Change Control · AI Change-Control System

Nothing reaches users un-reviewed

A new model, a tweaked prompt, a fresh integration every change is diffed, routed to the right approvers, and logged before it goes live.

change-control · review #4471
billing-agent · prompt.diff
-  refund up to $500 without approval
+  refund up to $50; escalate above

policy check:  finance-strict → requires risk + finance
reviewers:     risk ✓   finance ✓   eng pending
status:        awaiting 1 approval not yet live

Audit-ready for teams at

BroadfieldLumen HealthNordvikHalcyonVeratech

Treat every AI change like a production deploy

Diff everything

Prompts, model routes, tools, and permissions are versioned. See exactly what changed, line by line.

Route for approval

Policy decides who signs off. Risk, finance, or eng are pulled in automatically by the rules that apply.

Log for audit

Every approval, block, and rollback is signed and streamed to your SIEM. Evidence without the screenshots.

Instant rollback

Undo a bad change in one click

Every deployed version is retained. If a prompt regresses or a model misbehaves, roll back to the last approved state immediately and the rollback is logged like any other change.

Versioning & rollback
version history
v12 · prompt in review
v11 · model gpt-4o live
v10 · tool added approved
v9 · prompt rolled back ↩

Frameworks built in

Map controls to the standards you report on

Change-control records map cleanly to SOC 2, ISO 27001, and the EU AI Act. Export a control narrative for any agent and any window your auditors get the evidence in their format.

See report templates
compliance export
SOC 2controls mapped
ISO 27001controls mapped
EU AI Actrisk records
0gaps open

If it can change behavior, it counts as a change

The riskiest deploy in most AI programs is the tiny prompt tweak that ships at 6pm. Change Control kills that failure mode by treating every surface below as a reviewable unit no exceptions list.

Prompt text

Any character, system prompts included. A one-word edit can move a spend limit or drop a guardrail, so every character diff gets a reviewer.

Model route

Provider, model version, temperature and friends. Swapping to a cheaper route changes behavior; it goes through review like anything else.

Tools & integrations

Adding a tool is adding a capability, and it is reviewed like one with the scopes it requests spelled out in the diff.

Permissions & scopes

Data boundaries, spend limits, access grants. Widening any of them routes to the owners of whatever sits behind them.

The policy itself

Policy edits are the highest-risk change in the system, so they route to the strictest reviewer set by default. Nobody quietly rewrites the rules.

Environments & progressive rollout

Nothing promotes itself

Changes move draft to staging to canary to live, and every promotion is itself a policy-gated change with its own approvers. Canary runs the new version for a bounded slice of traffic and halts automatically on regression signals live means it earned it.

Environments & rollout
rollout · billing-agent v12
v12 · draft compiled
v12 · staging checks green
v12 · canary · 5% of traffic watching signals
v12 · live gated on canary

Evidence on demand

Auditors ask, you export

Control narratives are generated per agent and per window; every record carries the diff, the policy version, the approvers, and timestamps. Retention is configurable to your regulatory window, and exports ship as PDF for auditors or JSONL for your GRC tooling. The framework mapping above is the summary this is the record-level detail underneath it.

Narrative templates  How the log stays tamper-evident
$ pactflow export --agent billing-agent \
    --window 2026-Q2 --format control-narrative

collecting   214 changes · 214 approvals · 3 rollbacks
mapping      SOC 2 CC8.1 · ISO 27001 A.8.32 · EU AI Act art. 12
✓ billing-agent_2026-Q2.pdf written (jsonl alongside)

Move fast and keep the receipts.