Andreessen Horowitz says AI software makers will need to demonstrate that their products outperform customers’ internal alternatives as improving models make some features easier to recreate.
In an analysis published Nov. 2, partner Tugce Erten argues that access to an AI model alone may not justify a software premium. She points instead to reliable workflows, prepared business context and measurable results. Falling costs to run models may improve vendors’ margins, while stronger models and easier development tools can help customers reproduce some features themselves.
An August 2026 survey cited in the analysis found that 94% of 100 enterprise large language model (LLM) decision-makers expected their overall spending on the technology to rise over the following year. Thirty-seven percent expected it to at least double.
Repsol’s legal department is one example in the analysis. Andreessen Horowitz and Harvey say Repsol conducted a blind comparison of Harvey, Microsoft Copilot and an internal prototype called Lexia. The account says Harvey ranked highest for output quality. Fifty lawyers from 12 jurisdictions took part in an eight-week pilot that also assessed accuracy and time saved.
Harvey later reported that 96% of Repsol’s legal department adopted its product, with savings of four to six hours per lawyer each week. Repsol said in an April 2024 announcement that its lawyers remained responsible for reviewing and validating the system’s proposals; at the time, it estimated savings of up to three hours per lawyer per week during implementation. Harvey’s customer page does not describe how it measured the later figures.
Andreessen Horowitz also says Notion considered building customer-service software internally before choosing Decagon. Decagon’s case study reports that ticket resolution time improved by as much as 34%, but gives no measurement period.
A pricing example comes from Clay’s March 2026 memo, which separated data charges from orchestration work and set charges for complex reasoning models according to token use, without a markup.
Erten says vendors may reduce the per-customer effort of preparing business context by using it across clients; buyers may spend less than an internal effort to create and maintain similar resources would require. The cited examples do not disclose the fees Repsol or Notion paid or provide a full comparison with the cost of their internal alternatives.
