What Your Customers Really Think About AI Agency

Would you let an AI assistant get your money back? Would you let it spend your money? In Avaya's Customer Experience on Their Terms research, 41% of respondents would pre-authorize AI to initiate an eligible refund without asking again at the moment of action. Just 8% would pre-authorize it to purchase within a spending limit they set.

That difference offers a useful starting point for understanding AI agency. Consumers evaluate the job, its consequences, and how much control they retain. My reading of this research is that businesses can earn greater permission for AI by making each action's scope clear and useful.

What Does an AI Agency Mean in Customer Experience?

AI agency is an AI system's capacity to take actions toward a goal, such as changing an appointment or initiating a return. In customer experience, its practical value depends on the authority the customer and business grant it. Agentic AI describes the systems that can plan and carry out that work.

Two forms matter: a company's AI can act during a service interaction, and a customer's personal AI can deal with the company on their behalf. Both require a clear answer to the same question: what, exactly, has this agent been authorized to do?

Which Actions are Customers Ready to Delegate?

Willingness is strongest for several familiar service tasks with clear outcomes. Asked which actions they would be comfortable authorizing without another approval at execution, respondents selected the following examples:

Source: Avaya, Customer Experience on Their Terms. Selected responses shown; respondents could select all that applied. Base: 558 U.S. adults.

Yet when asked about the most advanced AI role they currently felt comfortable with, only 6% chose taking pre-approved actions, and 2% chose representation within limits. Those questions measure different things: comfort with specific tasks and comfort with a broader role. The findings suggest businesses should introduce delegation through concrete uses customers understand.

Start with a clear offer: authorize AI to process an eligible return under stated conditions. Let a successful experience give the customer a reason to consider delegating another task.

How Much Independence do Customers Want AI to Have?

Customers want different degrees of control, and have very little interest in unrestricted AI decision-making. In the report's control-model question, 33% preferred approving every action, 20% preferred setting rules or limits, and fewer than 1% selected independent AI decision-making. Another 22% would not allow AI to act for them.

The broader desire for control also appears in Pew Research Center's September 2025 findings: 61% of U.S. adults wanted more control over how AI was used in their lives. This is a separate survey of general attitudes, but it reinforces the importance of giving customers understandable choices.

Those choices need enforcement. OWASP's guidance on excessive agency recommends limiting an AI system's permissions and requiring human approval for high-impact actions. It also calls for authorization checks in downstream systems. For businesses, that means a promise to respect customer limits must be backed by controls where the action actually happens.

When Should an AI Agent Ask for Approval Again?

Approval should reflect the consequence of the proposed action. When asked which actions they would most want to approve immediately before execution, 59% selected purchases, 49% recurring charges or price increases, and 44% changes to payment or financial information. Respondents could choose up to three.

An AI assistant might be allowed to find a better service plan but still need fresh approval to enroll the customer in a more expensive one. Businesses should define that boundary before the interaction reaches it.

Google's Agent Payments Protocol announcement illustrates how the industry is addressing this need. It describes signed mandates that capture purchase authorization and, for delegated tasks, conditions such as price limits and timing. The principle is relevant beyond commerce: make the customer's instructions verifiable at execution.

Why Does the Ability to Undo an Action Matter?

Recoverability can influence whether customers will delegate in the first place. In Avaya's research, 37% said they would allow AI to act on its own only when the action could easily be reversed.

A separate question described an incorrect account change that could be fully corrected and caused no financial harm. Correcting it quickly was the leading response for restoring confidence, selected by 37%. That finding applies to the specific recoverable scenario presented.

For CX leaders, recovery belongs in the original service design. Decide how a customer can challenge an action, which changes can be reversed, and who owns the correction. Where reversal is difficult, require stronger approval before execution. Measure how quickly you restore the right outcome.

Will Customers Send Their Own AI Agents to Your Business?

Some are open to it, especially for time-consuming service chores. When asked which tasks they would most likely delegate to a personal AI assistant, 48% selected checking an order or request status, and 39% selected waiting on hold or navigating service systems. Respondents could choose up to two.

Readiness remains mixed. About 40% were comfortable with a personal AI interacting with a company's AI, while 42% were uncomfortable, even when the scenario specified authorized information and pre-approved actions. These are stated attitudes toward a hypothetical experience, rather than evidence of current adoption.

Businesses can prepare by making routine services accessible to authorized representatives. NIST's August 2026 guidance on agent identity emphasizes distinct agent identities and appropriately scoped permissions. Before granting access or allowing a change, a company needs to establish who an agent represents and the limits of that representation.

How is Avaya Infinity Designed to Support These Needs?

Avaya Infinity connects AI, enterprise systems, workflows, and human expertise in an orchestration platform. That foundation helps businesses coordinate the work behind an AI action and maintain continuity across the customer journey.

The research points to four priorities that align with Infinity's design:

• Connect answers to execution. Infinity orchestrates workflows across connected front- and back-office systems, helping a service request progress toward completion. Its Model Context Protocol support provides a standard way to connect AI with enterprise tools and data.
• Apply control where actions occur. Avaya's MCP implementation describes permission controls, an auditable gateway, tool-level controls, and explicit confirmation for high-stakes actions. These capabilities support workflows that reflect customer authorization and enterprise policy.
• Preserve understanding when people step in. Infinity carries interaction context across channels and AI-to-human handoffs. That matters when automation stalls: 64% of respondents wanted automatic connection to a person with their information carried forward after several unsuccessful AI attempts.
• Keep the architecture adaptable. Infinity's model-agnostic approach lets enterprises choose AI models for different uses and evolve the tools they connect as needs change.

Enterprises put this foundation to work by configuring consent, approval thresholds, recovery processes, and integrations for each service. Connections to customers' personal agents also require the appropriate identity and authorization design.

Where Should Businesses Begin?

Begin with a frequent, bounded task whose successful outcome is easy to verify. Define what the customer can authorize, when the AI must pause, and how a person takes over. Test both successful completion and recovery. Track completed resolutions, repeat contacts, and correction time to see whether delegation actually reduces customer effort.

The opportunity is to earn the next permission through a useful result. Customers should be able to give AI more work while remaining confident that the decisions they care about stay under their control.

Explore the full Customer Experience on Their Terms report and see how Avaya Infinity can help connect AI action with the context, workflows, and expertise your customers need.

About the Research

Avaya's Customer Experience on Their Terms study used an online U.S. consumer panel and was fielded in September 2026. All Avaya statistics cited here come from questions answered by 558 adults aged 18 and older. Questions 1-3 had 577 respondents. The findings provide directional insight into stated attitudes. Percentages are reported as published; questions and response options differ, and multiple-response percentages need not total 100%.