AI apps are being priced on proof now, not promise, and the model keeps raising the bar

The test

The argument from Andreessen Horowitz is blunt: to defend a premium, an AI vendor must outperform the customer's alternative across the full workflow, because model progress continually improves that alternative. An application is vulnerable, in this framing, when a customer can reproduce its core output with a general-purpose model and little ongoing effort. Lower inference costs can help margins, but they also lower the customer's cost of doing without you, so the premium has to be earned on workflow, not on access to a model. m:…7754

The firm's evidence is a legal pilot. In a blind evaluation, Repsol's legal team compared Harvey with Microsoft Copilot and its own internal AI prototype, Lexia, and Harvey delivered the highest-quality output. Harvey reports that Repsol lawyers now save four to six hours per week. The numbers are the vendor's own account, but the structure of the test is the point: a head-to-head against the customer's best alternative, judged on output quality rather than a feature list. m:…7754

Pricing follows the same logic. One vertical software company described in the piece had accumulated pricing benchmarks across roughly 3,000 companies and was exploring access through a Model Context Protocol server, turning proprietary context into the product. Clay separated Data Credits from Actions for orchestration, keeping fixed credit pricing for bounded workloads and passing through token costs for more variable reasoning models. When price is tied to recognizable work, a vendor can use routing, caching or better models to improve margins, instead of defending a token markup the customer can undercut. m:…7754

A robotics case study

Tutor Intelligence is the firm's live test of the same thesis in hardware. Andreessen Horowitz led the company's Series B, backing a robotics firm that deploys into real production environments ahead of full autonomy and uses human teleoperators, called tutors, to train its robot foundation models. The company is vertically integrated across its model, hardware and data collection, and both of its robot embodiments, Cassie for case picking and palletizing and Sonny for e-commerce fulfillment, run on the same model. m:…5947

The commercial terms are designed around the premium question. Tutor's hourly pricing model shifts spending from capital expenditure to operating expenditure, and implementation times run as low as 8 hours. The company plans to have more than 1,000 robots under contract by the end of 2027, and says nearly a third of customers who received robots in the past 12 months were existing customers adding robots, repeat behaviour suggesting the workflow, not the novelty, is doing the work. m:…5947

The company's chief executive, Josh Gruenstein, frames robotics as a systems problem requiring simultaneous innovation in AI, hardware and go-to-market strategy, and says the firm builds at scale in Watertown, Massachusetts. On the firm's podcast he describes a fleet of 100 Sunny robots in Boston used for internal data collection, an average of 40 to 60 days from first contact to a signed contract, and robots that work indoors and cannot operate outdoors. The United States deployed 38,000 industrial robot units in 2025, so Tutor's bet is not that robots are new but that a general-purpose fleet trained on real production work can reach customers single-purpose machines missed. m:…6942

Who keeps the judgment

Emergence Capital raises the question the pricing debate tends to skip: who owns the judgment the system accumulates. Every platform that captures the judgment behind a decision, the correction that improved a result, or the reasoning that closed a deal is accumulating something that used to leave with an individual, the firm argues, and enterprise contracts usually prohibit training on internal data while protection for behavioral traces and process knowledge is far less settled. Its proposed answer is a social framework for the intelligence economy, modelled on how radio and the internet settled norms before regulation hardened them, though the enforcement mechanism and any common cognitive passport format are not yet established. m:…6230

The stakes are visible in the firm's citations. A 2025 trial found undergraduates who used ChatGPT as a study aid scored significantly lower on a surprise retention test than students who studied the traditional way, while a Harvard trial of a tutor that withheld answers until students reasoned through found students gained more on tests right after the lessons than peers in a well-run active-learning classroom. Anthropic researchers found developers using AI scored 17 per cent lower on a quiz than those coding by hand. Systems that do the reasoning for the user erode the user's skill, while systems that force reasoning through preserve it, and that difference is what a premium product has to prove. m:…6230

For a venture reader, the two pieces fit together. McKinsey's 2026 survey found 89 per cent of organizations use AI in at least one business function but only 44 per cent report scaling it across the enterprise, and Andreessen Horowitz's survey of 100 enterprise decision-makers in August 2026 found 94 per cent expected total LLM spending to increase, with 37 per cent expecting it at least to double in the next year. Demand is not the constraint; proof is. A product that beats the customer's alternative on the full workflow, prices on recognizable work, and leaves the customer more capable rather than more dependent keeps its premium as models converge. m:…6230