Selling tools or selling outcomes? AI companies are moving toward two completely different futures

Authors: Variant; Compilation: Deep Tide TechFlow

AI companies are splitting into two species: one sells tools to law firms, and the other becomes a law firm directly. In this framework article, Variant breaks down why some industries should build for augmentation while others should go for automation—the key is responsibility, validation costs, and dependence on relationships. For founders and investors, it’s a practical standard for judging the ceiling of an AI company.

The first wave of AI companies mainly focused on workflow enhancement. Many have already become the most successful businesses in the history of enterprise SaaS. By 2026, as model capabilities reach a tipping point, startups are changing their playbook. They no longer sell tools to incumbent players; instead, they disrupt them directly through automation. Many of these companies don’t look like SaaS at all—they look more like software as a service itself.

The contrast between enhancement and automation aligns with Chris Dixon’s classic framework: strong tech versus weak tech. Weak tech is kind of “objectified”—it recreates old ways with new technology, usually just more efficiently. Strong tech starts from first principles, fully leveraging existing resources to design what technology is supposed to look like.

Take legal services as an example. Harvey sells a monthly SaaS subscription to law firms, while the competitor Crosby is itself a law firm. When AI does legal work, human lawyers only step in where people are most needed: building and maintaining customer relationships, validating the work, and taking responsibility when things go wrong.

Companies that sell enhancement tools don’t want to disrupt their clients and organizational structures—that’s their lifeblood. Companies that sell automation are signaling that the world is about to change dramatically. Like service companies, AI automation companies are more inclined to sell outcomes rather than seat counts, token counts, or subscriptions.

Chris notes in the article that both strong and weak tech can succeed; often it’s just that the timelines differ, and both paths have already produced—and will continue to produce—big results. The real question is which path matches the current state of specific jobs, the market, and the evolution of technical capability.

At Variant, we use a simple framework to decide which workflows are suitable for automation and which for enhancement. Following it, you can see what kinds of vertical companies may emerge and where venture-scale opportunities are. We look at three dimensions:

1. Responsibility: How dangerous and expensive is it when AI makes mistakes?

2. Validation: How fast and how cheap can users verify work quality?

3. Relationships: How dependent is the experience on human interaction?

Work that has high responsibility, high validation costs, and heavy reliance on relationships is suited to enhancement. This framework makes the contrast between Crosby and Harvey obvious. The internal legal function of large companies fits all three: high responsibility, high validation costs (many open-ended tasks), and high dependence on relationships. Harvey or Claude Cowork–type enhancement tools make sense in this scenario. By contrast, Crosby focuses on more routine, high-frequency legal services such as commercial contracts; responsibility is usually lower, it’s easier to validate, and dependence on relationships is weaker—the customer base is startups.

We learned similar rules when building autonomous systems on public chains: smart contracts can only take action on things that are verifiable. Anything that can’t be verified must be handed back to humans for enhancement off-chain, usually via governance. Crypto tokens follow analogous rules: they’re extremely good at rewarding verifiable quantities (compute, staking, liquidity), but not good at rewarding subjective quality. That’s why crypto successfully launched financial markets but hasn’t solved Airbnb or Uber. This experience is increasingly applicable to figuring out where truly autonomous systems can be scaled.

Today, “real automation” is more limited both in application scope and in the addressable market size. Startups that focus on automating workflows are often best suited to narrow beachhead markets—selling to startups or to niche markets that aren’t as averse to flipping the playbook.

Starting from narrow wedges is a way to avoid the “kill zones” of big AI labs, which focus on horizontal expansion to enhance larger existing markets.

As time goes on, we believe automation solutions will gradually take over markets that rely more on relationships and have higher responsibility, because model capability is expanding, validation costs are falling, human-AI interaction is crossing the uncanny valley, and cultural acceptance is improving. In that situation, the early data advantage gained from beachhead markets may compound—letting startups grow together with the market itself.

The balance and long-term pull of automation versus enhancement will keep diverging across industries. Consider education. Alpha School is a vertically integrated AI private school that replaces teachers with “guides” to enable a largely automated learning process. While we believe AI-assisted education will be extremely important, we think excessive automation and vertical integration won’t expand beyond niche audiences because parents worry about the student-teacher relationship, and schools worry about responsibility. Most of our education system is rigid (and mostly public institutions), which is unlikely to change quickly. Even top private institutions face the innovator’s dilemma. Therefore, enhancement-focused solutions seem more likely to deliver big results—but they also face competition from incumbents.

Hiring points in the opposite direction: we favor automation more. This vertical shows a Harvey versus Crosby dynamic: Juicebox sells candidate sourcing software to recruiters, while Prism is itself a recruiter—give it a job brief, it delivers candidates prepared for interviews, charging only when someone signs. Its key stages are easy to validate (response rates, interview performance, offer-to-sign rates), and responsibility is relatively low. Hiring is highly relationship-dependent, but it’s upstream (mainly finding candidates and converting them into interviews). It’s also a function companies often outsource, so the first customers are startups: they would rather buy outcomes than build a team. The largest enterprises have internal recruiters, and they tend to choose enhancement tools because they’re less willing to automate themselves away. The target customer is startups—this is a narrow wedge where automation embeds into data-rich processes, improves over time, and grows with the market.

There are two kinds of AI companies: enhancers and automators. Both can expand human autonomy, but they do it differently. Enhancers give people more leverage within existing institutions: more knowledge, more output, more agency. Automators widen access: expert services that used to be affordable only to enterprises become available to everyone, freeing up time for more ambitious things. The latest opportunity isn’t just selling software—it’s becoming the service itself.