Why an AI Agent Needs Customer Evidence More Than a Person

In February 2024 a British Columbia tribunal ruled against Air Canada after its website chatbot told a passenger he could apply for a bereavement fare after he had flown. The airline's policy did not allow that. The tribunal found the airline had not taken reasonable care to make sure its chatbot was accurate.

In April 2025 Cursor's support bot told a user who kept getting logged out that Cursor was designed to work with one device per subscription. Cursor's co-founder said on Reddit that there was no such policy. The company said the logouts came from a change to session security.

In both cases a bot stated, with confidence, a policy the company did not have.

What a person does first

A support person who isn't sure of an answer has options. They can open the account and read the last few tickets. They can ask a colleague who handled it before. They can say they will find out and come back.

Little of that is written into a process. It sits in the person, and it happens before the reply goes out.

An AI agent has the conversation and whatever it was given: a knowledge base, some account fields, maybe a few recent messages. If the answer isn't in what it was given, the reply doesn't say so. A wrong answer reads the same as a right one, and the customer has no way to tell them apart.

Why scale changes the cost

A person's wrong answer lives in one conversation. An agent's wrong answer to a common question can repeat in every conversation where that question comes up, until someone reads one of them and notices.

That is a reason to care about what the agent reads before it answers. It is also a reason to read what it said afterward, because the second check is the only one left when the first one fails.

What the agent needs with each answer

Three things, roughly in the order a person would reach for them:

• The record the answer rests on. The policy, the plan terms, the date. If the agent states a policy, the text of that policy should be what it read.
• The account's recent history, in date order. Tickets, calls and commitments from every tool the account touches, each with its source and date. A commitment made on a call isn't visible to an agent that only reads the help desk.
• A way to say it doesn't know. An agent that can see the record is missing, or that two records disagree, needs somewhere to go other than a confident guess. A handoff with the context attached is one such place.

The same evidence helps the person reading afterward. A reviewer checking an agent's answer needs the record the answer should have rested on, and finding it is the slow part of the review. Should Your AI Agent Use the Same QA Scorecard as Your Team? covers the criteria that catch the failures above: a statement that isn't true, an answer that should have been a handoff, a handoff that lost the context, and a loop.

A check you can run this week

Pull ten recent conversations your AI agent handled without a person. For each reply that states a policy, a price, a date or a fact about the account, find the record it should rest on, and sort the reply into one of three groups:

1. The record exists and matches.
2. The record exists and says something different.
3. There was nothing to check it against.

Read the third group first. A reply that stated something no record backs up is the one that behaves like the two cases above, and it will not look wrong on the screen.

Ten conversations can find problems. They will not tell you how often the agent has them, so treat the result as a list of things to fix.

What this doesn't settle

Two public cases are not a rate. Both involved consumer-facing chatbots built differently from a B2B support agent, and neither tells you how often an agent in your help desk states something untrue.

Evidence also doesn't make an agent right. A model can misread a correct record, and the record can be stale: when a product change isn't in the documentation yet, an agent can read the documentation correctly and still give a wrong answer. Giving an agent the account's history improves what it has to work with. It still needs someone reading its answers.

Where Resonant IQ fits

Resonant IQ's MCP server gives an AI agent read-only access to the same customer evidence its own product uses: conversations and activity from the tools you connect, with each answer tied to its source conversation and timestamp. Each answer also returns receipts that tie its claims to the rows that support them, so an agent, or a person reviewing it, can check a claim against its source. MCP and permissions covers why that matters.

Sources

• CTV News, "Air Canada's chatbot gave a B.C. man the wrong information. Now, the airline has to pay for the mistake", February 2024. Reports the decision of the British Columbia Civil Resolution Tribunal in Moffatt v. Air Canada, 14 February 2024.
• CX Today, "A Contact Center Chatbot Invents Company Policies, Now Customers Want Out", April 2025. Cursor's support bot and the co-founder's reply.