Automate What You Can, Augment What You Can’t

The first wave of AI companies largely focused on augmenting workflows. Many have been among the most successful enterprise SaaS businesses in history. In 2026, model capabilities hit a tipping point, allowing startups to take a new approach. Rather than sell to incumbents, this second wave aims to disrupt them through automation. Many of these businesses do not look like SaaS, but software-as-the-service.

This dynamic of augmentation vs. automation fits Chris Dixon’s classic pattern of strong vs weak technologies. Weak tech is skeuomorphic; it uses new technologies to replicate old ways of doing things, often more efficiently. Strong tech, by contrast, “build[s] from first principles, making full use of the available resources to design technologies as they ought to exist.”

Take legal services as an example. Whereas Harvey sells a monthly SaaS subscription to law firms, a rival startup, Crosby, is a law firm. AI does the legal work and brings in human lawyers for help where humans are most important: building and maintaining relationships, verifying the work, and providing accountability if something goes wrong.

Businesses that sell augmentation don’t want to disrupt the customers and organizational structures they sell to. In contrast, businesses that sell automation are suggesting that the world should look much different than it does today. And like services firms, AI companies that automate have a strong incentive to sell outcomes instead of seats, tokens or subscriptions.

As Chris notes in his piece, both strong and weak technologies can be successful, often on different time horizons, and both approaches have and will continue to create big outcomes. The question is which approach best fits the work, market, and capabilities of the technology as it develops.

At Variant, we’ve been using a loose framework for identifying which workflows are ripe for automation vs. augmentation. Downstream of this, we can start to glean what new kinds of vertical companies might emerge and where we think venture-scale opportunities may open up in the future. There are three dimensions we look at first:

1. Liability: How risky and expensive is it when AI gets it wrong?

2. Verification: How cheaply and quickly can the user verify the work was done well?

3. Relationships: How much does the experience depend on human interaction?

With work that has high liability, is costly to verify, and is deeply relational, augmentation is the way. We can see this framework play out in the Crosby vs. Harvey dichotomy. An incumbent organization’s internal counsel has all three attributes of the framework: high liability, costly to verify (many open-ended tasks), and deeply relational work. Tools for augmentation like Harvey or Claude Cowork make sense in that context. In contrast, Crosby focuses on more routine, high-volume legal services like commercial contracting, which are generally lower liability, more verifiable, and less relational, and can service startups.

We learned a similar set of rules building autonomous systems on public blockchains: a smart contract can only act on what it can verify. Everything it can’t gets handed back to humans to augment offchain, often through governance. Crypto tokens obeyed a similar rule: extremely good at rewarding verifiable quantities (hashrate, stake, liquidity), extremely bad at rewarding subjective quality, which is why crypto has successfully bootstrapped financial markets but has not yet tackled the likes of Airbnb or Uber. There a number of lessons learned here that are increasingly relevant as we look to the future of where truly autonomous systems can scale.

Today, true automation is more limited in terms of application scope and addressable market size. Startups focused on automating workflows are often best positioned in narrow beachhead markets, selling to startups or niche markets that are less averse to doing things differently.

Starting with a narrow wedge market is one way to avoid the “kill zone” of big AI labs, which are focused on horizontal expansion to augment larger, incumbent markets.

Over time, we believe automated solutions will slowly eat market share of even the more relational, higher-liability markets as model capabilities expand, verification becomes less costly, and human interaction with agents crosses the uncanny valley, becoming more culturally accepted. In that scenario, first-mover data advantages from beachhead markets may compound, enabling startups to grow with the market itself.

The landscape and long-term attractiveness of automation over augmentation will continue to differ across industries. Take education, for example. Alpha School, a vertically integrated AI private school, replaces teachers with “guides” who facilitate a largely automated learning process. While we believe AI-assisted education will be incredibly important, we think excessive automation and verticalization will not scale beyond a niche audience given parents’ concerns over teacher-student relationships, and schools’ concerns over liability. The bulk of our educational systems are inflexible (and mostly public) institutions, which is unlikely to change any time soon. Even top-flight private institutions have an innovator’s dilemma to disruption. Therefore, solutions focused on augmentation seem poised for bigger outcomes but also competition from incumbents entering the space.

Recruiting points the opposite way: we favor automation. The Harvey vs. Crosby of this vertical: Juicebox sells sourcing software to recruiters, while Prism is the recruiter – brief it on a role, and it delivers candidates ready to interview, charging a fee only when someone signs. Its critical phases are easy to verify (response rate, interview performance, hire rate) and relatively low liability. Hiring is deeply relational, but recruiting is further up the funnel (principally sourcing candidates and converting them into interviews). It’s also a function companies often outsource, which is why the first customers are startups: they’d rather buy the outcome than build the team. The largest enterprises have in-house recruiters and will choose augmentation because they’ll be reluctant to automate themselves out of a job. The target customer being startups is a thin wedge that lets automated solutions embed themselves into data-rich flows, improve over time, and grow with the market.

There are two types of AI companies: augmenters and automators. Both can expand human autonomy, but differently. Augmenters give people more leverage inside the institutions where they already work: more knowledge, more output, more agency. Automators expand access: expert services once priced for corporations become more accessible to anyone, and the hours they consumed get returned for more ambitious use. The newest opportunity isn’t just to sell software, but to become the service.