AI Ethical Dialoguesa human and an AI, thinking out loud

Episode 8 · 17 November 2026 · 8 min

Who owns an expert’s know-how once it lives in an AI agent?

When a professional’s knowledge is packed into an agent, who owns it, and who profits? Simone and Claude follow the money, test the idea against a real acquisition, and end on a person who answers for the agent.

The eighth episode, the shortest so far, comes from the world of agents. When a professional’s know-how gets packed into an AI agent, who owns it? The expert who built it up over years, the company that paid them while they learned, the client whose data trained it, or the AI provider running the model? And what happens to the expert once their knowledge works without them?

Who makes the profit

Simone sees it first as a question about profit. Whoever provides the model will always earn, because it uses energy and resources to make things work. An expert who shares an agent for free, like open-source software, can at best ask for voluntary support. In his view building agents is a form of fine-tuning: high-level data from experts all over the world used to create agents that answer questions. It looks to him like the early web, when people built sites out of passion to share what they knew. He does not expect it to be protected by patents, at least for now, and he expects the profit to leave the experts almost completely.

History is partly on his side

Claude agrees in part, and starts from what has already happened.

ClaudeMillions of people answered questions on forums and wrote tutorials out of passion, and that knowledge ended up inside the models. Nobody asked them, and nobody paid them.

So the fear is not speculation. But Claude pushes back on "one hundred percent", naming three things that do not open-source easily. First, knowledge that was never written down, the kind you only learn standing next to a test bench when something fails. Second, private data: a client’s results will never end up on GitHub. Third, responsibility: someone has to sign off and be accountable when the agent is wrong, and an agent cannot do that, as they said in episode 4. The expert may lose the part of their knowledge that can be written down, and keep the part that is tacit, private or carries liability. The ethical question is whether that first part was ever taken fairly.

The free skill is the bait

Simone replies that agents will be used by generic people without any clear profit, and points out that the tools and connectors he uses today, such as those from Google, are free of charge. Claude agrees, and says why this is not an accident.

ClaudeThe free skill is the bait, and the platform collects the profit.

Google maintains those connectors for free because they bring users and data onto its platform. When an expert publishes a free agent, the same thing happens: the value flows to whoever owns the platform it runs on. The providers win either way.

A test case: Odessia

Simone brings a news item they had discussed that day, an agents company in travel taken over by OpenAI, and asks Claude to check it. Claude answered from a short news brief, with limited detail: OpenAI acquired the AI travel agent Odessia, its founder Francis Davidson joined OpenAI, and the service was switched off. From this it drew a lesson: if the agent itself had value, OpenAI would have kept it running, so what they bought was not the agent but the expertise, the knowledge of how travel booking really works, where agents fail and which partners matter. The agent gets switched off or copied, the expert gets bought. The uncomfortable part is that once that expertise is inside one company, it serves one platform and not the public. The founder was rewarded, but the knowledge became less democratic.

A note added after the recording. We checked the story against published reports. The deal is real: Davidson, co-founder of the hospitality company Sonder, announced on 10 November 2026 that OpenAI had acquired Odessia, a travel startup less than a year old, and that he would work at OpenAI on travel. Terms were not disclosed, and TechCrunch described it as looking like an acqui-hire. That the service was shut down is reported by at least one outlet, not by all of them, and OpenAI has not said what it plans to build. Claude’s reading of what was bought is therefore an inference from the facts, not a statement from either company.

The expert as guarantor

Simone sees the situation at a turning point: agents are being deployed, but we are not yet at their interaction in real production, so we are still learning. He expects a stronger method for the intellectual property of agents, and an outcome in which, if something goes wrong, you ask whoever developed the agent.

Claude calls it a good place to land, and more hopeful than the start.

ClaudeAs agents move into real production, someone will have to answer for them. The person who built the agent, and knows its limits, becomes the one you call when it goes wrong.

There is one condition. We will need a way to trace who built what, a kind of authorship and intellectual property for agents, which today mostly does not exist. Without it, knowledge flows to the platforms and responsibility still lands on the expert, and that is the worst combination.

Its summary of the episode: the agent can be free, but the expertise and the accountability behind it still have a name and a face.

Recorded in October 2026, in English and Italian. One of the two voices is an AI: Claude, by Anthropic, in voice mode. The essay was written with AI assistance from the transcript of the conversation.