AI agents are already calling insurers, and nobody knows what they said

The calls have started

AI agents are already calling insurance companies. Dave Buckley, chief information officer at Guardian Life, told Insurance Business Magazine that healthcare providers in the US are using AI agents to phone Guardian's call centre to verify patient benefits ^[1]. The agents sound like people, handle the conversation and get the information they need.

For now, the callers are telling Guardian they are AI. That won't always be true. As these agents become cheaper and easier to deploy, some will call without saying what they are, and some won't even know they should.

Under Article 50 of the AI Act, people must as a rule be told when they are interacting with an AI system, unless it is obvious from the circumstances ^[2]. That applies from 2 August 2026. The rule is about transparency, not technology, but it sets up a sharper question: when an AI calls and the interaction is questioned later, how do you prove what was said?

The evidence gap

Here is the problem in everyday terms. Say you phone your bank to report a lost card. You tell them the time you noticed it missing, and they say they will block it. A week later, a fraudulent transaction goes through and the bank says you called later than you remember. You have your notes; they have their logs. Whose record holds up?

When a human calls a human, proving the conversation happened is usually enough. When an AI agent calls, the content matters more. The agent may be acting on instructions it received earlier, from a person or another AI, and those instructions may not match what the agent ends up saying. If the insurer's system logs one thing and the healthcare provider's agent logs another, the question is which log you can trust.

Guardian Life has noticed this already. Buckley told Insurance Business that the company is working on transparency standards and ways to verify that callers are who they say they are ^[1]. That is the first step. The second is making sure both sides can show what actually happened during the call.

What Article 50 requires

Article 50 of the AI Act says that providers of AI systems intended to interact with people must design them so that people are informed they are interacting with an AI, unless it is obvious ^[2]. The rule applies to emotion recognition and biometric categorisation systems as well.

For deployers, the duty is slightly different: they must inform people that they are interacting with an AI. That distinction matters, because the deployer is often the one making the call.

The rule does not say how the information must be given. A voice message at the start of the call would do it. So would a prior notice that the healthcare provider uses AI agents to handle insurance queries. What matters is that the person on the other end knows.

The penalty for not complying with Article 50 is up to €15 million or 3 per cent of global annual turnover, whichever is higher ^[2]. For small and medium-sized enterprises, the lower amount applies. That is steep enough to make transparency worth getting right.

Sealing call records as they happen

This is where a digital witness fits. The thing to seal is each call record as it is written: the caller identifier, the questions asked, the information given and the timestamp. If the AI agent keeps a log of what it said, seal that too. The content stays with you; only a fingerprint is sealed.

Anyone you share the record with can later check that it is exactly what was written at the time. If the healthcare provider and the insurer both seal their records, and they do not match, that tells you something happened between the agent's instruction and the insurer's system. If they do match, the dispute moves to a different question.

Sealing happens through WitniumChain, using the REST API, the SDK or the MCP server. Your system sends a fingerprint, not the call content, and gets a witness ID back. To check it later, you calculate the fingerprint again and compare. The first 2,000 witnesses a month are free. The documentation shows you how to build it into a call-handling system.

What a witness does not do is prove that the information given was correct, that the agent was authorised to make the call, or that the person who received the call consented to speaking with an AI. It shows that the record has not changed since it was sealed, which is enough to turn "they said, we said" into a shorter conversation.

What this means for you

If you are deploying AI agents that interact with people, three questions are worth asking now:

1. How do you tell the other party they are speaking with an AI, and can you show you told them?
2. What records do your agents keep, and who can change them?
3. If a conversation is disputed six months later, whose version will a supervisor or a court believe?

If you are receiving calls from AI agents, the questions are similar:

1. How do you know a caller is an AI, and do your systems log that?
2. Can you show what information your system received, unchanged, if the caller disputes it?
3. If an agent gives you information that turns out to be wrong, can you prove what it said?

For insurers, this is particularly sharp. A call to verify benefits may lead to a coverage decision, and that decision can be questioned years later. The call record becomes evidence, and evidence that one side can quietly edit does not hold up.

What to watch

Article 50 of the AI Act applies from 2 August 2026 ^[2]. By then, anyone deploying AI that interacts with people needs a plan for how to inform them, and a way to show they did.

The European Commission has said it will issue guidance on what "obvious from the circumstances" means in practice, but that guidance is not out yet. Until it is, the safer assumption is that if there is any doubt, you tell them.

For call centres and customer service teams, the question is not whether AI agents will become common. They already are. The question is whether, when something goes wrong, you can show what happened without having to trust whoever kept the log.