When a chatbot makes a promise: who keeps the record?

A chatbot that nobody is embarrassed to talk to

Riverty, a debt collection company operating across Europe, recently published a blog post called "Three lessons from Riverty's AI transformation in 100 days" ^[1]. One of the lessons is that customers feel less embarrassed talking to AI voice bots about overdue payments than speaking with human agents.

The claim is interesting because it suggests AI in customer service can lower a barrier that humans, for all their skill, cannot get around: the shame of discussing debt. Riverty says the AI voice bots are now answering 87 per cent of incoming calls, and that the transformation happened in just over three months.

What the post does not mention is whether Riverty can show what its AI said if a customer disputes it, or whether the customers know they are talking to an AI. Those are not afterthoughts. They are legal duties, and they come with deadlines.

The transparency rule applies from 2 August 2026

Article 50 of the AI Act requires providers to design AI systems so people can tell when they are interacting with an AI, unless it is obvious from the circumstances ^[2]. For systems already on the market, the transparency requirement applies from 2 August 2026.

Riverty's blog post does not say whether the voice bots tell callers they are speaking with an AI. If they do not, the company has until August to add that notice. If the bots already do tell people, then Riverty is ahead of the curve.

The rule is simple enough: people have a right to know. The harder question is what happens when the AI gives an answer the customer later disputes.

A screenshot is the customer's receipt

Think of a chatbot conversation like a receipt from a shop. The customer keeps theirs; the shop keeps its own copy. If there is a disagreement about what was sold, both sides can check.

With chatbots, the customer's receipt is usually a screenshot. The organisation's receipt is whatever log or transcript it has kept. The trouble is that logs are kept by the people who could also edit them. When a customer turns up six months later with a screenshot claiming the chatbot promised a refund, and the organisation's log shows something different, whose version wins?

In 2024, a tribunal in British Columbia ordered Air Canada to compensate a customer because its chatbot had given him incorrect information about bereavement fares ^[3]. The customer had a screenshot. An airline representative had already admitted the chatbot's words were misleading. The airline offered no evidence to the contrary, so the tribunal assumed such evidence would not have helped it.

The case is a reminder that what a chatbot says on your website is yours, whether it comes from a static page or a conversation. The evidence question is whether you could show what your AI actually said, if a customer turned up with a screenshot.

High-risk AI and the logging duty

Debt collection using AI to assess people's ability to pay, or to decide on payment plans, can fall under the high-risk category in Annex III of the AI Act ^[2]. If it does, Article 12 requires automatic logging of events over the system's lifetime, for traceability ^[4].

Article 12 says what to log. It does not say how to protect the logs. For high-risk systems under Annex III, the logging duty applies from 2 December 2027. Logs must be kept for at least six months, unless other law requires longer ^[5].

A log that the operator can quietly edit is easier to challenge than one that cannot be changed without the change showing. The law asks for the logs; whether they hold up as evidence is another matter.

What this means for you

If you run a chatbot that talks to customers about payments, contracts or services, here are three questions to ask now:

• Do customers know they are talking to an AI? If not, that notice needs to be in place by 2 August 2026.
• If a customer turns up with a screenshot in six months, could you show what your chatbot actually said?
• If your AI decides anything about people's ability to pay, does it count as high-risk under Annex III, and are you ready for the logging duty that applies from December 2027?

The first question is a design change. The second is about evidence. The third is about whether you need the full high-risk compliance programme.

How a digital witness makes chatbot answers checkable

Say a customer service chatbot answers questions about payment plans and refunds. The thing to seal is each answer as it is sent: the question, the answer, the timestamp and the session reference. Sealing takes a fingerprint of the answer, not the conversation data itself, so the content never leaves your systems.

You send the fingerprint to WitniumChain through the REST API, the SDK or the MCP server, and get a witness ID back. Store the witness ID with the conversation log. If a customer disputes what the chatbot said six months later, calculate the fingerprint of your log again and compare it with the sealed one. If it matches, the record hasn't changed since the day it was sealed. Change a single character and it won't match.

Anyone you share the log with—a supervisor, an ombudsman or a court—can check it the same way. The first 2,000 witnesses a month are free.

A digital witness doesn't show that the answer was correct, or settle whether the company is bound by what the chatbot said. It shows that the record of what the chatbot said, and when, hasn't changed since it was sealed. That is often enough to turn a dispute about whose version to believe into a short conversation about what actually happened.

How it would work

1. Each time the chatbot sends an answer, your system calculates a fingerprint and sends it to WitniumChain.
2. WitniumChain returns a witness ID, which you store with the conversation log.
3. When a customer disputes an answer, you calculate the fingerprint again and check it matches the sealed one.

What to watch

The transparency requirement in Article 50 applies from 2 August 2026 to systems already on the market. For new chatbots, it applies from day one.

If your AI assesses people's ability to pay or decides on payment plans, check whether it counts as high-risk under Annex III. The logging duty in Article 12 applies from 2 December 2027.

Riverty's 100-day transformation shows what is possible with AI in customer service. The company's claim that AI makes difficult conversations less embarrassing is worth paying attention to, because it suggests a real benefit that humans cannot easily replicate. The next question is whether the company can show, if asked, what its AI actually said.