing, invoicing. You want to do it as quickly as possible. As a result, it’s probably not done that well or that efficiently. The goal is to take tasks off their plate such as answering the phones, scheduling appointments, reminding customers about appointments, paying bills, payroll, and taxes and reconciliation and reporting

The tech to gain the wedge doesn’t have to be groundbreaking. Wrappers work. The more the product is used, the more it has to learn and improve to stay ahead of incumbents and other entrants. The barriers to try a solution like this has to be very low. Make yourself easily detectable and cheap. Do not strive for a perfect product from launch. Focus on the needs of your customers and get them relief from the pain as quickly as possible. Once you have acquired a meaningful community of these customers and they are finding daily delight in your wedge, you can “integrate and surround” the incumbent’s control point. Your goal is to integrate into the system of record.

Fending This Off as an Incumbent

Only the paranoid survive and as an incumbent you have to protect your flanks to fight off the “integrate and surround” strategy. Look for systems where merchants bring in revenue and engage with customers. Once you’ve identified these systems, lock them down. Ensure all your customers are using your version. Consider making them free through bundling or freemium models. No API access. Block all the little hacks and tricks, even if they seem initially harmless.

Think through the “job to be done” by your customers. Similar to a franchise system, you as the VMS are only healthy if the underlying vertical is healthy. You are looking for the the job to be done you can empower, and the administrative work you can automate away. You need to expand your scope beyond running the business, to understanding how your software can help its merchants do the work.

Similar to the DevOps movement, realize that you are now building products for users, not for owners. Make your product easy to try. Now is not the time win design contests. Move quickly. Good enough is good enough for now. As long as you are adding value, you can improve the product later, a wrapper is fine.

Also similar to DevOps, get used to being a product led growth company. This may require you to develop new GTM muscles. Yes, there will be friction. It’ll be tough to get the sales team to pay attention. Sales reps don't know how to sell it, and low price points mean low commissions. But the priority now is to get engagement, not money.

Ultimately, success hinges on who secures the system of action first. Burrow into the daily operations and critical workflows. Whether incumbent or native AI, controlling user engagement will determine who dominates the AI era.

Barbell VMS

So that is an argument for the incumbent large platforms of the world like PCOR, CCCS, TTAN, APPF and others. They get better because of this. On the other end of the spectrum, there are now thousands (millions?) of little niches where it used to be uneconomical for a small team to work on that is no longer the case. Which means tiny teams can work on a dozen different ultra niche problems for limited # of customers.

As Fractal put it in their 2023 State of Vertical SaaS report, “The accelerating availability of low-cost development tools and infrastructure will facilitate faster creation of specialized point solutions targeting the most price-sensitive segments in vertical markets. This escalation in competitive fragmentation for downmarket customers could incentivize leading vertical SaaS players to shift focus upmarket instead. Leveraging their greater resources, vertical expertise, and capability breadth, established vertical SaaS companies are well-positioned to deliver integrated, end-to-end solutions that are better suited for midmarket and enterprise customers.”

In the Spring 2019 issue of Graham and Doddsville, John Hempton was quoted as saying:

“In my case, [I usually want to own] a company that makes a widget that is a small yet important part of a bigger thing, has high switching costs, and has incrementally improved over time. I call this the trifecta. There´s an old saying for this: ´There are riches in niches.´”

The VMS I am really interested in are the truly hyper niche. I am interested in the less than $500m TAMs. Businesses like RinkNet (Hockey Scouting), Cayenta (Utilities Billing), Orion (Winery Focused ERP) or Tarantula (Telecom Site Management). The R&D and S&M spend that any small new entrant would ever commit to building a competing product is low for these tiny end markets, and incumbents that have been building the product for 10 years have the cumulative benefit of all that development and marketing spend behind them. In this way, VMS is similar to building a theme park. Once a theme park is built in a region the incumbent has distinct advantages because they could have the best location, get regulatory compliance/approval and develop partnerships all with high barriers to entry relative to the size of the market. The incremental returns on capital would be much different than the average returns on capital. Similarly, in VMS there is a distinct first mover advantage. VMS businesses have customized products that improve over time through an iterative process that relies on customer feedback. As the 10thman put it in their CSU write up, “In my view, customized VMS products improve much the same way a cook improves a brownie recipe: the vendor sells Version 1.0, customers try it, they provide feedback, and the vendor uses that feedback to create Version 2.0. By Version 13.2.A, the vendor has it nailed. They are so ingrained with their customers that they can almost anticipate customer needs before the customer can (which leads to even better products). For a new entrant to be competitive, they need to work through all of that trial and error internally before launching Version 1.0 of their competing product.”

Enterprising vendors with deep customer relationships and continuously improving products have also built a level of trust with their customers that is hard to replicate. So even if a new entrant has a similar quality product on Day 1, why would a customer switch if they were happy with - and trust - the current vendor? These are products that might represent 1% of customer revenue, so any potential savings are minimal. It is difficult to justify the risk of switching for mission-critical processes. Make it to the control point first and the incremental returns for a new entrant are not nearly as enticing as the existing returns for the incumbent.

Category Impact

It is also useful to think about the rate of change in a given category. At risk of being banal: not every category will adopt AI at the same pace. Categories such as ERP, HCM, financial, legal, payroll are going to be slow adopters. As Buck pointed out: “The second bucket [the high velocity bucket] is where there is likely to be a lot of change. Creative, sales, marketing, customer support, workflow automation, RPA and software development are examples of these markets.”

Why Seat Based Pricing Isn’t Dead

Tying software pricing to outcomes or usage inherently makes the cost of the product for end users more volatile and less predictable, additionally the moats for the dominant players today are tied to the number of users (user learning curves and switching costs). Broadly speaking, software companies with consumption models have been more cyclical with quarterly earnings seen as indicators of the health of IT budgets, given buyers’ ability to rationalise spend faster than multi-year seat-based contracts. Tying software spending to seats still makes more sense in most cases at the application layer. I think this will evolve to be a virtual seat based model + AI agent calls pricing model similar to what Agentforce/Zendesk/Jira/GitHub have done. In fact, 61% of SaaS companies implemented some form of usage-based pricing in 2022. The overall business will be a hybrid pricing structure. Companies will still charge per seat for human use, but charge differently for their AI product. As their customers shift to consume the software more through AI agents vs human agents, the mix shift of the overall business will almost tilt to consumption, but this will take time to play out in full.

As Software Synthesis lays out: Hybrid models are likely to prevail over pure consumption or seat-based models in the medium term. In that scenario, investors will need to assess top-line revenues that are a combination of recurring revenue (and all the beneficial properties that entails) and consumption revenue (with the volatility that brings). Enterprise software metrics predicated on ARR will need to be revised. In other words, analysts will have to scratch below the surface of a revenue multiple and parse out the relative quality of that revenue.

One Other Takeaway

One other tidbit i think is worth mentioning comes from a Steve Jobs quote

"In hardware you can’t build a computer that’s twice as good as anyone else’s anymore. Too many people know how to do it. You’re lucky if you can do one that’s one and a third times better, or one and a half times better… Then it’s only six months before everybody else catches up. But you can do it in software."

“The difference between the best worker on computer hardware and the average may be 2 to 1, if you’re lucky. With automobiles, maybe 2 to 1. But in software, it’s at least 25 to 1. The difference between the average programmer and a great one is at least that. The secret of my success is that we have gone to exceptional lengths to hire the best people in the world. And when you’re in a field where the dynamic range is 25 to 1, boy, does it pay off.”

Amjad Masad said on ILTB: “I think the bottleneck to human flourishness has been our output and how much we can get done. And the most productive people on earth should be able to do more, should be able to have more capital. We should double, triple down on the most productive people in the world. And I think AI really accentuates that.” I think if we get AI agents helping people generate software, build businesses, make decisions and automate tasks, the most productive people will get even more productive. I think a young Bill Gates or Bezos type person will be 1,000x more productive.

I think in the AI age you have to be willing to overpay for talent because it will be almost impossible to overpay for talent.

In summary

The argument against Constellation/VMS is this:

Constellation is able to reap value from vertical SaaS that is hard to displace and that customers never leave.
AI seems to imply that long-tail vertical markets will be easier to build solutions for.
These solutions by and large have low continued R&D expenditure. AI-native developers could theoretically create Constellation-type solutions for far cheaper.
This theoretically puts price pressure on Constellation companies.
Thus, if AI is actually going to put long term pressure on SaaS, we would expect it to show up in Constellation’s earnings and thus they represent a leading signal if and when this occurs.

If a customer is already willing to use a highly legacy vertical software solution that’s well past its expiration date, shouldn’t we expect them to continue using it? In other words, Constellation seems to specialize in industry applications that are important, but amongst a business subset that seems to not really care too much about being on the bleeding edge. Turns out, that’s a lot of businesses.

Links