In 1998, Pattie Maes’ group at the MIT Media Lab published a survey of “agent-mediated electronic commerce” in The Knowledge Engineering Review: software agents that identify needs, compare merchants, negotiate, and purchase on a person’s behalf. Read today, it is less a research agenda than a summary of this year’s business press. The vision sat waiting for twenty-five years because the agents couldn’t read messy websites, handle ambiguity, or be trusted with money. Then, within roughly eighteen months, all three constraints moved at once.
The infrastructure record of 2025 reads like an industry synchronizing its watches. In April, Mastercard launched Agent Pay — tokenized credentials bound to a specific agent’s identity with per-session limits — and Visa launched Intelligent Commerce one day later. In September, Stripe and OpenAI shipped the Agentic Commerce Protocol, whose core primitive — a shared payment token scoped to one merchant and one cart total — lets an agent pay without ever holding the buyer’s credentials. Amazon’s “Buy for Me” grew from about 65,000 items at its April pilot to over half a million by year-end; Perplexity wired PayPal into in-chat checkout before Black Friday.
The buyers arrived on schedule. By Cloudflare’s network data, automated traffic crossed 57% of web requests this year, passing human traffic for the first time — roughly eighteen months earlier than Cloudflare’s own CEO had predicted. Imperva’s annual bot report puts the figure lower (53% for 2025) on a different methodology, but the direction is not in dispute, and the composition is the story: the growth is agents — crawling, querying, comparing, and increasingly transacting.
Gartner has been calling these buyers “machine customers” — “custobots” — since well before the infrastructure existed, and its analysts’ reported projections (executives expecting 15–20% of revenue from machine customers by 2030) should be read as analyst forecasts, not measurements. The payment rails, however, are not forecasts. They shipped.
I run customer success for a living, and I have written before about CS for AI companies. This essay is the inverse question, and the more interesting one: a fast-growing share of every software product’s “users” are no longer people but agents acting on people’s behalf, and customer success — a discipline built entirely around human relationships — now needs a second half.
The academic literature on delegation spent those years assembling the operating manual. People delegate more readily to agents that feel more human — anthropomorphism measurably increases the willingness to hand over decisions. Trust in these systems behaves like relationship trust rather than tool trust, which cuts both ways: a 2024 study in the Journal of Business Research found that when a virtual assistant’s recommendation goes wrong, consumers experience it as betrayal, with damage to purchase intent that a mere tool failure doesn’t produce.
And on the vendor side, the B2B literature has matured past folklore: peer-reviewed work now links customer-success management and customer-health monitoring to retention outcomes in exactly the way practitioners always claimed. (While we are near the folklore: the beloved “5% retention lifts profit 25–95%” line traces to Reichheld and Sasser’s 1990 HBR article — real, but a consultancy’s client sample from the George H. W. Bush administration, not a law of nature. Cite it as heritage, not physics.)
Set the delegation research beside the traffic data and the strategic picture sharpens. The human doesn’t disappear from the account; the human recedes into the principal’s chair. The agent executes; the human sponsors, budgets, and feels betrayed when things go wrong. The account has two layers now, and they want different disciplines.
Usage-based health scoring silently assumed every session was a human realizing value. When half the traffic is agents, that assumption breaks in both directions: agent activity can inflate a “healthy” account whose human sponsors have quietly disengaged, and an account where humans delegated successfully can look like fading logins. The fix is the one running through this whole series: separate the signals, and tie health to outcomes the human layer would testify to, not to telemetry volume.
For the agent layer, onboarding means discoverability and reliability: clean APIs, machine-readable product descriptions, predictable failure modes. A vocabulary is forming here — hosting vendor Netlify has been evangelizing “agent experience” (AX), and Shopify now ships every storefront with agent-facing description and transaction endpoints by default, per trade reports. “AX” remains a vendor coinage — where analyst firms have picked up the term this year, they mostly mean something different (employee-facing agent experience), and the academic literature hasn’t formalized either sense — but the underlying observation is sound, and CS teams will feel it first: the agent that cannot parse your product routes its principal’s budget somewhere it can.
The betrayal study is the tell: when a delegated agent fails, the grievance lands with a person, amplified. Fewer human touchpoints will carry more relational weight, which makes the remaining ones — the escalation, the incident review, the renewal — less automatable, not more. The machine-customer era re-centers the human craft of the job while automating its clerical body, an arc that will be familiar to readers of my retention-engine essay: the system proposes, the human disposes, and the relationship is the part that was never the machine’s to keep.
The QBR of 2028 may well have an empty chair where the daily user used to sit. The user of record is a process, and a process doesn’t take meetings. The renewal will still be signed by the person who owns that process, on the strength of outcomes they can defend and failures that were handled like relationships rather than tickets. The agent economy is splitting customer success in two rather than replacing it, and the teams that notice which half they’re in will keep both.
References