AI application companies face a growing challenge: as general-purpose models become more capable and cheaper, customers can increasingly replicate their core features with minimal effort, threatening the premium these applications charge. A report published by venture capital firm Andreessen Horowitz warns that falling inference costs may boost margins, but they also make applications easier to replace. The central question isn't technical—it's commercial: what will buyers continue to pay for when their alternatives keep getting better?
The firm's analysis describes how vulnerability varies by product: an application is at risk when users can recreate its primary output using a general-purpose model with little sustained work. One purchasing executive at a worldwide professional-services company told researchers that when a direct prompt to a frontier model yielded superior results compared to a vendor's productivity platform, the organization saw minimal justification for paying the platform's markup. For projects involving confidential information, this firm had constructed an in-house AI system capable of operating models locally or on-site when clients prohibited third-party tools. The company was also testing open-source code and alternatives to commercial offerings. In an August 2026 survey of 100 enterprise large language model decision-makers conducted by the firm, 94% anticipated their total LLM expenditure would rise, while 37% expected it to at least double within the coming year.
The report identifies several capabilities that allow applications to defend their pricing: specialized knowledge in a domain, thorough integrations, dependable execution, and context that's difficult to obtain elsewhere. According to the authors, these elements function as pricing levers that can make purchasing substantially more appealing than building. Legal AI company Harvey provides an illustrative case: energy firm Repsol's legal division conducted a blind comparison pitting Harvey against Microsoft Copilot and an internal AI prototype called Lexia. Harvey produced the highest-quality results, then surpassed competitors on precision and time efficiency during an eight-week trial involving 50 attorneys across 12 jurisdictions. Following full deployment, lawyers now conserve four to six hours weekly, with adoption hitting 96% throughout the legal department. The advantages also reach future assignments: the team employed Harvey to spot recurring objections across historical agreements and enhance contract templates with backup provisions.
The analysis explains why specialized applications can retain worth even when customers possess access to powerful models: vendors must outperform the buyer's alternative throughout the entire workflow, not just one isolated step. For contract examination, that encompasses applying a corporation's legal criteria, maintaining consistent edits across a document, and retaining context as negotiations progress. One vertical software company the firm interviewed had gathered years of proprietary pricing benchmarks across roughly 3,000 organizations and was exploring access through a Model Context Protocol server, allowing AI systems to query them directly. The economics are straightforward: the vendor can gather, structure, validate, and refresh the information once, then distribute that expense across numerous customers. Each buyer can lease reliable, current context for less than reconstructing and maintaining a comparable resource independently. The vendor captures a margin by removing duplicated effort; the customer obtains superior answers at reduced total cost. When accumulated learning enhances each successive result, the application's edge can compound through usage.
The report outlines three rules for sustained success: operate the workflow better than the customer can, make context cheaper to rent than to rebuild, and prove the value before pricing to capture it. Founders should benchmark their most vulnerable workflow against the customer's most credible alternative, including implementation, human oversight, failures, and upkeep. They must measure the complete cost of finished work, demonstrate its worth against the buyer's options, and select a pricing unit the customer can forecast. Collaboration platform Notion considered building internally, then selected external solution Decagon so teams could concentrate on the core platform. After assessing quality, integrations, security, and partnership, Notion chose Decagon, with the published case reporting up to a 34% improvement in ticket resolution time. The report concludes that model advancement will continually strengthen the customer's alternative, meaning an application's advantage must expand faster to keep earning its premium. For businesses navigating the shift from model novelty to operational discipline, the challenge isn't whether AI delivers—it's whether your version delivers enough more than what customers can assemble themselves to justify the markup quarter after quarter.