Series continuity: from personalized to represented-by

Part 5 of the series examined what happens once a system knows a person well enough to serve or steer them. This article pushes that question one step further: what happens when the entity actually interacting with your product isn’t the person at all, but an AI agent acting on their behalf? Every principle this series has covered so far—legibility, calibrated trust, forgiveness, ethical personalization—assumed a human was on the other end of the screen, reading, deciding, clicking. That assumption is now optional.

The core concept: the rise of the machine customer

“Seamless autonomous collaboration and dynamic workflow orchestration.” — Anushree Verma, Analyst, Gartner

She was describing where enterprise software is heading, and the phrase quietly names the scale of the shift: products are moving from tools a person operates to infrastructure that other software operates on a person’s behalf. Gartner has a specific term for the entity now showing up in that role—the “machine customer”—an autonomous or semi-autonomous non-human economic actor that can negotiate, purchase, and transact without a human present in the loop for each step.

The statistics nobody in the room wants to say out loud

Gartner’s top strategic prediction for 2026 states that 90% of B2B buying will be AI-agent-intermediated by 2028, pushing more than $15 trillion through agent-to-agent exchanges—roughly half of US GDP moving through machines transacting with other machines. A Gartner survey of 646 B2B buyers in March 2026 found 67% now prefer a representative-free purchasing experience entirely, and 45% had already used AI during their most recent purchase, with AI often in the loop before a human sales rep was ever contacted. Separately, total transaction volume from autonomous machine customers could reach $30 trillion by 2030. And Gartner’s own contact-center forecast put a number on how soon this reaches customer service specifically: a fifth of all customer contacts handled by service teams were projected to come from machine customers, not humans, before the end of 2026. The user your product was designed for is, in a fast-growing share of interactions, no longer the one actually using it.

The India-specific reality: UPI becomes the first national rail built for machine customers

India is not waiting for this shift to arrive—it’s building the infrastructure for it directly into its payments backbone. The National Payments Corporation of India is developing a Unified Agent Protocol (UAP), reported in July 2026, that would let AI agents register, get verified, and transact autonomously across UPI—the same rail that already processes over 23 billion transactions a month—without altering the underlying payment infrastructure itself. The design is deliberately bounded: a fully delegated agent is expected to be capped at roughly ₹5,000 per transaction and ₹15,000 a month, building on UPI Circle’s existing delegated-payment feature, with a full audit trail and identity checks behind every agent-initiated payment. India isn’t waiting for the protocol either—Pine Labs launched P3P in June 2026, described as India’s first agentic payment protocol live in production on UPI, and Mastercard demonstrated the country’s first authenticated agent-led transaction that February, in partnership with Axis Bank, RBL Bank, and merchants including Swiggy and Zepto. A 2025 pilot between Razorpay, NPCI, and OpenAI had already tested ChatGPT-driven UPI payments months earlier. The user for these products, in a growing number of real transactions, is a piece of software with a spending cap—not a person looking at a screen.

The psychology: diffusion of responsibility, one layer removed

Social psychologists Bibb Latané and John Darley documented a well-established phenomenon called diffusion of responsibility—the tendency for an individual to feel less personally accountable for an outcome the more that outcome is distributed across a group or a process rather than resting visibly on them alone. Delegating a decision to an AI agent produces a strikingly similar effect, even when the “group” is just the person and their own software: a user who authorizes an agent to negotiate a purchase within a spending cap feels less personally responsible for the specific outcome than if they’d typed in the card number themselves, even though they set the exact boundary the agent operated within. That’s not necessarily a flaw—it’s precisely the convenience machine customers are built to offer. But it also means the psychological distance a well-designed agent creates between a person and their own decision is a real design force, capable of making people comfortable authorizing transactions and disclosures they would hesitate over if asked to do so directly.

Where this is already visible

NPCI’s still-forming Unified Agent Protocol is a direct design response to this exact tension—hard spending caps and mandatory audit trails exist specifically because diffusion of responsibility makes unlimited delegation dangerous by default. Pine Labs’ P3P shows the alternative already live: a single upfront authorization, after which the agent transacts independently within its defined scope. Mastercard’s Agent Pay framework, tested with Swiggy and Zepto, extends the same principle to everyday recurring purchases—the exact “frequent, low-value” use case regulators and platforms alike are converging on as the safest place to start. And Gartner’s machine-customer forecast for contact centers shows the pattern reaching well beyond payments, into every customer interaction where a service team may increasingly be talking to software representing a person, not the person themselves.

LucyUX applied

  • At Listen, the discipline shifts from listening to a human end-user directly to listening for what the human principal behind an agent actually authorized versus what they assumed the agent would do—a gap that widens the moment delegation feels convenient enough to stop double-checking.
  • At Understand, the job is modeling two distinct users at once—the AI agent executing the transaction and the human who remains legally and financially accountable for it—because designing only for the agent’s efficiency, the way most agent-to-agent protocols currently do, ignores the person diffusion of responsibility, which is quietly distancing from the outcome.
  • At Conceptualize, hard boundaries—spending caps, category restrictions, mandatory audit trails, the exact pattern NPCI is building into UAP—have to be first-class interface elements a human sets deliberately, not defaults an agent inherits silently.
  • At Yield, success is a human principal who, reviewing their agent’s activity a month later, recognizes every transaction as something they would have knowingly approved—not a growing list of purchases that technically stayed within a limit but drifted from actual intent.

Consequences

When the user isn’t there, accountability doesn’t disappear—it just gets harder to trace, and diffusion of responsibility means the person who set an agent loose within a spending cap will, psychologically, feel less implicated in what it does than the facts actually warrant. NPCI’s hard limits and audit-trail requirements exist because regulators have already concluded that convenience alone isn’t sufficient design—the boundary between delegated authority and unaccountable authority has to be engineered in, not assumed.

Up next in the series: The Black Box in Your Pocket. Users don’t need to fully understand how an AI system works, but they do need to trust the decisions an agent now makes on their behalf without asking first. Part 7 looks at why explainability is a design problem, not an engineering one.

References & further reading