Digital Agent as Agent-of-Record

The setup

In 2016, roughly a decade before agent was a common label for software, Eric Glyman's earlier company Paribus claimed price-drop refunds on behalf of its users by sending mail from each user's own email account.¹ As the company grew by roughly forty percent a month, one large retailer's customer service went from rarely fielding a price-adjustment request to, in Glyman's words, "tens to hundreds to thousands of emails every single day" on that single topic, and cease-and-desist letters followed.

The argument that held

Glyman's team concluded that, under both store terms of service and existing law, having an agent act on someone's behalf was already normal: a lawyer can represent a client, a sibling can return a shirt someone else bought. Paribus was acting as an agent in exactly that sense. "What was different? It was a digital agent." Nothing in any store's terms or in the underlying law required an agent to be human. The threats also probed the company's security posture, asking how the data was being stored, a question a clean AWS architecture answered, and Glyman's own assessment was that "there wasn't a great legal case" against the company.

The part that still hurt

Winning the legal argument did not matter, because retailers had a cheaper move available: deleting the policy. Rather than fight the claims, some retailers simply concluded it was easier to remove the price-adjustment guarantee altogether. On Paribus's internal dashboard this did not appear as a single overnight cut but as a slide, refunds slowing by two percent, then five, then ten, as individual merchants withdrew their guarantees one at a time. The company lost roughly eighty percent of its revenue.

Why it generalizes

The structural lesson Glyman draws for agent builders today has four parts. Legal permission is not the binding constraint: the agency relationship itself was defensible, but the counterparty's willingness to keep offering the underlying benefit was not. Agents change the economics of a policy: a price-adjustment guarantee is cheap when claiming it costs a customer thirty minutes on the phone, and expensive once an agent claims it automatically for everyone, so any consumer right or rebate that survives mainly because it is annoying to exercise is at risk of being withdrawn once an agent industrializes the claim. Scale is what triggers the response, not principle, since nothing happened while the company was small. And acting through the user's own credentials, in this case their own email account, is what made the agency framing hold up, and remains the design most consumer-facing agents still use. The broader implication is a pessimistic counterweight to more optimistic readings of an agent-driven economy: the likely equilibrium may not be agents claiming everything owed to consumers, but the things owed to consumers quietly ceasing to be offered at all.