Brief

When Your Customer Sends an Agent

When customers delegate service to AI, companies must decide whether to block, tolerate, or serve the agents at the other end of the line.

By Rohit Mangal, Jurre Nitert van Schijndel, Bharat Bansal, Adam Barrowclough, and Corrie Carrigan

First published on noviembre 10, 2026
min read


Executive Summary Personal AI agents remove much of the effort that has historically constrained service demand, increasing exposure to contacts, concessions, and lost recurring revenue. Our discussions with customer experience leaders at major service organizations revealed that companies can’t reliably identify agent traffic, leaving them unable to measure the shift or manage its cost. Agent-originated traffic will become one of the lowest-cost contacts a company handles when routed to a direct machine-to-machine interface instead of human-based chat and voice channels. As consumers delegate more transactions and purchasing decisions to AI agents, companies that make it easier for those agents to do business with them will gain a commercial advantage. Personal AI agents, such as Meta’s Muse, OpenAI’s dots, and Google’s Gemini Spark, are contacting companies on customers’ behalf to negotiate bills, pursue refunds, and cancel subscriptions. Most service models are based on assumptions that some customers won’t pursue valid requests and others will give up before reaching a resolution. These agents weaken both assumptions by making service requests easier to initiate and pursue. That creates financial exposure, but also a significant opportunity: Although companies will face more requests and higher concession costs, they will be able to resolve many of those interactions inexpensively. Companies that act early will turn that traffic into one of their lowest-cost contacts and become the company AI agents prefer to do business with. Companies that wait take on the same rise in requests and concessions without being able to identify or efficiently handle them. How companies manage that trade-off depends on the actions they take now. What is changing One of the authors of this brief had an AI agent negotiate his cable bill from $110 to $80 in a single 40-minute call, a conversation he had been putting off for five years. The agent brought him onto the line to authorize the interaction, then handled the negotiation while he went about his day. When the representative tried to sell additional services, the agent declined and went back to reducing the price. The customer saved over $350 a year, while the company lost that amount and incurred $30 in service costs. Service demand has always been limited by effort. Customers might decide a small billing error isn’t worth a call, miss discounts for which they qualify, or give up on refund requests after waiting on hold too long. Contact forecasts, staffing models, and concession provisions have been shaped by these behaviors. They reflect what customers historically pursued, not necessarily everything they were eligible to request. Agent-mediated service challenges three assumptions embedded in those economics: Customer effort falls dramatically: The AI agent does the heavy lifting. Customers can delegate issues they once decided weren’t worth pursuing. AI agents know more than a typical customer: They can search underlying policies for benefits and exceptions customers might never discover. AI agents can be far more persistent: They wait, escalate, and come back the next day without the frustration and inconvenience human customers experience. In speaking with customer experience (CX) leaders at major service organizations about their experiences with AI agents, we found that no one could reliably identify what share of their inbound contacts originated from personal agents. Even sophisticated service operations lack the infrastructure to measure agent demand, leaving companies facing a shift they can’t yet quantify. Agents also behave differently once engaged. They choose their own channel, picking whichever route completes the task, and will often reach for voice because that is where a human can still be persuaded. In recent interviews, leading healthcare payer executives told us that specialized AI agents already make a significant share of their incoming provider calls and that their own AI agents sometimes handle them. Two machines are already negotiating over a channel built for people. The risks As agents take on more service requests, companies face risks that extend beyond higher contact volumes to customer relationships and responsibility for agents’ actions. Cost to serve rises on two fronts: Volumes will rise beyond contact forecasts and vendor commitments built around how much effort customers have historically been willing to make. Claims will increase, too. Companies often forecast refunds, goodwill adjustments, and retention concessions based on requests customers actually pursued. AI agents can pursue more of the requests customers were eligible to make but previously abandoned. Each request may be valid, so traditional exception reporting could miss the cumulative effect across many small financial line items. Customer experience and loyalty erode: Service interactions give people a chance to explain a problem in their own words, letting representatives hear the hesitation or frustration and repair the relationship. An AI agent conveys the request without that context, so companies lose insight into how customers feel and opportunities to build loyalty. Customer experience also suffers for customers who aren’t using AI agents. Agent traffic competes for the same finite capacity, so queues lengthen and response times slip. Authority and accountability need clear rules: A customer may authorize an agent to act but dispute a particular outcome. Companies should decide what evidence of authorization they will require, what an agent can commit a customer to without a second confirmation, and what happens when a customer disputes what their agent did. The opportunity The shift to agent-led service requests will create two important advantages for companies that are prepared to act. Cost savings: Agent traffic will become one of the cheapest contacts a company handles. A verified request can arrive structured, bypass the hold queue, and be resolved automatically when the transaction and authorization are straightforward. Give that traffic its own machine-to-machine lane, and the cost to serve will fall well below assisted channels. Leave it in the human queue and companies risk absorbing the additional volume without realizing the efficiency benefit. Companies need to connect those requests to systems that can resolve them automatically. Ease of doing business: Consumers are beginning to delegate shopping, booking, and provider selection to AI agents. As that behavior expands, agents will favor companies that make authorized transactions easy to complete. Companies that force cumbersome handoffs or block agent access will lose business to competitors that can complete those transactions more easily. Actions to take now Companies should take five actions to prepare, regardless of how quickly consumer AI agents gain traction: 1. Size your exposure: Identify the customers who could claim a benefit today, measure how few actually do, and model what happens when asking requires little effort. Include the financial benefit granted and the cost of handling the request. The results provide a basis for deciding which policies to review and where to invest in a different service approach. 2. Set your AI agent posture by customer journey: Decide which interactions to block, tolerate, or serve. Billing disputes, account changes, and retention conversations each need different solutions. Bring service, fraud, legal, technology, and operations teams together to agree on the approach. Tell representatives what it means in practice, including what they can concede and when to ask for the customer directly. 3. Build the detection infrastructure: A company that can’t identify agent contacts also can’t reliably route them, determine their cost, or forecast their volume. Many AI agents may not be obvious, but there are behavioral clues, including form-fill speed, session velocity, and repetition. Add emerging ways to verify agent identity for those that authenticate themselves. Tag these interactions to track and report their volume. 4. Review and enforce concession policies: Many service policies are affordable only because a small share of eligible customers claims them. AI agents change that calculation. Find the widest gap between who qualifies and who claims. Decide what the company is willing to pay at higher claim rates, then review discretionary benefits and limits accordingly. Check policies are applied consistently across channels. 5. Build the agent-to-agent channel: Provide a machine-to-machine interface with authentication, clearly defined permissions, and records of each action. Route it separately from the human channel. The interface alone isn’t enough. Eligibility and concession rules need to support automatic decisions and execution within set limits. Otherwise, structured requests will still end up in a human queue and won’t deliver the cost advantage. The cost of waiting Companies that haven’t decided how to handle consumer AI agents leave those decisions to existing controls and frontline judgment. That can push requests into expensive channels without stopping the demand. Companies should set their approach now, before rising volumes make it harder to change.

Executive Summary

Personal AI agents remove much of the effort that has historically constrained service demand, increasing exposure to contacts, concessions, and lost recurring revenue.
Our discussions with customer experience leaders at major service organizations revealed that companies can’t reliably identify agent traffic, leaving them unable to measure the shift or manage its cost.
Agent-originated traffic will become one of the lowest-cost contacts a company handles when routed to a direct machine-to-machine interface instead of human-based chat and voice channels.
As consumers delegate more transactions and purchasing decisions to AI agents, companies that make it easier for those agents to do business with them will gain a commercial advantage.

Personal AI agents, such as Meta’s Muse, OpenAI’s dots, and Google’s Gemini Spark, are contacting companies on customers’ behalf to negotiate bills, pursue refunds, and cancel subscriptions. Most service models are based on assumptions that some customers won’t pursue valid requests and others will give up before reaching a resolution. These agents weaken both assumptions by making service requests easier to initiate and pursue.

That creates financial exposure, but also a significant opportunity: Although companies will face more requests and higher concession costs, they will be able to resolve many of those interactions inexpensively. Companies that act early will turn that traffic into one of their lowest-cost contacts and become the company AI agents prefer to do business with. Companies that wait take on the same rise in requests and concessions without being able to identify or efficiently handle them. How companies manage that trade-off depends on the actions they take now.

What is changing

One of the authors of this brief had an AI agent negotiate his cable bill from $110 to $80 in a single 40-minute call, a conversation he had been putting off for five years. The agent brought him onto the line to authorize the interaction, then handled the negotiation while he went about his day. When the representative tried to sell additional services, the agent declined and went back to reducing the price. The customer saved over $350 a year, while the company lost that amount and incurred $30 in service costs.

Service demand has always been limited by effort. Customers might decide a small billing error isn’t worth a call, miss discounts for which they qualify, or give up on refund requests after waiting on hold too long.

Contact forecasts, staffing models, and concession provisions have been shaped by these behaviors. They reflect what customers historically pursued, not necessarily everything they were eligible to request. Agent-mediated service challenges three assumptions embedded in those economics:

Customer effort falls dramatically: The AI agent does the heavy lifting. Customers can delegate issues they once decided weren’t worth pursuing.

AI agents know more than a typical customer: They can search underlying policies for benefits and exceptions customers might never discover.

AI agents can be far more persistent: They wait, escalate, and come back the next day without the frustration and inconvenience human customers experience.

In speaking with customer experience (CX) leaders at major service organizations about their experiences with AI agents, we found that no one could reliably identify what share of their inbound contacts originated from personal agents. Even sophisticated service operations lack the infrastructure to measure agent demand, leaving companies facing a shift they can’t yet quantify.

Agents also behave differently once engaged. They choose their own channel, picking whichever route completes the task, and will often reach for voice because that is where a human can still be persuaded.

In recent interviews, leading healthcare payer executives told us that specialized AI agents already make a significant share of their incoming provider calls and that their own AI agents sometimes handle them. Two machines are already negotiating over a channel built for people.

The risks

As agents take on more service requests, companies face risks that extend beyond higher contact volumes to customer relationships and responsibility for agents’ actions.

Cost to serve rises on two fronts: Volumes will rise beyond contact forecasts and vendor commitments built around how much effort customers have historically been willing to make. Claims will increase, too. Companies often forecast refunds, goodwill adjustments, and retention concessions based on requests customers actually pursued.

AI agents can pursue more of the requests customers were eligible to make but previously abandoned. Each request may be valid, so traditional exception reporting could miss the cumulative effect across many small financial line items.

Customer experience and loyalty erode: Service interactions give people a chance to explain a problem in their own words, letting representatives hear the hesitation or frustration and repair the relationship. An AI agent conveys the request without that context, so companies lose insight into how customers feel and opportunities to build loyalty.

Customer experience also suffers for customers who aren’t using AI agents. Agent traffic competes for the same finite capacity, so queues lengthen and response times slip.

Authority and accountability need clear rules: A customer may authorize an agent to act but dispute a particular outcome. Companies should decide what evidence of authorization they will require, what an agent can commit a customer to without a second confirmation, and what happens when a customer disputes what their agent did.

The opportunity

The shift to agent-led service requests will create two important advantages for companies that are prepared to act.

Cost savings: Agent traffic will become one of the cheapest contacts a company handles. A verified request can arrive structured, bypass the hold queue, and be resolved automatically when the transaction and authorization are straightforward. Give that traffic its own machine-to-machine lane, and the cost to serve will fall well below assisted channels. Leave it in the human queue and companies risk absorbing the additional volume without realizing the efficiency benefit. Companies need to connect those requests to systems that can resolve them automatically.

Ease of doing business: Consumers are beginning to delegate shopping, booking, and provider selection to AI agents. As that behavior expands, agents will favor companies that make authorized transactions easy to complete. Companies that force cumbersome handoffs or block agent access will lose business to competitors that can complete those transactions more easily.

Actions to take now

Companies should take five actions to prepare, regardless of how quickly consumer AI agents gain traction:

1. Size your exposure: Identify the customers who could claim a benefit today, measure how few actually do, and model what happens when asking requires little effort. Include the financial benefit granted and the cost of handling the request. The results provide a basis for deciding which policies to review and where to invest in a different service approach.

2. Set your AI agent posture by customer journey: Decide which interactions to block, tolerate, or serve. Billing disputes, account changes, and retention conversations each need different solutions. Bring service, fraud, legal, technology, and operations teams together to agree on the approach. Tell representatives what it means in practice, including what they can concede and when to ask for the customer directly.

3. Build the detection infrastructure: A company that can’t identify agent contacts also can’t reliably route them, determine their cost, or forecast their volume. Many AI agents may not be obvious, but there are behavioral clues, including form-fill speed, session velocity, and repetition. Add emerging ways to verify agent identity for those that authenticate themselves. Tag these interactions to track and report their volume.

4. Review and enforce concession policies: Many service policies are affordable only because a small share of eligible customers claims them. AI agents change that calculation. Find the widest gap between who qualifies and who claims. Decide what the company is willing to pay at higher claim rates, then review discretionary benefits and limits accordingly. Check policies are applied consistently across channels.

5. Build the agent-to-agent channel: Provide a machine-to-machine interface with authentication, clearly defined permissions, and records of each action. Route it separately from the human channel. The interface alone isn’t enough. Eligibility and concession rules need to support automatic decisions and execution within set limits. Otherwise, structured requests will still end up in a human queue and won’t deliver the cost advantage.

The cost of waiting

Companies that haven’t decided how to handle consumer AI agents leave those decisions to existing controls and frontline judgment. That can push requests into expensive channels without stopping the demand. Companies should set their approach now, before rising volumes make it harder to change.