87% of Swedish managers use AI daily, but *who keeps the record* of what it said?

A new normal

TNG's report on AI in Swedish work life, published in January 2025, paints a picture of adoption that has moved past the pilot stage ^[1]. 87% of managers and specialists use AI at least once a day. 81% ask AI instead of a colleague at least once a week. The tools are Copilot, ChatGPT and Gemini, and they are used for writing, research, analysis and coding.

The numbers suggest AI is no longer something people try on a Friday afternoon. It's in the workflow.

Yet 38% say they worry AI might make decisions feel less human, and the same share worry about AI's impact on decisions ^[1]. Those are not fringe concerns. They are the two sides of the same coin: when AI shapes what people see, recommend or decide, someone will eventually ask what it actually said, and why.

That is when the evidence question comes up.

The customer's screenshot problem

When a chatbot gives a customer an answer and the customer later disputes it, the customer's screenshot is often the only record. Think of it like a receipt you took a photo of in a shop, except the shop kept no copy and now says the receipt must be wrong.

In 2024, a tribunal in British Columbia ordered Air Canada to compensate a customer after its chatbot gave him incorrect information about bereavement fares ^[2]. The customer had a screenshot. The airline offered no evidence to the contrary. The tribunal assumed such evidence wouldn't have helped.

The case is a good example of what happens when one side keeps the record and the other doesn't. The customer had proof of what the chatbot said. The airline had an argument.

The law asks for transparency

From 2 August 2026, Article 50 of the AI Act requires people to be told when they are interacting with an AI system, unless it is obvious from the context ^[3]. The rule is about transparency: people have a right to know they are not talking to a person.

The marking makes it clear an AI is involved. It doesn't settle what the AI said, or whether the organisation is bound by it. That is a separate question, and it tends to come up later, when a customer, an auditor or a supervisor asks to see the record.

When AI is used to assess people, Article 26 requires users of high-risk AI to inform those affected by a decision ^[4]. For decisions with legal or similarly significant effects based on high-risk AI under Annex III of the AI Act, Article 86 gives the person a right to a clear explanation of the AI's role and the main elements of the decision ^[5]. For purely automated decisions under the GDPR, Article 22 limits when they can be made and gives the person the right to human intervention and to contest the decision ^[6].

The rules are about fairness and accountability. They assume the organisation can show what happened. That assumption is harder to meet when the record is a conversation or a recommendation that happened in real time and was never sealed.

What this means for you

If your organisation uses AI in customer service, recruitment, credit decisions, advice or recommendations, ask these questions now:

• When someone questions what the AI said or recommended, who has the record, and in what form?
• If a customer turns up with a screenshot six months later, what do you check it against?
• When you tell someone they were talking to an AI, do you also seal what the AI said?
• If a supervisor or an auditor asks to see a decision and the reasoning behind it, can you show it unchanged from the day it was made?
• Who in your organisation is responsible for keeping AI records, and do they know what the law will expect from 2 August 2026 and 2 December 2027?

The gap between using AI daily and being able to account for what it did is not a technical problem. It is a governance problem, and it is worth closing before someone asks you to prove something you no longer have.

How a digital witness fits here

The thing to seal is each answer or conversation as it happens: the user's question, the AI's response, the model version and the notice that the person was talking to an AI.

Sealing takes a fingerprint of the record, not the record itself, so the conversation stays in your systems. The fingerprint is sent to WitniumChain through the REST API, the SDK or the MCP server, and you get a witness ID back. When a customer, an auditor or a court later asks what the AI said, anyone you share the record with can calculate the fingerprint again and compare it with the sealed one. If it matches, the record hasn't changed. Change a single character and it won't match.

Say a customer claims the chatbot promised a refund the terms don't allow. You pull up the conversation, sealed at the time it happened, and show that the chatbot gave a different answer. The customer can check the seal themselves. The conversation becomes checkable, not just your word against theirs.

A witness doesn't show that the answer was correct, fair or within your authority to give. It shows that the record you are looking at is the one written at the time. That is often enough to settle the question, or at least to move the conversation from "did the AI say that?" to "should it have?".

The first 2,000 witnesses a month are free. Your system sends a fingerprint, gets a witness ID, and stores both with the record. To check, anyone calculates the fingerprint again and compares. The content never leaves your control.

How it would work

1. Seal each answer or conversation when it is sent or closed.
2. Store the witness ID with the record.
3. When questioned, share the record and the witness ID, and let the other side check.

What to watch

2 August 2026: Article 50 of the AI Act applies. People must as a rule be told when they are talking to an AI.

2 December 2027: High-risk requirements under Annex III of the AI Act apply, including logging (Article 12), human oversight (Article 14) and obligations for users (Article 26). The right to explanation under Article 86 also applies from this date.

2 August 2028: High-risk requirements under Annex I apply.

The deadlines are fixed. The question is whether your organisation will be ready to show what its AI did, said or recommended, when someone asks.