Before the checklist
Start with a defined first release
This guide assumes that you have chosen one real customer conversation to improve. If the purpose, approved next action or human owner is still unclear, complete Your First 30 Minutes before moving forward.
The checklist is designed to expose readiness gaps early. Completing an item means that the decision has been confirmed or the behavior has been tested—not simply that a setting was opened.
Have these six inputs ready
Unknowns may be recorded for follow-up, but critical facts and ownership decisions must be resolved before release.
- 01One starting conversation
- 02One approved next action
- 03Current source material
- 04Human owner and escalation destination
- 05Required account and channel permissions
- 06A private or controlled test path
Eight launch steps
The quick-start checklist
Work through the stages in order. Later steps depend on the decisions and evidence created earlier in the sequence.
Define
Confirm the first use case
Choose one repeatable customer conversation and one appropriate next action. Keep the first release narrow enough to test clearly.- Name the customer, channel or entry point.
- Describe the question or decision the conversation should support.
- Choose the booking, estimate, application, purchase path or human handoff that may follow.
- Record what is intentionally outside the first release.
Prepare
Gather and approve the knowledge
Assemble the smallest reliable source set that can support the use case. Resolve contradictions before the information reaches the AI Agent.- Collect current pages, documents, FAQs, policies and approved decisions.
- Confirm pricing rules, availability, eligibility and important exceptions.
- Remove expired, duplicate or unapproved material.
- Assign an owner for future source updates.
Configure
Prepare the AI Agent
Translate the first-release brief into a defined purpose, voice, boundaries, qualification logic and handoff behavior.- State the AI Agent's purpose in one clear sentence.
- Set the voice, response style and level of detail.
- Define questions used to understand or qualify the request.
- List facts, promises and decisions the AI Agent must not invent.
- Define active hours and what should happen outside them, when relevant.
Connect
Prepare the first entry point
Connect only the channel or customer context required for the first release. Confirm permissions and ownership before testing.- Confirm the correct business account, page, number or web property.
- Verify required permissions and administrator access.
- Use a private test path or approved test contact where possible.
- Keep unrelated channels and workflows unchanged.
Route
Configure the next action and handoff
Make the next step explicit and give escalations a real destination. A handoff is not complete until a person or monitored team owns it.- Confirm the approved customer next action.
- Choose the person, team or monitored destination for escalations.
- Define what context should be passed with the handoff.
- Verify that a person can pause, review or take over when needed.
Test
Run representative conversations
Evaluate the entire path—not only whether the AI Agent can produce a plausible answer. Test knowledge, behavior, routing and human control together.- Test common, ambiguous, difficult and out-of-scope questions.
- Test a missing or outdated fact.
- Test qualification and the intended next action.
- Test sensitive issues and human escalation.
- Correct critical failures and repeat the affected scenarios.
Release
Approve a controlled launch
Release the smallest stable version to a limited audience or context while a human reviewer is ready to monitor early activity.- Confirm all three release gates have passed.
- Choose the initial audience, channel and launch window.
- Name the person monitoring early conversations.
- Confirm how the experience can be paused or taken over.
Improve
Review early activity before expanding
Use real conversations to find knowledge gaps, unclear phrasing and routing friction. Stabilize the first path before adding more scope.- Review the first real conversations for accuracy and usefulness.
- Correct gaps, contradictions and missed escalation triggers.
- Confirm that next actions and handoffs reached the right destination.
- Expand only after the current path is stable.
Behavior coverage
Test more than the happy path
A plausible response to an easy question is not enough. Use representative scenarios to confirm that the AI Agent stays within scope, handles uncertainty and reaches the correct next action or human destination.
Common question
Uses approved knowledge and gives a useful next step.
Ambiguous wording
Clarifies instead of making an unsafe assumption.
Missing or outdated fact
Avoids inventing an answer and follows the defined path.
Out-of-scope request
States the boundary and redirects or hands off appropriately.
Sensitive complaint
Uses the required escalation path without overpromising.
Qualification
Asks only the necessary questions and routes the result correctly.
Human handoff
Passes useful context to a monitored destination.
Follow-up context
Handles a repeat or continued conversation coherently.
Note the input, expected behavior, observed result, correction and retest status. This creates evidence for the release decision.
Go or pause
All three release gates must pass
A first release is ready only when its information, behavior and human ownership work together. A strong score in one area does not cancel a critical failure in another.
Knowledge
The approved sources are current, relevant and free of unresolved critical contradictions.
Required: passBehavior
The AI Agent answers and acts within scope across normal, ambiguous and difficult tests.
Required: passOwnership
The next action, escalation destination and human take-over path are working and monitored.
Required: passIf any gate is not ready, pause the launch, correct the issue and repeat the affected tests.
First exposure
Launch narrow, observed and reversible
A controlled release limits the number of variables changing at once. Start with one defined customer context and make sure a human reviewer can see what happens.
Limit the scope
Use the selected channel, entry point, audience or launch window. Avoid enabling several new paths simultaneously.
Watch the activity
Assign a reviewer for the early conversations and give that person the context needed to judge quality.
Keep human control
Know how to pause, correct or take over the experience if a critical issue appears.
First conversations
Monitor early activity before you expand
Review the first real conversations closely enough to understand what customers actually ask, where the source material is thin and whether the intended next action feels natural.
Answer quality
Check accuracy, relevance, clarity and whether the response stays within the approved boundaries.
Knowledge gaps
Identify missing, outdated or contradictory information and update the approved sources.
Routing
Confirm that qualified requests, next actions and escalations reach the intended destination.
Customer friction
Notice repeated clarifications, confusing steps and abandoned paths—not only reply volume.
Add another use case, workflow or channel only after the current path is stable and its early corrections have been tested.
Pause the release
Stop conditions that require correction
Do not release—or pause an early release—when a critical condition below is present.
Record the decision
What to capture when the checklist is complete
A short release record makes later reviews easier and prevents the first scope, sources and ownership decisions from becoming informal team memory.
ScopeUse case, entry point, intended next action and exclusions.
KnowledgeApproved sources, source owner and last review date.
ConfigurationAI Agent purpose, boundaries, qualification and active hours.
OwnershipMonitoring owner, escalation destination and take-over path.
EvidenceTest scenarios, outcomes, corrections and retest status.
ReleaseApproval, launch context, monitoring window and first review point.
Continue the launch path