Before allowing AI to act, define who approves what. Plan waiting, rejection and changed requests, and agree on acceptance criteria before commissioning development.
When introducing an AI agent, evaluate the quality of its proposal separately from its authority to act. A specification that merely says “a human reviews it” leaves the evidence, approval scope and handling of changes unclear. Start with one workflow and define where execution pauses and who takes responsibility next.
What the official documentation says about confirmation
Google ADK provides Tool Confirmation to pause a tool workflow and receive confirmation from a user or another system. The documentation describes both yes/no confirmation and structured responses. At the time of review, the feature is experimental and has session-service limitations, so compatibility must be checked in the intended environment.
Google ADK: Action confirmations
The following is For f’s design guidance. It does not require a particular product; it helps business and development teams agree on scope.
Approve a specific action, not just an explanation
Show the transaction or customer, before-and-after values, recipients, attachments and reversibility in the review screen. A generic “May I proceed?” does not reveal the impact. Verify that the explanation matches the actual payload, and define when a changed payload requires renewed approval.
Define execution and approval permissions separately
Someone allowed to view a record is not necessarily allowed to authorize changes. Document the requester, approver and executing system, including delegation and absence. Require authorization checks at the execution boundary as well as appropriate controls in the interface.
Treat waiting, rejection and expiry as normal workflow states
Assume that stock or amounts may change while approval is pending. Let the business set the waiting period, and verify that expired or rejected requests cannot proceed automatically. Include notifications, rejection reasons and resubmission procedures so teams can identify why work is left pending.
Illustrative example: preparing and sending a quotation
Consider an AI-generated quotation draft that a sales representative reviews before sending. In the prototype, disable actual sending and evaluate the approval screen for products, quantities, discounts and recipients. Test whether changing a discount invalidates the earlier approval. This is a design example, not a For f deployment or performance claim.
Evidence to retain during acceptance testing
- Records linking the approver, approval time, approved version and execution result
- Tests showing that rejection, expiry, insufficient authority and changed content prevent execution
- A specification separating review-screen information from information permitted in logs
Requiring approval for every action can burden reviewers. Initially focus on consequential operations and record the queue size, rejection reasons and review time. Use those observations and the business’s acceptable risk to decide where automation can expand, documenting any conditions for bypassing review.
What to prepare before a consultation
Bring the current request-and-approval workflow, desired actions, required review fields and absence procedures. If requirements are still unclear, prototype the review screen and state transitions before building write operations, and check whether business reviewers can make an informed decision.
Is an approval button enough?
No. The execution system must verify who approved which version and under what authority. Include changes made after approval and insufficient information for reviewers in acceptance testing.
Discuss the scope with For f
We can start by defining the scope around your current workflow and what you need to establish. You can discuss what you know even if documentation is incomplete.
Discuss this topic / Book an online consultation
References and verification date
Publication and update dates were not available on the official pages.
Information checked on September 22, 2026. Product capabilities and conditions may change. Recheck the current official documentation and your environment before implementation.
The thumbnail is an AI-generated concept image, not an actual system screen or measured result.