Author: Variant

Translation: Shenchao TechFlow

Introduction by Shenchao: AI companies are diverging into two species: one sells tools to law firms, while the other operates as a law firm itself. This framework article by Variant breaks down why some industries should focus on enhancement while others should pursue automation—the key lies in responsibility, verification costs, and relationship dependence. For entrepreneurs and investors, this is a practical standard for assessing the ceiling of AI companies.

The first wave of AI companies primarily focused on workflow enhancement. Many of these have become the most successful businesses in the history of enterprise SaaS. By 2026, when model capabilities reach a critical point, startups will begin to change their approach. They will no longer sell tools to existing players but will directly disrupt them through automation. Many of these companies may not resemble SaaS, but rather the software as a service itself.

The opposition between enhancement and automation aligns with Chris Dixon's classic framework: strong technology versus weak technology. Weak technology is objectified, replicating old practices with new technology, usually resulting in merely higher efficiency. Strong technology, on the other hand, "starts from first principles, fully utilizes existing resources, and designs technology as it should be."

Taking legal services as an example. Harvey sells monthly SaaS subscriptions to law firms, whereas competitor Crosby is itself a law firm. AI does legal work, with human lawyers only intervening where human involvement is most necessary: building and maintaining client relationships, validating work, and assuming responsibility when issues arise.

Companies selling enhancement tools do not wish to disrupt client and organizational structures, as those are their sources of livelihood. Companies selling automation, however, suggest that the world should undergo significant changes. Similar to service companies, AI companies focused on automation are more inclined to sell results rather than the number of seats, tokens, or subscriptions.

Chris mentions in the article that both strong and weak technologies can succeed, often with different timelines; both paths have produced and will continue to produce significant results. The question is which path better matches the evolution of specific work, market, and technological capabilities.

At Variant, we use a simple framework to assess which workflows are suitable for automation and which are suitable for enhancement. Following this framework, we can identify what kinds of vertical companies may emerge and where opportunities for venture-scale investments might arise. We look at three dimensions:

1. Responsibility: How dangerous and costly is it when AI makes mistakes?
2. Verification: How quickly and cheaply can users verify the quality of work?
3. Relationships: How much does the experience rely on interpersonal interactions?

Jobs that have high responsibility, high verification costs, and are highly dependent on relationships are suited for enhancement. This framework is evident in the comparison between Crosby and Harvey. The internal legal departments of large companies meet all three criteria of the framework: high responsibility, high verification costs (many open-ended tasks), and high dependence on relationships. Enhancement tools like Harvey or Claude Cowork make sense in this context. In contrast, Crosby focuses on more conventional, high-frequency legal services, such as business contracts, where responsibility is generally lower, verification is easier, and relationship dependence is weaker, catering to startups.

When we built autonomous systems on public chains, we learned similar rules: smart contracts can only act on verifiable matters. Anything that cannot be verified must be handled by humans in an enhanced manner off-chain, typically through governance. Cryptographic tokens follow similar rules: they excel at rewarding verifiable quantities (computational power, staking, liquidity) but are not good at rewarding subjective quality; this is why cryptocurrency successfully launched financial markets yet has not managed to tackle Airbnb or Uber. The lessons here are increasingly applicable for judging where true autonomous systems can scale.

Today, true automation is more constrained in application scope and market reach. Startups focusing on automated workflows often fit best in narrow beachhead markets, selling to startups or niche markets that are less resistant to changing ways.

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

Over time, we believe that automation solutions will gradually encroach on markets that are more reliant on relationships and have higher responsibilities because model capabilities are expanding, verification costs are decreasing, human-AI interactions are crossing the uncanny valley, and cultural acceptance is rising. In such scenarios, the first-mover data advantage gained from beachhead markets may compound, allowing startups to grow alongside the market itself.

The patterns and long-term attractiveness of automation versus enhancement will continue to diverge across different industries. Take education as an example. Alpha School is a vertically integrated AI private school that replaces teachers with "guides," facilitating a fundamentally automated learning process. While we believe AI-assisted education will be extremely important, we think excessive automation and vertical integration will not expand beyond niches because parents worry about teacher-student relationships, and schools are concerned about responsibility. Most of our educational systems are rigid (and mostly public) institutions that are unlikely to change quickly. Even top private institutions face the innovator's dilemma. Therefore, focusing on enhancement solutions seems more likely to yield major results, but they will also face competition from existing players.

Recruitment, on the other hand, points in the opposite direction: we are more optimistic about automation. In this vertical field, Harvey compares to Crosby: Juicebox sells candidate sourcing software to recruiters, while Prism is itself a recruiter—give it a job brief, and it delivers candidates ready for interview, charging only when someone is hired. Its key performance stages are easy to verify (response rates, interview performance, hire rates), and the responsibility is relatively low. Hiring is highly relationship-dependent, but the recruitment process leans more upstream (mainly finding candidates and converting them into interviews). This is also a function that companies often outsource, so the first batch of customers is startups: they prefer to buy results rather than build teams. Large enterprises have internal recruiters that tend to choose enhancement tools, as they are less willing to automate themselves away. The target customers are startups, which is a narrow wedge, allowing automation solutions to integrate into data-rich processes, improve over time, and grow with the market.

There are two types of AI companies: enhancers and automators. Both can expand human autonomy, but in different ways. Enhancers increase leverage for people within existing institutions: more knowledge, more output, more agency. Automators expand access: expert services that were once only affordable for businesses become accessible to everyone, freeing up time for more ambitious pursuits. The latest opportunities are not just to sell software but to become the service itself.