Access proves who can act.
Authentication and permissions establish identity and scope. They do not evaluate whether a specific action is appropriate right now.
Zero-trust control for agentic operations
Never let AI mutate production data unsupervised.
Coatcheck is a control layer in development for AI agents operating across Salesforce and manufacturing systems. It checks risky AI actions at the door — intercepting high-impact actions, checking them against policy, and routing them for human approval before anything touches production.
POLICY DECISION / PRE-COMMIT
ACTORagent / procurement-copilot
INTENTrelease purchase order
TARGETERP / production
POLICYhuman approval required
HOLD_FOR_APPROVAL
Mutation paused before commit. A designated owner receives the context needed to approve or reject.
Illustrative request, not a live production event
The risk isn’t what agents say. It’s what they change.
Most AI safeguards stop at the conversation. Enterprise risk begins when an agent gains authority to alter orders, pricing, inventory, customer records, or financial terms.
Guardrails shape what agents say.
Prompt controls and model policies reduce harmful output. They do not govern the downstream transaction an agent is about to commit.
Coatcheck is designed to govern the mutation.
The control decision is designed to happen at the transaction boundary—after intent is known, before production state changes.
A control loop built around the moment of consequence.
Select each stage to inspect how an agent request moves from intent to a governed decision.
Step 01 · Intercept
Catch the action before the system of record.
The agent sends its intended mutation through Coatcheck rather than writing directly to the target platform. Context travels with the request.
Illustrative control event
eventmutation.requested
boundarypre_commit
contextactor + intent + target
stateintercepted
Watch the loop run.
Thirty seconds, end to end: an Agentforce agent asks to create a $42,000 purchase order, Coatcheck holds it for a human, the approver taps once in Slack, and the order completes. Recorded October 2026 in a demo Salesforce org.
Demo org · Acme Metals is fictional · no production data was touched
One agent, three verdicts.
Same agent, same policy, three requests: an account lookup sails through instantly, a $42,000 purchase order is held and approved, a $200 refund is held and denied. The policy reads every interaction on its own. Recorded October 2026 in a demo Salesforce org.
Demo org · Acme Metals is fictional · no production data was touched
The deny.
A $200 refund, held for a human. One tap: denied — nothing was changed. No is a complete answer, and a healthy system hears it regularly. Recorded October 2026 in a demo Salesforce org.
Demo org · Acme Metals is fictional · no production data was touched
#coatchecked: The new vendor.
The policy: POs under $500 flow freely — unless the vendor is new. The agent saw "$450, under $500" and ran it straight through, missing the unless. The policy didn't. Recorded October 2026 in a demo Salesforce org.
Demo org · Acme Metals is fictional · no production data was touched
The safety net.
The policy was loosened: refunds under $250 auto-approve. But a hardcoded guardrail in the code always holds refunds — no exceptions. The policy said yes. The safety net said no. Your policy is configurable. Your safety floors aren't. Recorded October 2026 in a demo Salesforce org.
Demo org · Acme Metals is fictional · no production data was touched
Own the middle—not every system around it.
Coatcheck is designed as middleware: one governance boundary between agent reasoning and the systems that hold production truth.
Requesters
AI agents
Agentforce, copilots, and automated workflows propose actions with their operating context.
intentactorscope
Control boundary
Coatcheck
Evaluate policy, hold risky actions, collect accountable approval, and return an explicit decision.
policyapprovalevidence
Systems of record
Production platforms
Salesforce, SAP, Oracle, and connected operations systems receive only permitted mutations.
CRMERPoperations
Controls made for enterprise operations.
The model applies familiar governance disciplines to autonomous software actors—without forcing every target system to reinvent them.
Default deny for consequential writes
A high-impact action does not pass merely because an agent has access. It must satisfy an explicit policy.
Human authority stays legible
Approvers see what is changing, why the agent proposed it, and which rule caused the hold.
Separation of duties survives automation
The system requesting an action is not the same authority that approves its own exception.
Every decision leaves evidence
The control path is designed to produce an audit-ready account of the request, policy result, human decision, and final outcome.
Design partner pilot
Bring us the agent workflow that makes you uneasy.
We’ll put a working governance layer in front of one Salesforce agent workflow and shape the product around how your team actually makes decisions.
You getPolicy checks, Slack approvals, and a verifiable audit trail
You bringOne real workflow and a weekly 30-minute feedback call
It costsNothing during the proof-of-concept pilot
Autonomy can scale. Accountability has to scale with it.
Coatcheck is being developed for manufacturing environments where agentic workflows cross CRM, ERP, and operational boundaries.
Exploring a governed agent workflow? Let’s talk. [email protected] →