The MAD Podcast interviewed Stripe's Emily Sands on the emerging commerce stack for agents: buying inside AI surfaces, merchant catalog exposure, consumer authorization, agent payment execution, and business discoverability.

Stripe frames agentic commerce as a spectrum, from autonomous machine payments to human-led purchases inside AI surfaces. Merchants need products, prices, and inventory to be readable and transactable by AI systems.

Sands positioned the token as wallet- and processor-agnostic. Strategically, Stripe is trying to define a protocol layer for agent payments, not merely provide one-time virtual cards.

One-time virtual cards worked for scoped single transactions. Link Wallet targets more complex agent work with budgets, scope, and approval rules. Payment shifts from credentials to policy.

Humans do not want to enter card details for tiny purchases, and card economics break at tiny amounts. With agents executing automatically and balances or stablecoins settling, microtransactions for inference, data, or research become plausible.

The episode highlights signup abuse, free-credit abuse, and trial-card abuse. Unlike classic SaaS, AI providers often incur real inference costs before failed billing is discovered.

Usage-based billing lets abusive users generate large costs and disappear at settlement. For AI companies, this is not only lost revenue; it directly consumes margin and compute capacity.

Major AI companies increasingly treat token abuse as a margin and scaling problem. Beyond model quality, abuse prevention, limits, billing, identity, and payment strategy become core AI infrastructure.