AI transparency law: a disclosure label is not transparency for an agent

What transparency law is actually asking for

Transparency has become one of the load-bearing words in AI regulation. The EU AI Act carries explicit transparency duties: people should be told when they are interacting with an AI system, and certain synthetic content should be labelled as such. Beyond that specific article, the broader direction of travel across jurisdictions is consistent. If an automated system affects someone, you should be able to explain what it did and show evidence for that explanation, rather than ask to be taken on trust.

For a passive system this is often satisfiable with a notice. A chatbot can disclose that it is a chatbot. A content tool can label its output. The obligation is real, but it is mostly about telling people the AI is there.

A label is not transparency for something that acts

An agent is a different thing from a chatbot. It does not just produce text for a person to read and judge. It takes actions: it calls tools, moves data, sends messages, changes records, spends money. A disclosure label tells someone an agent was involved. It says nothing about what the agent actually did on their behalf, and that is the part a regulator, an auditor, or a wronged customer will want to see.

The same gap shows up with explainability. A model card describes how a system was built and how it tends to behave in general. It does not tell you what this agent did at 14:32 last Tuesday, which record it changed, or whether a human signed off first. Transparency about a system in the abstract is not transparency about a specific action it took. For an agent, the second one is what counts, and it is the one a label cannot provide.

Three things an agent record has to show

If transparency for an agent means being able to reconstruct what it did, then a usable record has to answer three questions for every action that mattered.

What it did. The specific action, not a summary of intent: the tool that ran, the input it ran on, the result it returned. "The agent helped with billing" is not transparency. "The agent issued a refund of this amount to this account at this time" is.

On whose authority. Which policy allowed the action, and whether a human approved it. An action taken under a standing permission and an action a person signed off on are different facts, and the record has to tell them apart. Authority is the difference between an agent that acted within its scope and one that exceeded it.

With what result, provably. The outcome, recorded so it cannot be quietly edited afterwards. A log you can rewrite proves nothing the moment anyone has reason to doubt it. The record is only evidence if its own integrity can be checked, ideally by someone outside the organisation that produced it.

You cannot be transparent about what you cannot reconstruct

The uncomfortable part is that most agent deployments cannot answer those three questions after the fact. The actions happened inside a model's reasoning and a scatter of service logs that were never designed to be read together, if they were kept at all. When the transparency request arrives, from a regulator, an auditor, or a customer asking what the system did to their account, the honest answer is often that nobody can fully say.

That is the real content of transparency law for anyone running agents. The obligation is not satisfied by a notice that an AI is present. It is satisfied by being able to produce, on demand, a faithful account of what the agent did and the evidence that the account is true. If the system was not built to record that, the transparency duty is unmet no matter how prominent the label.

How Quox approaches it

Quox treats the record as the product rather than a side effect. Every agent action is captured as a witnessed event: the tool and its inputs, the policy that permitted it, whether a human approved it, and the result, written to a tamper-evident trail whose integrity can be checked independently and offline. The point is that the account of what the agent did is not something the operator can quietly revise later, which is exactly the property a transparency request leans on.

A record is evidence of what happened; it does not by itself make the action correct, and we are careful not to confuse the two. What it does is let you answer the transparency question with proof instead of assurance. The regulatory framing for European deployments is set out on the EU AI Act page, the compliance surfaces that build on these records live on the compliance suite, and the mechanism that makes the trail verifiable is covered in cryptographic audit trails for AI agents.