The Short Answer
The question isn't whether to include a human in the loop, but precisely where. Approving every step eliminates savings and creates fatigue, leading to automatic sign-offs. Conversely, failing to require approval for irreversible actions results in the kind of incidents that reach executive management.
The practical approach begins by breaking down the process into individual actions, classifying each action by its reversibility and impact, and selecting an appropriate approval pattern—blocking, post-hoc, or sampling. Then, design the decision screen so that approval is a genuine judgment, not just a click.
The governance framework within which these decisions are made is detailed in AI Governance for Agentforce.
Decision Matrix: Reversibility vs. Impact
| Low Impact | High Impact | |
|---|---|---|
| Easily Reversible | No approval, monitoring only | Sampling of a percentage of cases for review |
| Reversible at Cost | Post-hoc review | Blocking approval before execution |
| Irreversible | Blocking approval before execution | Blocking approval + written justification + Audit trail |
Reversibility is measured by three questions: how long it takes to undo, how much it costs, and who is exposed before it's undone. A message sent to a customer is irreversible, even if a correction can be sent—the impression has already been made. Updating an internal field is reversible in a second and therefore doesn't warrant a stop.
Three Approval Patterns
Blocking approval stops an action until a human decides. It is the most time-consuming pattern and is therefore reserved for irreversible or high-impact actions. The rule: if we stop, the human must receive everything necessary for the decision on a single screen.
Post-hoc review allows the agent to act and submit the result for review within a defined timeframe. Suitable for low-cost, reversible actions, it requires two conditions: a short timeframe and a genuinely functional undo button.
Sampling reviews a percentage of cases, not all of them. This is the correct pattern for high-volume, low-impact actions, and it also serves as a quality mechanism that later justifies removing a blocking approval elsewhere.
Designing the Decision Screen
This is the part that determines whether the Human-in-the-Loop is genuine or merely formal. A screen that only displays the proposed action will receive automatic approval. A screen that requires opening three tabs for verification will cause delays and workarounds.
Four components should appear together: the proposed action in one line, the business-language explanation, the source the agent relied on with a link to the extracted section, and the confidence level or flags that were triggered. Below that, three options, not two: Approve, Reject with Reason, and Amend before Execution.
The reason for rejection is the most valuable asset in the process. It generates the training and testing set for the next version, so it should be a short list of common reasons, not an open text field that no one fills out.
How these decisions are monitored over time is explained in Observability for AI Agents.
Preventing Rubber Stamping
Approval fatigue is an expected failure, not a surprise. Three warning signs: average decision time drops below a few seconds, rejection rate approaches zero, and approvals are concentrated at the end of a shift.
The solution isn't training, but design. Reduce the number of approvals by removing unnecessary stops for reversible actions, add post-hoc sampling that creates accountability, and provide feedback to the approver on the quality of their decisions. When approvers know that a percentage of their decisions are reviewed, they start paying attention again.
Key metric: If the correction rate is zero for months, either the agent is excellent and approvals can be reduced, or no one is reviewing. Sampling distinguishes between the two.
When and How to Remove an Approval Point
Removal is a data-driven decision, not a gut feeling. Three cumulative conditions: sufficient volume of documented decisions in the category, a consistently low correction rate for several consecutive months, and a genuinely tested undo or rollback mechanism.
Removal is staggered. First, to a narrow subset—for example, only amounts below a certain threshold, only customers in a specific segment, only during business hours. Then, measure for a month. Only then expand. Simultaneously, post-hoc sampling remains even after removal, otherwise there's no way to detect deterioration.
There also needs to be a rollback condition: if the error rate crosses a predefined threshold, the blocking approval automatically resumes. Without a written rollback condition, the decision to remove becomes irreversible itself.
Scenario: A Service Center That Removed the Wrong Approval
A telecommunications company deployed an agent (Agentforce) that prepared responses to billing inquiries. In the first version, every response required agent approval, including simple status updates. Handling time decreased only slightly, and agents complained they were reading the same text repeatedly.
The team analyzed 4,000 approvals and found that 78% were for informational responses only—completely reversible and low impact. The correction rate in this category was below one percent. In contrast, the correction rate for credit adjustments was significantly higher.
The decision: Remove blocking approval for informational responses, with sampling of a percentage of them for weekly review; reinforce approval for credit adjustments and add a mandatory written justification above a certain amount. Handling time decreased significantly, and rejections in the credit adjustments category actually increased—a sign that approvers were paying attention again.
Risks and Preventive Actions
| Risk | What it Looks Like | Preventive Action |
|---|---|---|
| Approving Everything | Value is nullified, users bypass | Impact-Reversibility Matrix for every action |
| Automatic Sign-off | Decision time in seconds, zero rejections | Post-hoc sampling and personalized feedback for approver |
| Screen without Justification | Approver cannot truly judge | Display justification, source, and confidence level on one screen |
| Premature Removal | Production incident after a few good weeks | Measurable removal thresholds, phased expansion, and rollback conditions |
| No Rejection Documentation | Nothing to learn for the next version | Short list of rejection reasons and weekly review |
Metrics
| Metric | What it Reveals | Frequency |
|---|---|---|
| Stop Rate | Percentage of actions that required approval | Weekly |
| Correction Rate | Percentage of cases the approver changed or rejected | Weekly |
| Median Decision Time | Is the approver actually reading? | Weekly |
| Sampling Findings | Errors that passed approval | Monthly |
| Actual Undoing Time | Does the rollback mechanism work? | Per version |
For guidance in planning approval points and aligning them with regulatory requirements, the Agentforce and AI Service is the practical path forward.
Checklist
- ☐ Process broken down into individual actions
- ☐ Each action classified by reversibility and impact
- ☐ Approval pattern selected for each action: Blocking, Post-hoc, or Sampling
- ☐ Approver defined according to existing process authority
- ☐ Decision screen displays action, justification, source, and confidence level
- ☐ Option for amendment exists, not just approval or rejection
- ☐ Rejection reasons documented in a short list
- ☐ Measurable thresholds for approval removal and rollback conditions defined
- ☐ Rollback mechanism is practically tested, not just planned