Earned Intuition #003: Vertical AI: What Would You Never Ask a Human to Do?
Just "selling the work" is not enough. The best vertical AI companies will invent new types of work.
I. “Sell the Work” Is a Starting Point, Not an Endpoint
The dominant strategy in Vertical AI today is straightforward - Take an existing task performed by humans, automate it with LLMs, and charge accordingly.
An oft-repeated Henry Ford quote comes to mind: “If I had asked people what they wanted, they would have said faster horses.” If we restrict ourselves to existing human workflows, the best we’ll achieve are faster humans.
In nearly all cases of emergent technology, the buyers don’t really know what they want until they see it. Therefore, a grave mistake we make in evaluating the use case of LLMs is to box ourselves in with the framework of - Where can I replace humans today? Instead, we should ask ourselves - What would I never have asked a human to do, but now becomes possible?
II. Why “Sell the Work” Resonates and the “Faster Horse” Fallacy
For many of these industries, software spend is insignificant compared to labor spend. And so the prevailing logic emerged: If technology companies want to reach the other 90% of GDP, you have to attack labor spend. Thus creating massive TAM expansion for many software categories. This approach has already produced credible companies. But it also creates a trap.
Most Vertical AI companies today are building faster horses - streamlined versions of workflows we already understand, already staff, and already budget for. If your product is pegged to labor replacement, labor cost reduction also becomes your price ceiling. If your product unlocks workflows that humans could never do, labor cost becomes your price floor. The additional value you deliver now becomes your ceiling. This is true unbounded TAM expansion!
III. The Alternative: Build What Humans Cannot
There is a scene from the movie Gladiator where Maximus and his fellow captured gladiators from foreign countries first arrive in Rome. One of them looks upon the Colosseum for the first time and says: “I did not know men could build such things.”
Our goal should be to create and transform industries that otherwise could not exist under the constraints of human scale. That are only achievable through the computational power we have unlocked. To make this concrete, here’s a non-exhaustive framework of what AI-native workflows may exhibit:
Impossibly high coordination complexity
Tasks with infinite permutations that overwhelm deterministic systems
Continuous learning loops where outcomes improve perpetually with more data and never asymptote in terms of improvement
Negative ROI for humans to perform manually, regardless of how low we bring costs
IV. Business Model Implications: Price What You Unlock
AI software spans the gamut of usage (token) based pricing to per agent/per month but usually benchmarks itself to the alternative labor costs saved.
However, if your product delivers a new outcome versus simply a cheaper one, you are free to price based on extra value delivered. This is how you meaningfully expand TAM instead of operating within the confines of a pre-existing one.
The pricing model for AI applications will bifurcate:
Workflow automation-based products will be priced against FTEs costs (floor)
Novel outcome-based products will be priced against business value delivered (ceiling)
Outcome-based pricing, where revenue is tied to measurable performance improvements, is already showing up in enterprise AI. Palantir, for example, often gets paid not just for deploying Foundry, but for driving operational value. Upside participation is embedded into the contract.
V. Some Examples I’ve Seen or Am Thinking About
Energy & Utilities
Sam Altman has said that he is targeting 250 GW of compute capacity by 2033. That is 1/3rd of peak power in the US dedicated to generating more videos of Sam! At the same time, electrification is reshaping load curves across transportation, housing, and industry. How do we fundamentally reinvent energy operations with a software-first approach for both energy consumers and generators? How do we get humanity back on the Henry Adams curve?
Legal
In law and other white collar services industries, the early use cases are obvious: automate research, summarize transcripts, draft documents. That’s not transformation. That’s speed.
The deeper opportunity is to change the unit economics and operational throughput of the firm itself.
EvenUp, began with demand letters for PI lawyers - a textbook “sell the work” wedge. But that was just the start of changing the “work” and fundamentally enabling a different set of operations and revenue potential for PI firms through our subsequent “system of work” products. Many of our clients have grown their top line by 50%+ without adding any headcount.
This is true TAM expansion driven by new business made viable by new capacity.
In transactional law, my bet would be on Legora to ultimately become a bigger company than Harvey for the simple fact that Legora seems to be enabling entirely new firm operations and contract workflows. Harvey on the other hand, looks to be replicating a genius associate with its focus on fast, deep legal research. Maybe the company should have been called Mike instead.
Financial Services
One of the core principals of running a financial services/fintech business is customer segmentation. How do you differentiate profitable versus unprofitable customers based on interchange spend, delinquency/default rate, account balance etc. Then how do you service them differently based on their projected LTV.
It’s why JP Morgan Private Bank and Chase Bank treat their customers differently. Retail checking/savings customers are often loss leaders and it’s not worth the bank’s effort to try to introduce them to other products. But what if that didn’t have to be the case anymore? Imagine agents that treat all the customers the same regardless of account balance or spend potential and could monetize the long tail of retail customers effectively. What if every single customer received an expert financial advisor watching over their accounts and proactively offering bespoke services?
Consumer Facing
While I’ve spent the majority of this post thinking about B2B applications, I believe that the surface area for these types of applications may be even larger when we start to explore consumer-facing applications.
One area I’ve been noodling on is what an AI-native healthcare system would look like. Some companies today such as Doctronic in PCP and Thea in Dermatology have started to explore the edges of this concept. They use phrases such as “like a real doctor”, “clinical grade”, or “world-class physicians”. Frankly, I’m more interested in a future AI doctor that makes how we practiced medicine in 2025 look barbaric. To start, care should be continuous, proactive, and deeply personalized - not episodic, reactive, and time-constrained.
For more thinking on this, my friend Anne Lee Skates @ Parable wrote a piece on Trusted Recommenders that you should definitely check out.
VI. Build Colosseums
The first generation of vertical AI will sell the work. The generation that defines it will build workflows that work was never scoped for.
Don’t build faster horses.
Build colosseums.
I’d love to hear from you. DMs are open!
Thank you to Anne Lee Skates , Nikunj Kothari , Coyne Lloyd for feedback on the initial draft!
Love this!