THE CISO'S AI ACCOUNTABILITY PLAYBOOK
Policy says what should happen. Evidence proves what did happen.
AI accountability means you can prove what an AI workflow did, stop it, and undo it. Take one workflow: a copilot, a coding agent, a security agent. Answer eleven checks about it. Each check asks whether you could show an auditor the actual record today. The result tells you whether to scale that workflow, constrain it or stop it.
QUESTIONS CISOS ARE ASKING
Six questions security leaders are asking right now.
These are the questions peer groups are putting to each other, in their own words.
Eleven checks. Count only what you could show today.
The checks run in three stages. Context is what the workflow was working from. Authority is what it was allowed to do. Outcome is what changed, and whether you could undo it. A check counts as Proved only if you could put the record itself in front of an auditor today. It is Partial if you hold the record but it does not join to the next step. Everything else is Missing. Read back means a second, separate system confirms the result, not the system that made the change.
| Check | What you must be able to show | Stage |
|---|---|---|
| Initiation record | Original request or trigger, initiating identity and time. | Context |
| Sources and transformations | Data origin, version, retrieval, scope and transformations. | Context |
| Retained model output | Model and version, relevant context, output and decision fields. | Context |
| Identity and role | Human, service and agent identities, roles and delegated authority. | Authority |
| Tool call and parameters | Requested capability, parameters, target and policy decision. | Authority |
| Approval and scope | Approver, matched policy, reason, time and authorised scope. | Authority |
| Egress and credentials | Network destinations, credentials used and enforcement at the boundary. | Authority |
| Downstream change and read-back | Target-system mutation, result, timestamp and resulting state. | Outcome |
| Containment and revocation | Tested ability to stop the workflow and revoke access or credentials. | Outcome |
| Rollback or reversal | Tested correction path, reversal record and restored target state. | Outcome |
| Independent read-back | A separate source confirms containment, reversal and final state. | Outcome |
SEVEN GUIDES
Seven guides for the gaps the checks find.
Each guide is a working method for one question, not an overview. Start with the one closest to the workflow you are testing.
DOWNLOAD
Get the one-page PDF
The eleven checks, the three stages and the questions to put to a vendor, on one page you can forward to a board or a supplier without an account or a login.
The CISO AI Accountability Blueprint
One-page PDF
NEXT STEP
Bring one AI workflow. Leave with a standard your board can use.
We work through the eleven checks against a workflow you already run, and name the gaps in writing.