Marginal Cost of Knowledge to Zero

Two framings of the same force

The idea surfaced on Cheeky Pint in two related forms. Alex Rampell's version is concrete: his wife used an AI model to argue with a company over a billing dispute, the company used an AI model to argue back, and the dispute resolved in her favor. His framing is that the marginal cost of arguing has gone to zero, and once it does, everybody argues. Card networks carry breakage models where, if a price drops, a merchant technically owes a customer money, but nobody used to file a five-dollar dispute because the effort exceeded the payout. When the cost of arguing approaches zero, that changes.¹

Eric Glyman offers a sharper abstraction of the same force: it is not really the marginal cost of arguing that is collapsing, it is the marginal cost of time and knowledge. Most Ramp customers do not have a single software engineer, let alone one dedicated to finance, and the bottleneck was never desire, it was the cost of the cognitive work itself. A task that used to cost roughly five dollars of human time, and so went undone, now costs pennies of tokens.¹

Why it matters

This is the mechanism behind selling outcomes rather than software seats: once knowledge work approaches free, the addressable market shifts from software licenses to work actually performed. Glyman's example is Ramp selling "sets of work," expenses done, accounting done, embedded inside a financial operating platform, capturing a small slice of a large amount of customer value while doing work that customers would never have paid a person to do directly. Ramp's own policy agent, processing roughly one hundred thousand expenses a day at 99 percent accuracy, is that mechanism already running in production.¹

It also predicts an explosion of previously uneconomic activity: disputes, audits, reconciliations, and micro-negotiations that were never worth a human's time become worth an agent's time. Whether both sides arming agents against each other nets out to noise or to a fairer equilibrium, where small parties finally pursue what they are actually owed, is unresolved. And it sharpens the question of who captures the resulting value: if arguing itself is free for everyone, the edge is not the agent, which commoditizes quickly, but the proprietary data and outcomes the agent is able to act on.

Tensions

If every party's agent argues with every other party's agent, it is not obvious whether the system settles into a fairer equilibrium or simply burns computational resources in an arms race that merchants and issuers eventually price back into their models. And Glyman's own reframing matters here: the durable value is doing the underlying work, not winning any particular dispute. Disputes are a visible special case of cheap cognition, not the main event.