Selling Tools or Selling Results? AI Companies Are Heading Toward Two Completely Different Futures

Author: Variant

Edited by: TechFlow

TechFlow Editor's Note: AI companies are splitting into two species: one sells tools to law firms, the other acts as a law firm directly. This framework article by Variant breaks down why some industries should pursue augmentation and others automation—the key lies in liability, verification costs, and relationship dependency. For entrepreneurs and investors, this is a practical standard for judging the ceiling of AI companies.

The first wave of AI companies mainly focused on workflow augmentation. Many of them have become the most successful businesses in the history of enterprise SaaS. By 2026, model capabilities reached a tipping point, and startups began changing their approach. They no longer sell tools to existing players, but directly disrupt them through automation. Many of these companies look less like SaaS and more like the service itself.

The opposition between augmentation and automation aligns with Chris Dixon's classic framework: strong technology vs. weak technology. Weak technology is skeuomorphic, using new technology to replicate old practices, usually just more efficient. Strong technology "starts from first principles, fully utilizes existing resources, and designs what technology should look like".

Take legal services as an example. Harvey sells monthly SaaS subscriptions to law firms, while competitor Crosby is a law firm itself. AI does the legal work, and human lawyers intervene only where humans are most needed: building and maintaining client relationships, verifying work, and taking responsibility when things go wrong.

Companies selling augmentation tools do not want to disrupt their customers and organizational structures; those are their bread and butter. Companies selling automation are implying the world should change drastically. Like service companies, AI companies doing automation prefer selling results rather than seats, tokens, or subscriptions.

Chris mentions in the article that both strong and weak technologies can succeed; often the timelines are different. Both paths have already and will continue to produce major results. The question is which path better matches the specific work, market, and evolution state of technical capabilities.

At Variant, we use a simple framework to judge which workflows are suitable for automation and which for augmentation. Following this further, we can see what kind of vertical companies might emerge and where VC-scale opportunities will appear. We look at three dimensions:

1. Liability: How dangerous and expensive is it when AI makes a mistake?
2. Verification: How fast and cheap is it for users to verify work quality?
3. Relationships: How dependent is the experience on human interaction?

Work with high liability, high verification costs, and high relationship dependency is suitable for augmentation. This framework is clear in the comparison between Crosby and Harvey. Internal legal counsel at large companies fits all three items of the framework: high liability, high verification costs (many open-ended tasks), and high relationship dependency. Augmentation tools like Harvey or Claude Cowork make sense in this scenario. Conversely, Crosby focuses on more routine, high-frequency legal services like commercial contracts, where liability is usually lower, easier to verify, and relationship dependency is weaker, serving startups.

We learned similar rules when building autonomous systems on public chains: smart contracts can only act on verifiable things. Anything that cannot be verified must be handed back to humans for off-chain augmentation processing, usually through governance. Crypto tokens follow similar rules: extremely good at rewarding verifiable quantities (compute power, staking, liquidity), extremely bad at rewarding subjective quality. This is why crypto successfully launched financial markets but hasn't cracked Airbnb or Uber yet. The lessons here are increasingly applicable to judging where true autonomous systems can scale.

Today, true automation is more limited in application scope and accessible market size. Startups focusing on automation workflows are often best suited for narrow beachhead markets, selling to startups or niche markets less resistant to changing the game.

Starting from a narrow wedge market is a way to avoid the "kill zone" of large AI labs, which focus on horizontal expansion to augment larger existing markets.

Over time, we believe automation solutions will slowly encroach on markets that are more relationship-dependent and have higher liability, because model capabilities are expanding, verification costs are declining, human-AI interaction is crossing the uncanny valley, and cultural acceptance is rising. In that case, the first-mover data advantage gained from the beachhead market may compound, allowing startups to grow along with the market itself.

The landscape and long-term attractiveness of automation vs. augmentation will continue to diverge across different industries. Take education as an example. Alpha School is a vertically integrated AI private school, replacing teachers with "guides" to facilitate a basically automated learning process. Although we believe AI-assisted education will be extremely important, we believe excessive automation and verticalization will not expand beyond niches, because parents worry about teacher-student relationships, and schools worry about liability. Most of our education systems are rigid (and mostly public) institutions, which is unlikely to change soon. Even top private institutions face the innovator's dilemma. Therefore, solutions focusing on augmentation seem more likely to achieve major results, but will also face competition from existing players entering the field.

Recruiting points in the opposite direction: we are more bullish on automation. The Harvey vs. Crosby in this vertical is: Juicebox sells candidate sourcing software to recruiters, while Prism is a recruiter itself—give it a job brief, it delivers candidates ready for interview, charging only when someone signs. Its key stages are easy to verify (response rates, interview performance, signing rates), and liability is relatively low. Recruiting is highly relationship-dependent, but recruitment is more upstream (mainly finding candidates and converting to interviews). This is also a function companies often outsource, so the first customers are startups: they prefer buying results rather than building teams. The largest enterprises have internal recruiters and will choose augmentation tools, because they are less willing to automate themselves away. The target customers are startups, this is a thin wedge, allowing automation solutions to embed into data-rich processes, improve over time, and grow with the market.

There are two types of AI companies: augmenters and automators. Both can expand human autonomy, but in different ways. Augmenters give people more leverage within existing institutions: more knowledge, more output, more initiative. Automators expand access: expert services that were once only affordable for enterprises become accessible to everyone, and time consumed is released for more ambitious things. The latest opportunity is not just selling software, but becoming the service itself.