When signing off on AI output isn't human oversight

The summit that asked the right question

At a recent AI summit in Bavaria, digital minister Dr. Fabian Mehring raised a point that cuts to the heart of how organisations will use AI in practice: if hundreds of AI agents prepare documents, decisions or analyses for you, and you sign them off without the time or the means to check them, are you supervising anything at all? ^[1]

The question matters because the AI Act will soon require organisations using high-risk AI systems to have human oversight in place. That means a natural person who can intervene in, override or stop the system. ^[2] The law doesn't say how much time that person needs, or what tools they should have, but it does say the oversight must be effective.

Think of it like the co-pilot's seat on a plane. The person sitting there can take over, but only if the instruments show what the autopilot is doing and the person knows how to read them. A co-pilot who just watches the autopilot without being able to check or intervene isn't supervising the flight.

What the AI Act says about oversight

Article 14 of the AI Act requires high-risk AI systems to be designed so that natural persons can oversee them while they operate. ^[2] The oversight must enable the person to understand what the system is doing, interpret its outputs correctly and decide when to intervene or stop it.

For systems listed in Annex III of the regulation, these requirements apply from 2 December 2027. ^[3] That includes AI used in recruitment, credit scoring for individuals, pricing life and health insurance, managing employees and similar decisions about people.

The person overseeing the system must have the authority, the competence and the means to intervene. ^[2] If they can't realistically check what the AI has done before signing off, the oversight is procedural rather than real.

When approval becomes rubber-stamping

The scenario raised at the Bavarian summit is not far-fetched. AI agents can already draft emails, summarise documents, prepare reports and propose decisions at a pace no human team could match. When the output lands on someone's desk for approval, that person faces a choice: spend hours checking the work, or trust that the AI got it right.

If the volume is high and the time is short, approval turns into rubber-stamping. The person becomes a liability shield rather than a supervisor.

The AI Act doesn't ban that arrangement outright, but it does require the person to be able to intervene, and to understand the system's outputs and limitations well enough to know when to do so. ^[2] That only works if the person has access to the information the AI relied on, the reasoning it followed and any flags the system raised along the way.

Without that, signing off on AI-prepared work is like signing a document you haven't read.

What meaningful oversight looks like

Meaningful oversight starts with design. The system must be built to give the supervisor the information they need, in a form they can use, at the moment they need it. ^[2] That might mean:

• showing which data the AI used and which it didn't,
• flagging cases where the AI's confidence was low or the input was unusual,
• recording each decision the AI made and each action it took,
• making it possible to trace how the AI reached a particular output.

The supervisor also needs enough time and authority to act on what the system shows them. If every decision must be approved within minutes and the supervisor can't realistically check the work, the oversight duty isn't being met.

For organisations, the practical question is this: if a decision made by your AI is challenged six months later, could the person who approved it explain how they checked it at the time?

How a digital witness fits here

When AI agents prepare work at volume, the organisation needs a way to show what each agent did, when it did it and on whose instruction. That's where a digital witness comes in.

Here, the thing to seal is each agent action: the instruction it received, the tool it called, the data it used and the result it produced. Sealing takes a fingerprint of the record, not the data itself, so the content stays in your systems. The witness adds a permanent marker that anyone you share the record with can later check.

If a supervisor signs off on an AI-prepared decision, and someone later questions it, the organisation can show what the AI actually did that day and what information was available to the supervisor at the time. The witness doesn't show that the decision was right, but it does show that the record hasn't been quietly rewritten since.

That matters when the question is whether the person approving the work had enough to go on.

How it would work

1. Each time an agent acts, your system sends a fingerprint of the action record to WitniumChain through the MCP server, the SDK or the REST API.
2. The witness seals the fingerprint and returns a witness ID.
3. To check, anyone you share the record with calculates the fingerprint again and compares it with the sealed one.

The content stays with you. The first 2,000 witnesses a month are free. Read the documentation.

What this means for you

If your organisation uses or plans to use AI agents that prepare work for human approval, ask these questions now:

• Can the person approving the work realistically check it, given the volume and the time available?
• What information does the supervisor see: the AI's output, the data it used, the reasoning it followed, or just the final result?
• If a decision is challenged later, could you show what the AI did and what the supervisor knew when they signed off?
• Do you have a record of each agent action that nobody, including you, can quietly edit afterwards?

The high-risk requirements apply from 2 December 2027 for systems in Annex III. ^[3] That leaves organisations time to design oversight that works, rather than oversight that looks good on paper.

What to watch

2 December 2027: High-risk requirements, including the human oversight duty in Article 14, apply to systems listed in Annex III of the AI Act. ^[3]