Companies are putting AI agents at the counter. They cancel flights, explain bills, change addresses and process refunds, and the company tells its customers what the agent can do for them.
How agents get tested today
The testing happens on the company's side. Intercom's Fin Simulations lets a team "run fully simulated customer conversations from start to finish" before an agent reaches real customers, and most agent vendors now ship something similar. The certifications in this market, standards like SOC 2 and ISO 42001, describe the vendor's process.
That work tells a company whether the system runs and whether it has a process around it, and it's useful work. It's built from the company's chair, by people who work for the company.
The question the customer asks
The customer has a different question, which is whether they got what they were promised. Retailers have hired mystery shoppers for decades to answer that from where the customer stands, by walking in as a customer and writing down what happened.
An agent at the counter deserves the same check. Someone who stands where the customer stands tries to cancel the flight or dispute the charge, and records what actually happened, for whom, and what the agent did when the request went off script. The company's claim about its agent becomes something a person can test.
The rules are arriving on the customer's side
Regulators have been looking at this from the customer's chair for a few years. In June 2023, the Consumer Financial Protection Bureau's report on chatbots in consumer finance, drawn from complaints people filed, warned that chatbots can fail to recognize when a consumer is invoking federal rights and can trap people in a "doom loop" with no path to a person. California's regulations on automated decisionmaking technology, finalized by the state's privacy agency in July 2025, give consumers a right to opt out when automation replaces human judgment in a significant decision, with compliance starting January 1, 2027.
When both sides are agents
Soon the customer sends an agent of their own. The conversation then happens between two pieces of software, one working for the company and one working for the person, and the person who asked holds the only position that can see both: what their agent was told to get, what the company's agent said it did, and what showed up afterward.
That record is where a vigilant customer does the checking.
Sources
- Intercom, Fin Simulations
- Consumer Financial Protection Bureau, Chatbots in consumer finance, June 2023
- Hunton Andrews Kurth, CPPA Finalizes CCPA Regulations on Automated Decision-Making Technology, July 2025