Seal

Seal lets an AI payment agent handle changed instructions while keeping execution tied to a person’s exact approval. Finance teams want agents to finish useful work. The difficult moment is when a supplier changes the amount or bank account after the original approval. Understanding that message should produce a new proposal, not new authority. Our demo uses Meta Muse Spark Open through Baseten to read an invoice and a later supplier message. Muse cites the exact source lines, proposes the payee, account, amount and currency, and asks for clarification when needed. Seal shows what changed. A person reviews the exact revised details, then explicitly continues execution. In our accepted hosted run, Muse recognised a EUR36.80 credit note and a replacement account against a EUR250,000 invoice. It proposed EUR249,963.20 and requested review. Approval alone left the ledger unchanged. Explicit continuation recorded one synthetic payment to the replacement account. Reviewers can test the result themselves: verify the signed record, change one cent in a copy and see verification fail, then retry a completed payment and see repeat execution refused. One-use authority is checked and consumed in the same transaction as the simulated payment. The event contribution combines Muse’s source-cited interpretation and clarification with a reviewable proposal, exact-value approval and explicit continuation. It builds on earlier Seal infrastructure and verifier work. All documents and payments are synthetic. No bank is connected. Try the demo: https://agent.aqta.ai/?view=run