The Better the Models Get, the Harder It Is for AI Apps to Make Money? 94% of Enterprises Are Still Increasing Budgets

With every model upgrade, the per-run cost of an AI application falls, and it becomes easier for customers to replace it. In the article, a16z's Tugce Erten offers a somewhat uncomfortable contrast: 94% of enterprise LLM buyers still expect spending to rise. Cheapness and spending more are happening at the same time—so how do AI applications earn a premium?

The chart puts two paths side by side. When customers build their own stack, they have to choose models and routing, connect company data, evaluate outputs, and maintain the system. When they buy from a vendor, they are buying workflow orchestration, context and integrations, evaluation and control, and ongoing model optimization. In the end, both are compared on the work completed: quality, speed, total cost, and risk. Value lies in the workflow, not in the model layer.

Tugce Erten offers three rules.

First, operate the workflow better than customers can themselves. Most customers don't want to build it; they want someone to run it and be accountable for the results. Legal AI platform Harvey beat general-purpose AI assistants and energy giant Repsol's in-house tool in blind tests. Blind tests provide a result customers can check, which is tougher than slogans. Notion once considered building its own support AI, then bought Decagon's AI support agent so its team could keep focusing on the core product. This is also the real-world version of Rule 1: what customers want is someone to run the support workflow. A commenter asked whether Rule 1 means winning AI applications will quietly turn into service companies. That question is not trivial. The closer you get to outcomes, the more you look like a service.

Second, make renting context cheaper than rebuilding it. Company history, processes, permissions, and evaluation criteria are expensive to rebuild. One commenter said nobody wants to teach a new tool company's entire history all over again. Stickiness comes from context, not from model parameters.

Third, prove value, then price to capture it. One commenter noted that Rule 3 is hardest for small teams. Without numbers customers can verify, a premium is hard to defend. Harvey's blind test is an example. The pricing anchor should be the quality, speed, total cost, and risk of the completed work, not the per-token price.

The counterintuitive part is this: model progress will keep improving customers' alternatives. An application's advantage has to expand faster; otherwise, the stronger the models get, the easier it is for customers to build it themselves or switch away. A commenter working on agents said token costs have actually gone up, with each upgrade making workflows more complex and burning more. A 94% expected increase in spending is not surprising. Cheaper per-call pricing can lead to more calls, more complex workflows, and higher verification costs.

Selling tools is not as good as selling validated workflows. But don't treat becoming a service business as a fallback. Serviceization requires that you can actually operate, deliver, and be accountable for results.

Original article from a16z's Tugce Erten: link

Reply citation: link