AI agents are beginning to take on everyday tasks — booking travel, shopping for products, managing our online accounts. But who will these agents work for?
With major platforms racing to build general-purpose agents, now is the time to demonstrate that trusted, user-aligned alternatives are both possible and feasible. Consumer Reports (CR) and Stanford’s Digital Economy Lab are joining forces to explore a new kind of AI agent: a consumer-authorized agent that acts on behalf of you, serving your interests first and foremost. We imagine a future where agents are loyal by design, helping people navigate the market, understand their options, and make informed choices.
To start building toward this vision, we’ve defined a set of foundational use cases a consumer agent could soon support. These tasks follow a complete customer journey: researching a product, making a purchase, submitting feedback, checking account details, and initiating a return.
You can read the full set of use cases here. Below, we walk through the high-level journey and our plan to bring it to life.
Customer Journey
Imagine the journey of a consumer interacting with an AI agent to aid with a common household purchase task:
“My old dishwasher finally gave out, and I wanted a replacement fast. I asked my AI agent to help me find top-rated models based on Consumer Reports’ testing, something energy efficient that fit the dimensions of my kitchen. After reviewing a shortlist and comparing prices, I chose a highly rated model that could be delivered within the week: the Dishwhiz 9000. I asked my agent to order the dishwasher from a major retailer and schedule the delivery for this weekend.
After the installation, I noticed a musty smell and inconsistent drying. I asked my AI agent to send feedback to the manufacturer describing the issue. They responded with a cleaning procedure and suggested checking the rinse aid level.
Later, I wanted to double-check the return policy and whether I was still eligible to return this model if the problem continued. My agent retrieved the return policy from the manufacturer and confirmed that the model was still eligible for return based on my purchase date.
After a few more weeks of no improvement, I decided I’d rather return this dishwasher and go with a different model. My AI agent initiated the return process with the retailer, confirmed I was eligible for a full refund, and even helped schedule the pickup. I saw the refund appear in my bank account, and the pickup happened on time.
Then I asked my agent to help me choose a better-rated alternative from the original CR shortlist, clarifying that I’d be willing to spend a bit more for higher reliability. I also made sure my CR agent knew about the issues I had with this dishwasher, so that other members could take my experience into account.”
Implementation
We divided this customer journey into discrete use cases:
- Researching a purchase
- Making a purchase
- Submitting feedback
- Inquiring about an account
- Executing a return
We plan to develop against these use cases in a modular, technically progressive fashion, beginning with the least complex (unauthenticated) agent tasks, and gradually layering in authentication, permissioned data access, and transactional flows.
Our development goal is to build reusable primitives that can support many types of interactions over time. By starting simple and building outward, we will lay a foundation for more complex agentic interactions.
Development will focus on “happy paths,” the ideal version of each task working as intended. We also plan to deliver an evaluation suite to confirm that the system performs as expected. In addition, we will note “unhappy paths” as we identify them, so that they can be considered and developed against in the future.
Our foundational use cases are transactional, meaning they focus on task completion and do not accommodate negotiation. We would like to explore agent-driven negotiation once transactional flows have been established. The use cases also presume a single consumer agent and do not accommodate multi-agent interactions – though we consider multi-agent to be another area ripe for future investigation.
Development Roadmap
| Use case | Complexity | Notes |
|---|---|---|
| Use case: Researching a Purchase | Complexity: Low | Notes: Read-only, unauthenticated MVP that does not involve a counterparty when managed by CR |
| Use case: Submitting Feedback | Complexity: Low/Medium | Notes: Atomic, one-way write interaction that can be unauthenticated or authenticated. Introduces customer lookup logic but avoids the complexity of reading account data or managing transactions. |
| Use case: Inquiring About an Account | Complexity: Medium | Notes: Requires the agent to securely access customer data—which requires scope-based permissions, data freshness concerns. |
| Use case: Making a purchase | Complexity: Medium/High | Notes: Introduces financial risk and needs robust confirmation. Involves payment gateways, cart systems, identity verification. |
| Use case: Executing a Return | Complexity: High | Notes: Involves conditional logic (return window, refund eligibility), coordination with both retailers and logistics systems, and reconciliation of payment and fulfillment records. |
If you’re building agents, funding ecosystem infrastructure, or developing trustworthy AI, we’re eager to hear from you. The Loyal Agents initiative is a joint effort between the Stanford Digital Economy Lab and the Consumer Reports Innovation Lab, and we’re actively looking for contributors across industry, academia, and civil society. Reach out at innovationlab@cr.consumer.org.
