How could an AI agent's revenue per transaction grow over time?

Jaipuria, writing in his newsletter, lays out four ways a personal AI agent can get paid — convenience fees from the customer, merchant or partner fees, affiliate-style referral fees, and advertising — then singles out transaction fees as the two "most interesting to quantify." He frames take rates as a spectrum already visible in commerce, running "from 0% on the low end... to maybe 30% if someone is purchasing something through Apple App Store or Steam," and argues an agent's own rate will depend on "the value that you are providing," how involved it is in customer acquisition, how much fulfillment, trust, and post-order work it does, and "how strong are the alternatives merchants have if they don't use you."

His mechanism is staged, not fixed. An agent starts by negotiating "some commission off the take rate that someone like a Stripe would be earning" — a payments layer he estimates at "sub-0.5%... maybe even on the order of 0.2%." It then layers in affiliate or referral-program fees, which vary sharply by category — "anything from ~0% on flights to 1% on Amazon grocery purchases to 4% on Expedia hotels and 8% on Viator experiences" — net of which he models roughly 4%. Finally, in categories where it's economical, the agent could push toward an aggregator-like role — taking on "inventory access, accurate prices, reliable bookings," becoming "merchant of record," and handling cancellations — which he models at roughly 12% net, citing Booking's aggregator take of "14% to 15% of GMV" against affiliate fees that "net out to ~5-7%." Combining the layers, he projects a blended take rate moving "from ~0.8% to 2% to 4% over time," while cautioning "there will always be a long tail without a commercial agreement in the medium term."

None of the nearest notes quantify take-rate structure directly. The closest is Will brands succeed by building their own AI shopping agents?, which argues from the brand side that agents become a decision layer merchants must win placement in; Jaipuria's framework describes the revenue mechanics behind that same decision layer, showing what winning or losing placement would cost a merchant at each stage. It also sits upstream of Can an AI agent serve both merchant and user interests fairly?: Azhar treats merchant-funded referral fees as already compromising an agent's loyalty, while Jaipuria's staged model shows referral fees are only the middle layer of a trajectory that, in his telling, ends with the agent closer to a booking aggregator with even larger stakes in which option it recommends.

These figures are Jaipuria's own illustrative modeling — "suppose you get 0.2%... as below" — not reported earnings from an operating agent. The per-category affiliate and aggregator percentages are described as published by existing programs and companies, but the blended three-stage progression is a scenario he constructs, not a measurement of any agent's actual revenue; the excerpt names two real data points elsewhere (Zuckerberg's stated intent to "profit by taking a small fee," and Instinct's reported >$1B annualized GMV) without tying either to a specific take rate. If agents do follow this staged path, it implies that today's near-zero commission fees are a starting point rather than a ceiling, and that competition for affiliate and aggregator status — not the initial payments fee — is where the larger revenue fight will happen.