In June, Anthropic had to temporarily block access to some of its models for political reasons. Companies that had already built software and workflows on top of them were left stranded. For us it was immediately clear that this had to be read as a business risk. Especially because we were in the middle of developing Nuclos 5.

The media then quickly discuss origin and ownership of models. That is not wrong. For companies, the first question is rather: what happens in purchasing, production, or finance if a model is gone tomorrow?

We need a figure you can know and test. I call it, perhaps a little grandly, “Time to Validated Switch”. How long does it take to move business-critical AI support onto another model and run it again in production with comparable quality, security, and compliance?

With Nuclos 5 we decided: an ERP has to stand and work without AI. Too much in the company hangs on the ERP to justify even a few hours without it. Processes, screens, rules, rights, data. If a model goes offline, a working ERP remains. If a company no longer wants to use AI, or is not allowed to, a working individual ERP remains. Nuclosian is remarkable assistance inside Nuclos’s low-code frame, but it is not the pivot of day-to-day operations.

If you want to attach local models, you should be able to do that without binding the system to outside vendors. MCP is the tool layer for developers. Nuclosian uses the same logic in the client, for process owners and configuration.

Open source and open interfaces help with substitutability. They are not enough if process logic and security rules live only inside the model. Data access, flow, and checks should stay in the system. Then the model can change. The process keeps running.