AI agents now make consequential decisions inside relationships that carry fiduciary duties. The duty does not move to the machine or the vendor; it stays with the human professional who already owes it. AI made the predictable work fast and automatic, and what is left is taste, judgment, and someone who answers. What is missing is the proof: asked to show an agent acted loyally and with care, there is no record that proves it.
This is one line of work approached from three sides: a working prototype, a doctrinal argument, and a peer-reviewed result. Each column is the same idea further along: first built, then argued in front of legal scholars, then peer-reviewed and presented at ICML. The record is how you prove the someone was there.
Named co-author, MIT FutureTech Delphi study of 272 international AI-risk experts (arXiv 2606.04490, 2026), which measured the gap this work answers: the people most exposed to AI harm are not the ones responsible for preventing it. Findings ↗
The Certificate of Action, running. A tamper-evident record built across an agent pipeline: break any step and the chain fails. Framed first as a product problem at Docusign, then prototyped at the LQ002 hackathon.
The doctrinal case, presented to a room of fiduciary-law scholars at NYU's Fiduciary Duties and AI workshop. Four duties, loyalty, care, disclosure, and confidentiality, mapped onto four verifiable layers.
The technical proof. A trust primitive for verifiable autonomous legal AI workflows that satisfies concrete record-keeping and oversight requirements. Presented at the ICML 2026 AI for Law workshop in Seoul, July 10. The paper publishes in the workshop proceedings later this month.
Proceedings link coming when the paper is live