Chadi Abi Fadel
Peaklight AI
Research
Sovereignty over AI agents
Independent research on how organisations keep control of the AI agents that act on their behalf, grounded in what I run in production at Peaklight.
The question
Organisations are handing real authority to AI agents: answering customers, touching records, triggering actions. Most of those agents run inside someone else's walled garden. The research question is simple to state and hard to answer: how does an organisation remain sovereign over AI agents that act on its behalf, and how would you measure whether it is?
Why "sovereignty" and not "safety"
Most public debate about AI sovereignty is national: where the chips are, where the data centres sit, which government can compel access. That matters, but it is not the question an operating organisation faces. A bank, a hospital or a ministry needs an operational answer: can we see what the agent did, can we stop it, can we change provider, and does the work continue when a provider fails? That is organisational sovereignty, and it is measurable in a way that legal sovereignty is not.
The approach
Rather than trusting a model to behave, the work treats sovereignty as something an organisation can specify, enforce and measure in conventional software that sits outside the language model. Policy lives in code you can inspect. The evaluation framework is unpublished and will appear here once it has been through peer review.
This is independent work in progress. Results will be published here and on ORCID as they are ready. Nothing on this page implies endorsement by any institution.
From research to practice
The research question is also the design brief behind every AI employee that Peaklight runs. If you want to see these ideas applied rather than described, the Peaklight trust page and documentation are the practical counterpart to this page.