Skip to content

Free agency resource

Client AI agent launch checklist

A blank worksheet for the people, permissions, costs and acceptance tests behind a client deployment. Includes a fictional example you can adapt to your own workload.

Download Markdown

No signup required. Copy, edit or print it.

Use this worksheet when an agency or independent builder prepares an AI agent for a client. Complete it for one workload at a time. It works as a document or spreadsheet; it does not assume the hosting platform supplies every control listed here.

Start with the client's expected result. A successful deployment is evidence that something started. The acceptance record below establishes whether the intended task works, reaches the right person and stays within its agreed permissions.

The checklist and decision rule are a proposed operating method. Tailor them to the actual workload and agreement. Store references to credentials, never secret values.

1. Name the workload and its owners

FieldFill in
Client name
Workload name and stable internal ID
Business outcome
Business owner and acceptance contact
Technical owner
Backup operator
Client escalation contact and channel
Expected operating hours and timezone
Human fallback when the agent is unavailable
Target launch date
Document owner and last reviewed date
  • ☐ A named person can approve changes to the task and its permissions.
  • ☐ A named person receives and acts on operating failures.
  • ☐ The client knows how to use the human fallback.

2. Define exactly what the agent may do

DecisionFill in
Accepted inputs and their source
Required output and delivery destination
Allowed reads
Allowed writes
Actions requiring human approval
Actions the agent must refuse
Expected behavior when input is incomplete
Completion deadline or delivery expectation
How the destination confirms completion
What happens when a request arrives twice
  • ☐ Required approvals are reflected in connected-service controls where available.
  • ☐ A prompt change alone cannot silently expand the agreed authority.
  • ☐ The output is useful enough for the client to evaluate without reading runtime logs.

3. Record the actual deployment boundary

ComponentIdentifier or referenceOwnerWho can access it?
Hosting account and organization
Project and engine/runtime
Machine/container, where exposed
Agent/profile
Workspace and state locations
Connected services
Human access roles
  • ☐ The team has identified which component provides the workload's actual security boundary.
  • ☐ Separate names, profiles or folders have not been treated as proof of access isolation.
  • ☐ Authorized and unauthorized access cases appear in the acceptance matrix.
  • ☐ The access arrangement is appropriate for the client and any other workloads sharing infrastructure or administration.

For cloudrun, the isolation guide distinguishes the engine boundary from Hermes profiles and explains the limits of the dated evidence.

4. Assign account and credential ownership

Service/accountBusiness ownerAuthorized scopeSecret-store referenceRenewal ownerRevocation procedure reference
Model provider
Source system
Delivery channel
Other integration
  • ☐ Connected-account ownership matches the client agreement.
  • ☐ Only required permissions are enabled where the service supports them.
  • ☐ Renewal and authorization failures have an owner.
  • ☐ Revocation is documented in the external service as well as the hosting configuration.
  • ☐ The worksheet contains no API keys, tokens, passwords or recovery codes.

The BYOK guide explains the difference between removing a key from configuration and revoking it at the provider.

5. Record cost and stop decisions

Cost or decisionApproved amount, rule or referenceOwner
Hosting
Model usage
Connected-service usage
Setup and operating time
Notification threshold
Stop threshold and action
Where actual usage is checked
Review frequency
  • ☐ The client has an understandable estimate of the complete workload cost.
  • ☐ Someone can observe usage in the relevant services.
  • ☐ A documented action exists when usage reaches the agreed threshold.

Use the always-on agent cost guide to distinguish hosting from model usage, other services and operating work. The budget fields above are for your actual agreement; no default amount is implied.

6. Freeze the configuration to be accepted

ConfigurationVersion or reference
Runtime
Model/provider configuration
Instructions/prompt
Skills and custom tools
Integration permissions
Input/output schema, if applicable
Schedule and timezone
Channel routing
Known working configuration to restore
  • ☐ The accepted configuration can be identified later.
  • ☐ Required files and tools exist on the target host.
  • ☐ The test environment uses harmless fixtures or appropriately authorized data.

7. Run the client acceptance matrix

Write expected behavior before testing. In each evidence field, keep the run ID, timestamp and destination reference that let an operator verify the outcome. Keep confidential payloads in their authorized system.

Mark a case Pass, Fail, Not run or Not applicable. Explain each Not applicable decision. Do not count a skipped case as a pass.

CaseInput or conditionExpected behaviorObserved result and evidenceStatusReviewer
Ordinary taskNot run
Missing required inputNot run
Incorrect or conflicting source dataNot run
Unauthorized requesterNot run
Other client's data or credentialNot run
Action requiring approvalNot run
Disconnected/expired integrationNot run
Model/provider unavailableNot run
Duplicate request or deliveryNot run
Scheduled run and timezoneNot run
Destination receives the resultNot run
Usage threshold or stop ruleNot run
Restart with pending workNot run
Pause and controlled resumeNot run
Recovery of required stateNot run

Acceptance tests concern behavior under the conditions tested. They are not an exhaustive security audit or a promise about every future model response.

8. Prepare the operating handoff

  • ☐ The operator knows where to see the last successful task and any pending work.
  • ☐ Alerts identify the workload, failed stage and action owner without exposing secret or client content.
  • ☐ A missed expected run has a detection method.
  • ☐ The incident procedure identifies how to stop new actions and preserve relevant evidence.
  • ☐ Before retrying a failed action, the operator checks whether it already changed the destination.
  • ☐ A model, prompt, tool or runtime change triggers the relevant acceptance cases again.
  • ☐ Support hours, escalation and exclusions match the actual client agreement.
Operating itemProcedure or referenceOwner
Health/outcome check
Failure alert
Stop new work
Inspect external actions before replay
Restore known configuration
Approve and roll out a change
Incident communication

9. Rehearse recovery

State neededIncluded in which export/backup?Last verified copyRestore procedureTest result and date
Instructions/configuration
Skills/custom tools
Required files
Memory/history required by the agreement
Credentials or reauthorization steps
Schedules and channel routing
External-system records
  • ☐ A configuration template and an installation image are not being represented as a complete customer backup.
  • ☐ The team knows which required state is omitted from each export.
  • ☐ Recovery was tested in a disposable destination with outbound actions controlled.
  • ☐ The recovered setup passed a representative acceptance case.
  • ☐ Any stated recovery time or acceptable data loss follows the agreement and tested procedure.

Read the backup and recovery guide before relying on a platform or runtime export.

10. Make launch a recorded decision

Proposed decision rule: Launch only when every required acceptance case has passed, the client has accepted the output, and the operating and offboarding owners are assigned. Resolve failures or reduce the scope and rerun the affected cases. Record a reason for every case excluded from the agreed scope.

Launch decisionRecord
Required cases and any excluded cases with reasons
Open issues and effect on the agreed scope
Accepted configuration reference
Final decision: launch / hold
Client acceptance contact, decision and date
Technical owner, decision and date
First post-launch review date

11. Plan offboarding before you need it

Handoff itemRecipientExport/transfer procedureAccess to revokeCompletion evidence
Instructions, skills and configuration
Required files and agreed history
Model account
Connected services
Channels and schedules
Hosting administration and billing
Backups and retained copies
  • ☐ The agreement says what the client receives and in which format.
  • ☐ Pending tasks and external changes will be reconciled before shutdown.
  • ☐ Only the intended destination remains active after a transfer.
  • ☐ The recipient verifies that the delivered material is usable.
  • ☐ Agency access is removed from each relevant external account.
  • ☐ Resource deletion, subscription cancellation and external credential revocation are checked separately.
  • ☐ Retained copies have an agreed purpose, owner and removal date.

The agent migration guide provides a more detailed cutover and rollback sequence.

Fictional worked example: a weekly draft brief

This example is fictional and has not been executed. It illustrates how to fill the worksheet. “Example Client,” the role names and the workflow are invented. No test result, customer outcome or cloudrun integration capability is implied.

Workload: Prepare a weekly draft summary from an approved project-notes folder. Deliver the draft to the client's reviewer. The agent must not publish it or change the source records.

FieldExample entry
ClientExample Client
Workload IDexample-weekly-project-brief
Business ownerClient project lead
Technical ownerAgency automation engineer
Backup operatorAgency duty operator
InputApproved sample project notes in a test folder
OutputDraft summary with source references and a missing-information section
Allowed actionsRead the approved folder; deliver the draft to the reviewer
Approval boundaryReviewer decides whether and where to publish the draft
Account ownershipClient owns source and delivery accounts; agency receives the approved access
Credential recordReferences to the client's chosen secret store; no values in this worksheet
Schedule proposalMonday morning in the client's agreed timezone; exact trigger to be configured and tested
Failure fallbackNotify the named operator; reviewer prepares the brief manually if needed
Recovery approachKeep a reviewed configuration copy and verify the host's export/restore support before launch
BudgetAwaiting client agreement; no launch budget approved in this example

Example acceptance cases

CaseInput or conditionExpected behaviorStatus/evidence
Ordinary taskHarmless notes with three completed project updatesDraft identifies all three updates and links to their sourcesNot run; no evidence
Missing informationA note lacks an ownerDraft flags the missing owner without inventing oneNot run; no evidence
Unauthorized sourceRequest to read an unrelated client's folderAccess denied; no unrelated content in the draftNot run; no evidence
Publication requestMessage asks the agent to publish the briefAgent leaves publication to the reviewer; write access remains unavailableNot run; no evidence
Duplicate triggerThe same scheduled task arrives twiceBehavior follows the agreed duplicate rule; reviewer receives no unintended duplicateNot run; no evidence
Disconnected sourceSource authorization is revokedFailure is visible to the operator; agent does not present a stale brief as currentNot run; no evidence
DeliveryDraft generation succeedsThe intended reviewer receives the draft onceNot run; no evidence
Offboarding rehearsalAgency access is removedClient retains agreed artifacts; agency can no longer use the connected accountsNot run; no evidence

Example decision: Hold. The budget is unapproved, the integration path has not been validated and acceptance cases have not run. The next step is to configure a harmless pilot, verify the source and delivery connections, agree on cost and execute the matrix.