High variance kills traditional SaaS unit economics

www.anjalishriva.com/pricing_paradigm

That’s a section of the super clear and insightful article by @anjali Shrivastava, “A broken pricing paradigm”. It was published more than 2 years ago and yet it is more relevant than ever in the “agentic era”. We have witnessed rate limits and skyrocketing costs at the big model labs (Anthropic, OpenAI) because a token is neither an atomic unit of cost nor an atomic unit of compute. “Traditional SaaS pricing mirrors the physics of stable software,” but AI introduces high or even infinite variance to the pricing equation: if task requirements AND number of tasks per day are high variance (fat tails: a small number of tasks and/or users dominate the costs) economic theories fall apart with possibly unlimited risk.

The shock to the SaaS model requires a complete change in the pricing mindset. It is not like SaaS companies that have been thriving can now compete without AI. Gone is the zero marginal cost. Plus, the cost distributions per user are power laws with long, long tails.

Now, layer on top of that the new agentic paradigm, even in its simplest form, one agent with multiple tool use. The agent must discover the tools it needs, find out where the data it needs might be and in what form, explore an undefined number of dead-ends and spend an unpredictable number of tokens. Imagine what happens if it must interact with an unpredictable number of other agents that will also consume an unknown number of tokens and possibly interact with more agents, and turtles all the way down. The product of possibly many, possibly infinite variance distributions does not bode well for pricing.

AI agents, for all of their promises, have a number of issues to address (security, privacy, quality assurance, stability, determinism, ...) but perhaps the biggest of all is the unpredictability of resource use. Some early adopters discovered that letting agents loose could lead to outrageous invoices. The current solution is to have limits on how much an agent can use, but predicting resource utilization is a necessity.