From AI demo to production: the readiness checklist

A convincing demo proves that a model can complete the happy path. Production readiness means the surrounding system remains useful when inputs are messy, dependencies fail and people behave unpredictably.

Define the authority boundary

List what the agent may read, propose, change and communicate. Give each action an explicit approval rule. Low-risk, reversible actions may run automatically; consequential or irreversible actions should begin with human approval. Autonomy is a design variable, not the default.

Build an evaluation set

Collect representative tasks, edge cases and known failures. Define what a good answer looks like and score every significant change against the same set. Without repeatable evaluation, teams optimise by anecdote and discover regressions through customers.

Control knowledge and tools

• Limit retrieval to authorised sources and preserve source references.
• Grant the smallest possible permissions to external systems.
• Validate inputs and outputs at every tool boundary.
• Use deterministic software for rules that must always hold.
• Protect system instructions and sensitive context from untrusted content.

Design the failure path

Models time out, integrations return partial data and ambiguous requests occur. Decide when the agent retries, asks for clarification, hands work to a person or stops. Preserve enough state for a person to understand what happened without reconstructing the entire session.

Make behaviour observable

Record the request, relevant context, tools called, approvals, output, latency and cost while respecting privacy requirements. Monitor task success rather than model availability alone. A system can be online and still quietly produce poor work.

Prepare the operation

Name an owner for quality, incidents and source material. Establish how users report problems and how unsafe behaviour is disabled. Document data retention, access review and vendor dependencies. Train users on both the capability and its limits.

The release gate

Before launch, the team should be able to answer: What is success? What can the agent do? Which actions require approval? How do we detect a bad result? How do we stop it? Who responds? If any answer is vague, narrow the scope or strengthen the system before increasing autonomy.

The strongest production agents are rarely the most theatrical. They are constrained, measurable and embedded in a workflow that knows when software, a model or a person should take the next step.