ACC Agent Gate Campaign v8 · aurora-mkt-prod · Black Friday 2026 Synthetic demo dataGate active
AI agent governance for Adobe Campaign

The agent asks. The policy decides. A human signs.

An AI agent got a Black Friday brief for 8 EU markets and made 28 calls to Campaign. The gate ran 13, explained 1 without running it, sent 5 to a human, and blocked 9 before they reached production.

Recipients sent 0Held by QA 0

How a tool call travels

Adobe’s CX Enterprise agentic tools list Campaign Classic capabilities that include query execution, workflow control and SOAP/JS execution [1]. The gate is an MCP server the agent talks to instead. Approvals are not a tool the agent can call.

AI agent Claude · Cursor · any MCP client tool call ACC Agent Gate policy/default.yaml · fail closed SQL / JS analyzers 9-check pre-send QA PII masking · row caps · write limits allow mask dry human block only if allowed Adobe Campaign Classic v8 instance (mock in this demo) Human approver role-checked, 1 or 2 audit.jsonl · sha-256 chain
Policy lab

Loosen a rule and watch the same session play out

Every change re-runs all 28 agent calls through the gate in your browser. Simulated approvers sign whatever reaches them, except the production reminder they reject for a business reason. That is the point: humans alone miss what policy catches.

Policy

Data & queries
Sends
Pre-send QA
Workflows & scripts
Other server JS
Session
Sent 0Defective sends 0Held 0

Playground

Submit your own call to the gate

Write SQL or server-side JavaScript, request a send or a workflow action. The gate uses whatever policy is set in the lab, and every decision is hashed into a live audit chain.

Pick a preset or write your own, then submit.

Approval inbox

Requests waiting for a person. Approve as different roles to see role checks, dual sign-off and the re-check at execution time.

Live audit chain

SHA-256 of each entry includes the previous hash, computed in your browser.

Pre-send QA board

Four of eight markets would have shipped a defect

Checks follow Adobe’s own pre-send guidance [3][4][5][7][8]. Select any cell for the detail. Apply a fix and the lab and playground re-run against the corrected delivery.

Audit log

Forge any entry and the chain gives it away

The recorded session wrote 34 entries, blocked calls included. Select an entry to rewrite its decision, as someone hiding a mistake might. Every entry after it stops verifying.

verify(audit.jsonl)
Chain intact

Policy file

Rules live in a reviewed file, not in a prompt

This file follows the lab. Lines you changed are highlighted.


    
  • 01
    Fail closedA tool not in the policy is blocked. New capabilities are opt-in.
  • 02
    Anything that looks like a write is a writePattern analyzers over-block on purpose. A false positive costs one click; a false negative costs a market.
  • 03
    Re-check at execution timeIf a proof is withdrawn while a send waits for sign-off, the approved send is still stopped.
  • 04
    Separation of dutiesRole-checked approvers, two people above a volume threshold, nobody approves twice.
  • 05
    Defence in depthOperator groups and named rights stay the outer wall [6]. The gate adds intent checks those rights cannot express.
Sources

Grounded in Adobe Experience League

Adobe’s documentation names the capabilities and the checks. Where it is silent, such as which agent capabilities are read-only, the gate assumes the worst case.