Edition 3

If you are building AI tools for companies, the biggest mistake is automating the current process.

"Don't take a 12-step workflow and add AI to each step. You end up with 12 LLM calls per transaction, unpredictable outputs at every stage, hallucinated data in your system of record, and higher cost, lower trust."

Instead, the entire process needs to be rethought from the end goal backwards.

The frame used in this article is that every business tool has four layers: 1) the system of record, where one right answer is needed, 2) logic, that can be expressed as if/then, 3) deterministic workflows, that can follow a predictable path of scripts, triggers, and automations, and 4) AI / LLM, where genuine judgment is required.

These layers govern the steps actually needed. Using a purchase order for an example, the wrong way would be to take the existing 12-step workflow and AI to all of them, ending up 12 steps, 12 LLM calls, hallucinated data, and higher cost.

The right way would be to start from the end goal, work backwards, and realize most of the 12 steps are workarounds, not requirements, that exist because the original software was bad. You can kill them and end with three steps: a database validates, a script routes, and AI handles the one ambiguous classification.

To summarize, the wrong question is 'How do we add AI to this process?' The right question is 'What's the end goal and the minimum path to get there?'"

Similar to the last story from a different angle, this article makes the case that the prevailing "agent" and "co-worker" framing is harmful. This sells AI as labor replacement, which anchors buyers to headcount cuts and leaves the underlying system untouched.

It also makes products illegible: when everything is an "agent," the least tech-savvy, most buzzword-allergic buyers in vertical markets can't tell what a product actually does.

The cautionary tale is GM's 1980s "factory of the future," which put $40B-plus of robots exactly where human workers had stood, and kept the assembly line and job roles unchanged. It ended up being a massive failure, actually driving costs up, and producing robots that sometimes painted each other instead of cars or welded their doors shut.

Toyota, on the other hand, used the same technology but asked what becomes possible when a new capability enters the system, and redesigned the plant around it. Even further back, the same lesson was learned in factory electrification, where owners bolted an electric motor onto the old steam drive shaft and saw nothing for thirty years until they rebuilt the floor around distributed power.

The article ends by saying the real prize is capability, not substitution, i.e., "the work not done: bids not submitted, calls not answered after hours, patients not seen, inspections not performed." That shows up as revenue growth, not a headcount-reduction model.

"When you frame AI as an 'agent'—a digital worker—you implicitly promise substitution... When you frame it as a capability, you promise expansion."

AI cuts cost and helps the Rule of 40, but most of it - today - does nothing for the customer relationship that actually holds retention. Too often, it is internal cost reduction that never touches the customer.

"Just because you can cook rice infinitely at no cost doesn't make you a Michelin-starred restaurant. It's all the other aspects of these integrated activities that make you who you are, not the AI tool."

During the software selloff, the median public SaaS name fell 32%, EV/Revenue went from 9.1x to 4.8x, and 86% of companies compressed. Some of the common takeaways may not be as straightforward as publicized.

Tom Tunguz claimed vertical software trades at the steepest discount in software because they are slowing down. The median LTM growth rates of vertical and horizontal software, however, are effectively the same, and returns have been weakly correlated (0.07) with revenue growth.

Tom also summed up the popular view well by saying "Vertical software has fallen 43% this year. DevTools, just 21%. The gap between them… tells you what markets actually believe about AI." While facially accurate, it implies that the market concluded that vertical software itself — by nature of being industry-specific — was easier to replicate with LLMs than other software.

The real performance gap had more to do with a handful of horizontal SaaS companies that were "graded" by the market as being not just defensible but poised to benefit from the rise of AI, i.e. the "AI picks-and-shovels" treatment.

Vertical SaaS still trades at a premium to horizontal peers, but traded much higher to begin with. The market, however, has correctly repriced terminal value, simplifying the moat that matters to "Are you agentic infrastructure?"

Another similar lesson can be applied to the myth that all vertical moats are eroding equally. This article classifies the 57 vertical SaaS players by source of defensibility: 1) proprietary data moats (e.g., Verisk, Veeva, FICO) still commanding a 72% premium (down from 220%), 2) regulatory barriers without compounding data (e.g., Tyler Technologies, ADP) falling from a 120% premum to a 15% premium, and 3) the vertical halo (e.g., ServiceTitan, Toast) falling from a 41% premium to a 40% discount.

The final through-line, the article argues, is that markets are pricing the destruction and the near-term acceleration but ignoring the longer-term value creation, because AI still needs domain-specific data and decision context that "exists nowhere on the public internet" and lives instead in the workflows, transactions, and institutional knowledge vertical platforms capture.

The AI-native services thesis is hot. Sell the work/outcome, not the tool/software. For every dollar spent on software, six are spent on services, and so on.

This investor, while agreeing there are many appealing features of the model, is more cautious regarding the scope (which markets this applies to), the addressable market (how large a business you can realistically build), and venture fundability (the proportion of AINS startups that are a good fit for venture capital).

They're seeing a broad embrace of startups tackling services without clear paths to automation or without a natural venture-scale second act, forgetting why software is such a hallowed category of investment, making all-too-liberal assumptions around the ease of automating humans out of an industry, the down-market valuation of services revenue, or the right to grow a product suite horizontally.

I recommend reading the full piece - an estimated 18-minute read - if you find the subject intriguing. The overall point they make, however, is not that AI Services built on LLM infrastructures are inherently a bad initial wedge (or business model). Rather, it's that these startups must go beyond the wedge to be internal platforms rather than external vendors. I.e., they must build internal customer workflows, and the dominant mode of winning in vertical AI will still be software with software defensibility, regardless of how it's positioned or priced as outcomes rather than software, or by whom it is used.

Entrata, a property-management software company, filed its S-1. Entrata is not a hot AI story, but has strong metrics for its scale: Rule of ~40 with ~20% margins and 20%-plus growth, 97% gross revenue retention, and 117% net revenue retention at $509M of ARR in 2025.

The last private mark was $4.3B (Blackstone, May 2025), but Entrata's closest comp, AppFolio, is down 23% in the period since.

Entrata is as close to a classic profitable-at-scale software profile as it gets, so it will be interesting to watch what valuation it fetches and whether it reopens the IPO window for the backlog of "pretty good" software companies waiting to file.