Summary

Eric and John open with a question John raised over lunch: is AI actually too expensive for some businesses? It sounds simple, but the answer turns on three distinct problems most companies never separate: whether people actually know how to use AI well, whether you can honestly measure the return, and what you are actually paying versus what you think you are paying.

They work through each one in order. On the usage side, they argue that buying licenses and hoping for adoption is a recipe for low ROI. Power users are rare, and the gap between someone who uses AI constantly but ineffectively and someone who uses it to think better about hard problems is enormous. On the ROI side, they draw a sharp line between cost savings (which are measurable) and revenue attribution (which is often fuzzy), and point to prospect research and faster creative iteration as two of the clearest paths to a direct revenue connection.

The conversation lands on the cost structure itself. Most businesses default to the most powerful and expensive models for every task, without realizing that cheaper models handle routine work just as well and can cost orders of magnitude less. John's story about using a flagship model to rewrite prompts for a cheaper one captures the whole episode's argument: with the right approach, AI is rarely too expensive. Without it, you are paying full price for a fraction of the value.

Key takeaways

• AI without adoption is just a sunk cost: Buying licenses does not create leverage. Most employees will not use AI well without deliberate training and incentives, and the power users tend to already be power users of other software.
• Using AI to think is the highest-leverage move: The biggest gap is not between people who use AI and people who don't. It is between people who use it to execute tasks and people who use it to think through bigger, harder problems.
• ROI has two sides, and cost is the easier one: Measuring hours saved and seat count reductions is straightforward. Attributing revenue gains to AI is harder because process improvements and business discipline often deserve as much credit as the tool itself.
• Start ROI tracking with use cases that have a clear line to revenue: Prospect research, faster creative iteration, and personalized sales demos are examples where the connection between AI effort and business outcome is concrete enough to measure.
• The default model is almost always the most expensive one: AI providers set flagship models as the default, and most business users never change them. Simpler tasks like reading a PDF or summarizing text work fine on models that cost a fraction of the price.
• You can use a smarter model to optimize for a cheaper one: If a task fails on a lower-cost model, asking the expensive model to rewrite the instructions for the cheaper one often solves it, and then you run all future instances on the cheaper version.
• Businesses on prosumer plans are sitting on a narrow window: Individual and small-business tiers are still heavily subsidized by providers preparing for IPO. That subsidy will shrink as these companies move toward profitability.

Notable mentions and links

• Klarna is the go-to example of high-profile AI cost savings: the company announced its AI assistant had replaced the equivalent of 700 customer service roles, then later reversed course and began rehiring human workers, illustrating how easy it is to overclaim AI ROI.
• Claude Cowork is mentioned as an example of a tool that lets users connect AI directly to local files on their computer, making analysis of large or disparate file sets faster and more effective than routing everything through cloud-based integrations.
• OpenAI's AI Advisors program and Anthropic's service partners represent the consulting arms both companies launched around the time of this episode, which Eric and John read as confirmation that most businesses need hands-on help to actually get ROI from AI, not just access to the tools.
• Colossus, the massive AI supercomputer built by Elon Musk's xAI in Memphis, comes up because Anthropic struck a deal to rent compute capacity there, an arrangement John calls "strange bedfellows" given that xAI and Anthropic compete directly in the AI market.
• "Token maxing" is a term the tech industry uses to describe the behavior of incentivizing employees to use as much AI as possible, which Eric and John acknowledge can be wasteful but still argue is better than no adoption incentive at all.
• ChatGPT and Gemini are discussed alongside Claude as the three major AI platforms businesses evaluate, with John noting that each has taken a different approach: Google optimizes for speed and scale, Anthropic focuses on enterprise, and OpenAI tries to serve everyone. These trends are visible in production data: the Vercel AI Gateway production index shows Anthropic leading on spend while Google leads on token volume, exactly the premium-vs-cheap model split Eric and John describe.