Artificial intelligence applications must outperform customers’ alternatives to sustain premium prices as improving models make some software functions easier to replace, investment firm Andreessen Horowitz said in an analysis published Nov. 2.
Lower costs to run models can improve application margins. They can also make competing tools easier to build, author Tugce Erten said in “How AI Applications Earn Their Premium.” The analysis draws on company examples, buyer discussions and an August 2026 survey, rather than a study covering a single period.
In Andreessen Horowitz’s survey of 100 enterprise decision-makers responsible for large language models, 94% expected spending on those models to rise. Another finding was that 37% expected spending to at least double over the following year.
Erten said customers need evidence of how much work an application completes, its quality and its total cost compared with alternatives. A buyer can weaken a vendor’s pricing power by bringing just one frequently performed task in-house, without replacing the entire application, she said.
At Repsol, a blind evaluation compared legal AI provider Harvey with Microsoft Copilot and an internal prototype called Lexia, the analysis said. Harvey produced the best output and then beat the alternatives on accuracy and time savings during an eight-week pilot involving 50 lawyers across 12 jurisdictions.
Repsol subsequently rolled out Harvey across its legal department, the analysis said. Lawyers now save four to six hours a week, and adoption has reached 96%, Harvey said. Erten said the team also used past agreements to identify recurring objections and improve contract templates.
Erten said specialised applications can justify higher prices by applying company standards, keeping document edits consistent and retaining context through negotiations. She said repeated use can improve subsequent results when an application learns from completed work, but vendors need to measure those gains.
Access to information and the work needed to prepare it offer another basis for charging customers, Erten said. One unnamed industry software company had collected years of proprietary pricing benchmarks across roughly 3,000 companies. The analysis said spreading the cost of maintaining that information across customers could make buying access cheaper than rebuilding it individually.
Erten said vendors should count implementation, human review, failures and maintenance when comparing their applications with customers’ alternatives. In another example, Notion chose Decagon after considering an internal build. Decagon’s published case study reported an improvement of up to 34% in support ticket resolution time.
Erten said prices should reflect recognisable completed work, such as a contract review or resolved support conversation, with failed attempts included in delivery costs. The analysis cited Clay’s experience that fixed AI prices overcharged simple workloads while making complex ones unprofitable.
Clay separated charges for data and workflow coordination, keeping fixed prices for more predictable AI work and passing through variable model costs without a markup, the analysis said. Erten said cheaper delivery can support better margins, lower prices or more work, provided the complete offer stays attractive against customers’ improving alternatives.
