Production AI workflow governance is not a policy document wrapped around an agent. It is a set of product boundaries that shape what the workflow can do.
Those boundaries include who can start work, what data each step may see, which systems it can touch, how secrets are handled, when people review, and what evidence is retained.
Kora is built around the idea that AI agents should participate inside those boundaries, not bypass them.
Security is one of the reasons Kora starts from a workflow operating system model instead of an unconstrained agent model.
1. Auth must fail closed
Security-sensitive configuration should fail closed.
If required security configuration is missing, the system should reject requests instead of silently disabling validation. If no approved workflow version is live, a workflow start should fail closed. If a required system connection or model setting is missing, the workflow should not proceed as if nothing happened.
This posture matters because AI workflows often sit near systems of record, documents, provider APIs, and human approvals. A permissive fallback can turn a configuration mistake into an operational incident.
2. Customer boundaries are a product invariant
For customer- or business-unit-scoped work, product actions must stay inside the right boundary before any AI step receives context.
Kora treats that boundary as a product responsibility, not an instruction that an agent is expected to remember. Identity, permissions, and workflow scope should decide what a user, workflow, or agent can see.
That is important for AI because context can be broad. A workflow system must not rely on a model, prompt, or agent instruction to enforce customer data boundaries. The product has to enforce them before data reaches the AI step.
3. Keep bearer secrets out of the workflow content
Secrets should be operational configuration, not workflow prose.
Credentials for ERP, email, ticketing, storage, model providers, and other business systems should live in managed configuration. They should be scoped to the step that needs them and kept out of prompts, generated source, logs, and review messages.
That rule is especially important for agentic workflows. If credentials are placed in prompts or source files, they become difficult to rotate, audit, scope, and redact.
The safer pattern is explicit binding: a workflow step declares the system access it needs, and the product provides only what that step is allowed to use.
4. Environment scope is governance scope
An environment is not just a label. It is a security and operations boundary.
Production and staging can have different settings, credentials, integrations, policies, model configuration, and live workflow versions. A reviewed workflow becomes operational only when it is made live in the right environment.
This gives governance a concrete surface:
- which workflow version is live in production?
- which integrations are available there?
- which credentials and settings are configured?
- which publish attempt failed or succeeded?
- which run came from which live version?
Without environment scope, teams often end up with global settings that quietly change live behavior.
5. Agent authority should be bounded
An agent task should have a bounded role in the workflow.
The workflow should define what the agent may do, what tools it may use, what format or quality the answer must meet, and when a person must review. The agent should not receive general authority to decide what the business process is allowed to do.
This is the difference between "the agent has tools" and "the workflow permits this agent task to use these tools for this purpose."
The second model is more restrictive. It is also more operable.
6. Audit the right actions
Not every event belongs in an admin audit log.
Kora should keep high-level audit records for changes that affect access, live workflow behavior, protected data, or external side effects. Detailed step history belongs with workflow evidence. Integration activity and health signals can be tracked separately for investigation and operations.
This separation keeps audit readable while preserving the detailed evidence needed for workflow inspection.
The NIST Cybersecurity Framework is a useful external reference because its current framing emphasizes governance as part of cybersecurity risk management. In workflow terms, governance needs product-enforced boundaries, not only documentation.
7. Public links need narrow authority
Human review links and artifact delivery links can be useful, but they need narrow authority.
Public review or artifact links should be narrow, expiring invitations to perform one action or view one approved artifact. They should be revocable, auditable, and scoped to the specific review or delivery job.
The principle is that a public link should grant only the specific action it is for. It should not become a portable session or broad system credential.
8. Governance should enable adoption
Good governance does not prevent AI adoption. It lets adoption happen safely.
When the workflow has clear boundaries, teams can start with narrow agent assistance, keep human review where needed, inspect evidence, and expand automation based on outcomes.
Without those boundaries, every AI improvement becomes a risk debate. With those boundaries, teams can move one reliable step at a time.
That is the practical security posture for production AI workflows: explicit scope, explicit authority, explicit evidence, and no hidden trust in the prompt.