For advisory and assurance firms
Your clients' AI agents will need sign-off. Give your teams something they can test.
Every client rolling out agents will ask you two things: help us do this safely, and tell us we can rely on what the agents produce. Today there is little to test beyond a log console, interviews and a sample. Rootz Receipts gives every piece of agent work evidence your teams can test independently, without access to the client's systems.
From a sample to every item
Test every agent output, not a sample.
Teams test a sample because checking everything by hand costs too much. Evidence that verifies in milliseconds changes the economics: your tools can check every agent output in the population and bring your people only the exceptions.
Controls tested through the client's console
- Evidence
- Logs and screenshots the client's engineers produce on request.
- Coverage
- A sample, chosen because full testing is too expensive.
- Independence
- You rely on the system you are testing to report on itself.
Evidence attached to each result
- Evidence
- Signed when the work happened: who authorized it, what was asked, what was used, the controls in force.
- Coverage
- Every item, checked automatically; people review the exceptions.
- Independence
- Checked with a free verifier, offline, without the client's systems or ours.
What your teams test
Five assertions, each with its own result.
Every result carries the same five lines. Weak evidence is shown as weak: a line that was only declared, or could not be evaluated, says so and says why. There is no single pass mark to hide behind.
| Line | The assertion your team tests |
|---|---|
| Authorized by | The rules the agent ran under were approved by an officer of the company who held that authority on the day it was given. |
| Specification | This is exactly what the agent was asked, and it was allowed under the controls in force. |
| Materials | These are the sources and earlier results the agent used before it answered. |
| Process | The controls in force were the ones approved; it states what the platform measured and what it only declared. |
| Integrity | The result is in the signed record of the session, complete and unaltered. |
Practice opportunities
Three kinds of engagement it opens.
Design the authority model
Who in the client may approve an agent's rules, the boundaries each class of agent must stay within, and which changes need a second officer.
Roll it out with the client's agents
Two extensions on the client's OpenShell, with no changes to their agents. It fits into the rollout you are already running.
Test controls and outputs
Test against evidence signed at the time of the work. Designed to map onto the frameworks your teams already use, such as ISO/IEC 42001 and ALCOA+ data integrity.
Shape the assurance method with us.
During the pilot program we are working with a small number of firms on what evidence your teams need, in what form, and how your tools should read it. The verifier is free and will stay free, so your teams never need our permission to check.