Skip to content
· 5 min de lectura

What an agent authorization record should contain

A payment receipt tells you money moved. An AI agent transaction receipt has to tell you something much harder: whether it should have. Those are different products and only one of them is what a card terminal prints.

The question that decides the dispute


A payments practice published a five part series between April and May 2026 on what happens when an AI purchase goes wrong. Two sentences from it shaped how I build this. The first: operational logs are not sufficient, evidence quality records are. The second: the party with the better record will not merely win, it will control how the dispute is framed, which category it lands in, and whether the fight reaches a court at all.


That is the whole commercial argument for a record, and it is stronger than anything I could invent. If the quality of the record decides the framing, then the record is not documentation. It is the product's position in a room it will be dragged into later.

The four questions, and the fields that answer them


They will ask: who acted, what were they allowed to do, what did they actually do, and what happened next. Every field below exists to answer one of those without going back to a log.


  • The principal. The person or company the money belongs to, identified by something more durable than an account. `who`
  • The agent. The identity that acted, and the deployment it belongs to. Not a session token. `what`
  • The mandate. Its version, its scope, its expiry, and a reference to the instruction the user wrote when they granted it. `under which authority`
  • The proposed action. What was asked for, normalised before evaluation, so the record shows the request that was judged and not a summary written afterwards. `what was requested`
  • The decision. Allow, deny or escalate, with the rule that produced it, named by version. `what was decided`
  • The evaluation. The conditions that held at the moment of the decision: amount after the running total, count inside the window, price unchanged from the quote, clock. `what was true when it decided`
  • The chain. The hash of this record and the hash before it. `what came before`
  • The consequence. The charge, the refusal, the confirmation that followed, and the time it took. `what happened next`

Notice that four of those eight are about state at the instant of the decision, not about the outcome. The most common mistake in this area is recording that a purchase was allowed. The useful record says what the boundary was, what the running total was, and which rule turned those two into an answer.


One more structural point, which is the difference between an audit trail and a receipt archive. A receipt archive is written for one reader at one moment. An audit trail is written for a reader who arrives later with a different question, and often with adversarial intent. That means the same row has to serve a user who wants a friendly sentence on a phone and a lawyer who wants the raw decision, the rule, and the state. I build the friendly sentence from the row rather than storing it separately, so the two can never disagree.

Retention is not a detail


Card networks already set the floor. Visa's Acceptance Risk Standards require acquirers to investigate and retain dispute investigation details, including the event description and analysis, for a minimum of two years. Two years is the minimum for a human initiated transaction in a closed loop. Agent initiated transactions generate disputes about what an agent was told, so the window in which that question can be asked has to be at least as long, and the record has to survive export.


My ledger is append only for exactly this reason. Rows are hashed into each other at write time. Nothing rewrites history; corrections are new rows that reference the row they correct. Deletion is not available as an operation on a decision, and when retention requires removal, it removes the payload and leaves the chain intact, so a reader can see that something was removed and when.


It is worth being concrete about who reads this, because the four readers want different things and the format has to serve all of them. The user wants to know what the agent bought and why, in one sentence, on a phone. The support team wants the decision, the rule, and the state at the time. The auditor wants the chain and the ability to recompute. The regulator wants all three plus the identity of the principal, in a form that does not depend on our database being online.


The second structural point is that there are two events and one row. An authorization decision and the charge that follows it are not the same fact. They happen at different times, they can diverge, and when they do the divergence is the interesting part. So the row records the decision, and the consequence is either in that row or in a row that points at it, and a reconciler compares the two series. A record that only logs successful charges cannot answer the question a dispute actually asks, which is what was attempted and refused.

Denials are records too


This is the design choice I would defend in a code review. A denied authorization leaves the same row as an allowed one, with the same fields, because the denial is often the more interesting document three months later. It shows what the agent was trying to do, what the boundary was at the time, and how close it came. Systems that discard denials are discarding the only free early warning they will ever get about a bad instruction.


One vendor's documentation puts the intent neatly: the audit trail exists before the dispute does. That is a product claim about their own system, but the sentence is the right one, and it is the standard I measure myself against.


I am building this at Noesia. The property that makes it recomputable is in deterministic authorization for AI agents, the reason it has to be tamper evident rather than tamper proof is in why agent payments need an append-only audit trail, and the commercial consequence is in who is responsible when an AI agent buys the wrong thing. If you have been in a dispute where the logs were not enough, I would like to hear what was missing. What an external audit found in mine is what a security audit taught me.

Building Know Computer, a personal context layer for the AI era. Reply via [email protected].

Más

Notas