AI Economics is a Design Problem

AI advantage will come from redesigning workflows around cost per outcome rather than maximizing AI usage or minimizing token spend.

• ICONIQ’s survey found that 71% of R&D employees have access to AI tools, 62% use them daily, but only 39% are power users, showing that broad access has not translated into deep, repeatable adoption.
• The most promising workflows are high-volume and repeatable, have digital inputs and outputs, a measurable business result and quality threshold, and are constrained by queues, handoffs or routine review rather than irreplaceable judgment.
• Ramp’s two-person Deal Desk manages thousands of contracts each month for more than 600 sellers and over 70,000 customers, with AI handling intake, pricing, terms, document generation and signature routing while humans handle exceptions.
• AI spending varies sharply: the top 1% of companies spend about $7,500 per employee per month, the top 10% about $630, and the median company just $12; 68% of surveyed builders say internal AI costs are very or moderately difficult to predict.
• Model selection should follow the workflow: 82% of surveyed companies use Anthropic and 71% use OpenAI, while roughly 4% to 7% use open-source models from DeepSeek and Moonshot AI; the average number of providers rose from 3.1 to 3.3 in six months.
• Coinbase says it halved AI spending while token usage increased by using routing, caching and lower-cost open models, illustrating that better system design can increase useful work without increasing the operating envelope.

Takeaway: Investors should assess whether an AI company can identify and redesign economically meaningful workflows, measure outcomes and route tasks across people, software and models; token volume, model preference and per-employee spending are inadequate standalone indicators of AI quality or value.