The Customer Has an AI Agent

For thirty years, customer experience has rested on an assumption so basic that nobody bothered to write it down: the customer is a person. A person saw the ad, ran the search, compared the options, filled in the form, called support, and decided whether to come back. Every channel we built — web, mobile, store, contact center, chat — was a place for that person to participate.

We designed customer experience for participation. AI is changing it to representation.

That sounds like a small shift. It isn't. It changes who can act on each side of the relationship, and that changes most of what companies have spent three decades learning to optimize.

Representation isn't new. Abundance is.

Very little that matters in modern life happens through direct participation. We hire lawyers to negotiate for us, advisers to invest for us, brokers to buy houses for us. Companies send procurement teams to buy on their behalf. Representation is one of the oldest ways people and organizations act beyond their own time and expertise.

What has been rare is access to it. Representatives are expensive, so most people handle most transactions themselves, and most companies save their best representatives for their biggest customers.

AI removes that scarcity. Every customer can now have a representative that searches, compares, negotiates, and completes work. So can every enterprise. Software acting on a person's behalf isn't a new idea — Pattie Maes was writing about agents that reduce work and information overload in 1994. What is new is that the capability is becoming cheap, general, and present on both sides of the same transaction at the same time.

Three questions

The shift is easiest to see as three questions. Most of the current conversation only asks the first one.

What happens when the enterprise has an agent? The company starts to participate through an actor that can understand a request, apply policy, change a booking, approve a refund, or decide to escalate. That agent becomes a representative of the brand, and its behavior communicates the company's judgment as surely as its advertising does. The design question stops being what the interface says. It becomes what the actor is allowed to understand, decide, do, disclose, and hand back to a person.

What happens when the customer has an agent? The customer stops visiting every page and filling in every form. An agent that knows their constraints, budget, history, and accessibility needs does it for them, across every company they deal with. That agent may never see the landing page anyone designed. It reads structured data, policies, availability, and track record. The person is still the beneficiary, but the direct user of the company's surface is now software acting on their behalf.

What happens when the customer is an agent? Sometimes no person is present at all. A vehicle buys charging. A production line reorders parts. A software service provisions more capacity. Gartner has been calling these machine customers for years, and they are real — but they are one condition in a larger picture, not the whole of it.

Agency on both sides at once

Put the three questions together and a pattern appears that none of them shows alone: agency is increasing on both sides of the relationship at the same time.

Enterprises are deploying agents to sell, serve, and resolve. Customers are acquiring agents to search, compare, and negotiate. The customer relationship is becoming a bilateral agency system — an interaction between two systems of representation, each with its own goals, data, permissions, and principal.

That breaks one of the most tempting ways to talk about this shift, which is as a ladder: human, then assisted, then delegated, then autonomous, as if every company climbs toward the top. It doesn't hold up. A company's agent serving a human customer and a customer's agent dealing with a human-run company aren't steps in a sequence. They're different configurations, and they exist side by side, often in the same industry, sometimes in the same week.

The better picture is a map. One axis is how much agency the customer holds. The other is how much the enterprise holds. The configurations fall where they fall.

A few things follow from drawing it this way. Human-direct experience doesn't disappear — it stays in its corner and remains the right mode for high-emotion, high-consequence moments. A customer can be fully represented while the company on the other side is still entirely human-operated, which is where many industries are heading faster than they expect. And the multi-party cases, where no single company can deliver the outcome alone, aren't a higher level. They're a boundary around all of it.

Omnichannel solved the last continuity problem

The last great organizing idea in customer experience was omnichannel. Its insight was that one customer uses many channels, and the relationship should follow them from store to web to app to contact center without starting over. Most enterprises still haven't fully delivered it. The idea was still right.

Bilateral agency creates the next version of the same problem. The question is no longer whether the relationship survives a change of channel. It's whether it survives a change of representative.

In a single week, the same customer might research something with an assistant, hand the purchase to their own agent, let that agent negotiate with the company's agent, step in personally for the one decision that matters, and leave a device to handle replenishment. Five modes, one customer. If the relationship resets every time the actor changes — if identity, intent, constraints, authority, and prior commitments don't carry across — the company hasn't built an agentic experience. It has built five disconnected automations.

I think of this as the omniagent problem. Omnichannel gave us one customer across many channels. The next era needs one relationship across many representatives.

Here's what it looks like. A traveler who uses a wheelchair is flying internationally. Their objective isn't to book a flight and a hotel. It's to arrive, stay independent, keep their medical and mobility needs intact, and recover fast if something goes wrong. They start with an assistant to research options and draft questions. They hand the booking to their own agent with hard constraints: step-free access, a roll-in shower, a transfer vehicle they can board without leaving the chair, power for their equipment. That agent works directly with the airline's, the hotel's, and the transfer company's agents. Then the flight is cancelled, and five organizations have to re-plan around the same constraints. The traveler should experience one recovered trip, not five service tickets.

More autonomy is not the goal

The same trip contains the counterexample that matters most.

When the hotel offers a substitute room, the traveler may want to confirm it themselves. Not because their agent can't handle it, but because hotel accessibility data is so often wrong, and the cost of trusting it is arriving at a room they can't use. At that moment the best experience is less autonomous, not more.

That's the principle I think the whole field needs to hold onto. The goal isn't maximum autonomy. It's agency fit: deciding, moment by moment, who should understand, who should decide, who should act, and who remains accountable. The person. The person with an assistant. The customer's agent. The company's agent. Or the two agents working it out between them.

None of this is new to human-factors research. Sheridan and Verplank were describing levels of automation in 1978, and adjustable autonomy has a long literature behind it. What customer experience adds is the relationship. The right fit depends on this customer, this intent, and these stakes, and it has to be able to change inside a single interaction. An agent should proceed on its own while confidence is high and the action is reversible, stop and ask when it meets a real trade-off, and hand control back the moment the consequences rise or the customer wants it.

It avoids two failures everyone has lived through. One is making a customer approve every trivial step. The other is leaving a system running on its own when the situation obviously needs a human.

The agents don't want the same thing

Most writing about agentic customer experience quietly assumes the two sides cooperate. Often they won't.

A customer's agent optimizes for its principal: lowest total cost, best fit, least risk, most privacy — and immunity to the persuasion techniques companies have spent decades refining. It doesn't see the landing page. It doesn't respond to scarcity cues. It compares every offer, every time.

A company's agent optimizes for its principal: margin, retention, utilization, and risk. Put the two in the same transaction and you don't get a journey. You get a negotiation.

The consequences aren't subtle. The funnel compresses, because the agent skips the stages that used to shape consideration. Brand splits in two: an emotional preference that may survive as an explicit instruction — prefer this airline if the whole journey stays accessible and the fare is within fifteen percent — and an operational reputation that has to be verifiable to count at all. Behavioral data thins out, because the customer's agent holds the context and the company sees only the request. And companies will be tempted to block, throttle, or charge for agent access, which is how the next round of platform fights begins.

The companies that depended on customers not comparing will feel this first. The companies whose value is real — reliable, fair, well described, easy to verify — may find that an agent-mediated market rewards them in ways the old funnel never did.

Trust has to become executable

When the customer's representative is software, trust can't stay a feeling. An agent can't act safely because an experience seems trustworthy. The system has to be able to establish, every time, who is acting, on whose behalf, with what authority, toward what outcome, with what consent, based on what evidence, under whose control — and who is accountable when it goes wrong.

Some of those layers are technical. Several aren't. Whether an agent can bind its principal to a contract is a legal question. Whether consent to share a preference extends to a negotiation is a policy question. Who pays when one company's agent makes a booking error on another company's system is a commercial question that most current contracts don't answer. The organizations that treat trust as metadata are going to discover it's liability.

What this asks of experience leaders

For decades, the job of an experience leader has been to represent the human inside business and technology decisions. That job hasn't gone away. It has grown.

It now includes designing for a participant that isn't a person: an authorized agent that needs to discover, understand, verify, and transact. Mathias Biilmann at Netlify has made the case for calling that agent experience, and he's right that it's a real design surface. It includes deciding, deliberately, where human judgment belongs. And it includes keeping one relationship intact while the actors inside it keep changing.

The practical agenda follows from the argument. Map where agency already sits on both sides of your customer relationships, including the parts nobody labeled AI. Decide agency fit for the moments that carry real consequences, instead of automating whatever happens to be automatable. Make your products, policies, and track record legible to the agents that represent your customers, because they will judge you on exactly that. Protect continuity across representatives the way the last decade tried to protect it across channels. And decide what you're willing to expose, negotiate, and let an agent commit to on your behalf, before the market decides for you.

The question that organized the discipline was what the customer needs. It still matters. The question underneath it now is who should act for them, and how that authority should work.

The future of customer experience is not human or machine. It is the deliberate design of the relationship between them.