AI agent implementation checklist: what must be ready to launch?

2. Confirm sources, access and maintenance

List the sources and systems required for the first workflow. Obtain approved test material and record how examples may be used. Check source ownership, quality, currency and withdrawal. Keep missing and conflicting information visible.

• Verify access for each user role and record type.
• Test that retrieval, outputs and logs preserve intended permissions.
• Define what is stored, for what purpose and for how long.
• Name the person who updates or retires source material.
• Record supplier and processing dependencies for the responsible review.

For staff answers, the internal knowledge solution expands source readiness. Document security and privacy decisions for the actual use case; this checklist is not a certification.

3. Prove each operation and its boundary

Inventory the permitted tools with example inputs, outputs, effects and failures. Use test environments or controlled fixtures before enabling real operations. Confirm that server-side checks enforce user authority independently of model output.

• Separate reading, preparing and executing actions.
• Preview the actual proposal when human approval is required.
• Invalidate approval after material changes or expiry.
• Handle duplicate requests and uncertain execution outcomes.
• Set time, retry and usage limits with a visible stopping path.

The approval guide explains the decision record. Use the integration guide to resolve destination-system behaviour before the release review.

4. Test outcomes the business recognises

Create representative cases before tuning and keep some separate for final evaluation. Include ordinary work, ambiguity, missing sources, conflicting evidence, unavailable tools, malicious instructions and requests outside scope. Define the expected outcome for each.

Anthropic’s agent evaluation guidance distinguishes the final environment state from what the conversation says happened. Verify that the intended record or action exists, not just that the response sounds successful.

• Have the business owner review evidence and material failures.
• Test permission denial, human rejection and stopping behaviour.
• Check desktop, mobile, keyboard access and relevant channel conditions.
• Record the tested model, prompts, tools and source configuration.
• Agree which failures block launch rather than hiding them in an average score.

5. Prepare the people who receive real work

Assign the support owner, exception queue and escalation path. Show operators enough context to continue a task without replaying every conversation. Provide a way to pause the workflow and explain how existing tasks will be handled.

• Make failed, waiting and pending-confirmation states visible.
• Verify alerts reach the person responsible for acting.
• Prepare a recovery procedure and test it.
• Control access to logs and retained evidence.
• Review accepted outcomes, corrections, unresolved age and total operating cost.

NIST’s AI Risk Management Framework provides a voluntary framework for managing AI risks across use and evaluation. Apply review to the actual business context and responsibilities.

6. Launch the smallest complete workflow and learn

Hold a release review with the business owner and delivery team. For each unresolved item, record its consequence, owner and next step. Do not postpone a dependency that the first useful outcome requires, such as receiving-team ownership or a working permission check.

Hypothetical release decision: a service-intake agent may launch with reviewed clarification drafts while automatic external sending remains deferred. It can still complete the intended first slice if staff have an effective review queue and the agreed intake outcome is verified.

Estimate implementation with the team around that slice. AI coding tools can assist development and test preparation, but source readiness, integration access, user acceptance and operating support remain real dependencies. Avoid a universal budget or duration promise.

After launch, compare results with the baseline and review changes before widening the task or tool scope. Use AI agent ROI for the business case and LCR’s AI Agents page for the delivery conversation.

Questions

Frequently asked questions