The Scorecard Built for Humans
In the third and final post in this series about customers’ AI agents and the asymmetry of AI disclosures, let’s talk about what happens when a customer’s AI misbehaves or abuses the human agent: who can they sue?
I want to be upfront. I am not a lawyer. I have worked in a call center and have designed multiple applications that support the ecosystem. So I am looking at this from the human and design side, and leaving the doctrine to people who know it. This is purely a thought piece.
Four parties and no obvious defendant
Let’s start with who is involved when a customer’s AI agent interacts with a human agent.
There is the customer, who instructed their AI to achieve an outcome. There is the AI agent, built on a model by a provider to follow instructions. There is the human agent who is handling the escalation path. And then, there is the employer of the human agent. Sometimes the employer is a vendor to a company whose customers the human agent is handling.
The human agent has no binding relationship with anyone else except their employer. The agreements that allocate risk in agentic interactions are between model providers and their users, and the recurring point in all of that writing is that only parties to an agreement can invoke it. Someone who is harmed but is not party to those agreements generally cannot reach the protections inside them.
Protecting Employees
Employers already have a duty to protect employees from harassment by anyone who is not employed there, where it is based on a protected characteristic. Under the EEOC standard, an employer can be held liable for how a customer treats an employee. Two things have to be true: the employer knew about it or should have known, and did not act to correct the situation.
The law is not written with AI agents in mind. It was written for the customer who screams at someone for twenty minutes, and it exists because we decided that a person should not have to absorb that as a condition of employment.
I want to be careful about how far this stretches, because the version of this argument that overreaches is easy to dismiss. As I understand it, harassment in the legal sense generally requires a protected-class basis. An AI agent that is simply relentless is not harassing anyone in that sense. It may even be programmed to be polite while it does it. None of that makes the call any easier for the person taking it.
An employer sets the conditions of the job, which means an employer owns them.
Methods and Procedures
This is a design problem.
The human agents are measured on handle time, resolution rate, customer satisfaction, and adherence to policies. Every one of those assumes the person on the other end eventually gets tired, or satisfied, or gives up. That assumption has held for as long as contact centers have existed. An AI agent does none of it. A human agent who draws a queue full of these conversations will watch their numbers drop, and nothing in the measurement will show why.
I was once having a rough day on the floor. Almost every call that day was harsh, and then one got disconnected while the customer was mid-rage. My supervisor pulled me aside to ask if I had ‘hung up’ on him. I had not. I got dinged for ‘potential disconnect’ anyway.
The policies dictate whether a human agent is allowed to end a conversation. Most contact centers have a disconnect policy for abusive callers, and without an explicit authority, people stay on the line for fear of it being scored against them. Those policies were written for humans. A human disconnects eventually. An AI agent does not.
So what should employers do?
There are a few options:
To start, have training and policies around identifying agentic interactions and what a human agent can do.
Give explicit authority to end the conversation when the AI agent is borderline abusive, and don’t score the human agent poorly for that interaction.
Determine how agentic conversations will be scored against a human to avoid penalizing people for conditions they cannot control.
Continue conversations with human agents as the tech landscape evolves.
And, I can’t emphasize the last point enough. Asking the people doing the work what they are seeing is the single most important way to understand the agentic behavior against human agents.
The uncomfortable version of this is that the employer has the duty and also owns the metrics creating the pressure. The duty and the scorecard were both built for humans, and need to be revisited for AI agents.
If you run one of these organizations: do your human agents know what they are allowed to do when it isn’t a person on the other end?