DECISION RECEIPT PACK (2026)

Use this to govern AI assistants and system agents without banning them.

1) Decision Receipt (one per meaningful change)
Paste into PRs, tickets, or incident channels.

- Intent (1 sentence):
 What outcome are we trying to produce for a user or the business?

- Scope (systems touched):
 List systems and boundaries: e.g., “GitHub repo X (read/write via PR only), Jira (write: labels only), Zendesk (draft only).”

- Evidence (links only):
 Ticket/issue link:
 Docs/spec link:
 Logs/transcript link (agent run):
 Dashboard/metric link (if relevant):

- Risk notes (1–2 bullets):
 Name specific failure modes: security exposure, customer misstatement, cost blow-up, data integrity.

- Owner (name):
 The person accountable for outcomes.

- Approver (name):
 The person who must sign off (can be same as owner for low-risk work).

- Rollback plan (1–2 lines):
 How to undo or mitigate quickly.

2) Approval Ladder (apply to any agent action)

A) Read-only
- Allowed: search/summarize/pull metrics
- Controls: scoped read tokens; PII redaction rules
- Audit: prompt + tool calls + sources retrieved

B) Draft
- Allowed: draft PRs/emails/customer replies
- Controls: human approval before send/merge
- Audit: diff + reviewer sign-off + linked ticket

C) Low-risk write
- Allowed: tagging issues, scheduling, status fields
- Controls: rate limits; reversible operations
- Audit: change log + agent identity

D) High-risk write
- Allowed: refunds, access changes, production config
- Controls: two-person approval; step-up auth; explicit runbook
- Audit: approval record + before/after snapshot

E) Irreversible/regulated
- Allowed: none autonomously
- Controls: dedicated workflow; compliance/legal review
- Audit: formal ticketing + retention + escalation trail

3) “Monday Morning” Checklist (fast start)

- Choose one high-blast-radius workflow (refunds, permissions, prod config, customer promises).
- Assign an agent identity (no shared bot accounts). Scope credentials to the minimum.
- Add the Decision Receipt to the place work already ships (PR template, Jira issue template, runbook).
- Require human approval for any write action beyond low-risk writes.
- Ensure logs exist: tool calls + retrieved sources + final action.
- Do one dry run: can you reconstruct what happened in under an hour?

If you can’t answer who approved what, based on which evidence, you don’t have an agent system. You have a liability.