Last updated September 18, 2026
4 min read
A fictional engineering manager approves 37 low-risk prompts in a morning. The 38th asks whether to bypass an authorization check for a migration. It looks like the same button. Human presence exists, but attention, evidence, and decision design have failed. This is an explicitly fictional composite, not a customer result.
In this article
Approval rate is not decision quality
Anthropic reported roughly 93% approval in sampled permission prompts and declining attention as prompts became routine. That is vendor telemetry from one system, not proof that every approval gate fails. It still shows why prompt count is a poor safety metric.
Anthropic on human approval and containment bounds this claim. Vendor telemetry from one permission design, not every human gate.
| Evidence point | Interpretation |
|---|---|
| 1 | 37 routine prompts |
| 2 | 1 risk-bearing decision |
| 3 | Weak gate: continue button |
| 4 | Strong gate: evidence and authority |
Treat the table as a working diagnostic. Replace illustrative labels with repository evidence before making a budget, staffing, or release decision.
A decision packet earns human attention
Present the question, affected modules, options, trade-offs, recommendation, evidence, uncertainty, allowed edits, owner, and expiry. A reviewer should understand why the decision matters without reopening 19 chat turns or reconstructing a hidden assumption.
Order matters because later checks inherit the intent and scope established earlier. A skipped step does not remove uncertainty. It hands that uncertainty to a reviewer under time pressure.
Authority must include the ability to stop
A human gate is ceremonial when the person can only click continue. The owner needs permission to reject the approach, request new evidence, narrow scope, or stop the feature. General automation-bias research explains why monitoring demands can weaken attention, especially when alerts rarely matter.
Automation bias and complacency review adds a second evidence boundary. General human-factors evidence, not coding-specific prevalence.
A procedural fix can be inspected, rehearsed, and improved. A warning without an owner or artifact usually disappears during the next busy sprint.
Reduce prompts and strengthen the remaining gates
Automate deterministic low-risk permissions through explicit policy. Bundle related choices. Route only risk-bearing ambiguity to people with context. Then measure gate count, response quality, reopened decisions, and incidents tied to ignored uncertainty.
paqad-ai v1.67.0 was the current public release when this article was checked on July 21, 2026. Its repository describes risk routing, specialist roles, structured state, deterministic checks, documentation sync, and local audit records. These are product mechanisms, not independent outcome proof.
Use the AI workflow audit guide to map current practice, compare the bar with production-ready AI code, and use the AI coding consultant guide when the team needs outside ownership.
Decision rule: If a prompt carries no meaningful choice or stop authority, remove it from the human gate and handle it through reviewed policy.
Run the human-decision packet in a 45-minute working review with the product owner, module owner, and final reviewer. Do not turn it into a score. Read the decision rule aloud: If a prompt carries no meaningful choice or stop authority, remove it from the human gate and handle it through reviewed policy. Record disagreements, missing evidence, and the person responsible for the next action. Repeat the review after one real feature so the team can compare whether the artifact reduced ambiguity or only added paperwork.
Frequently Asked Questions
What decisions need a human gate?
Use one for product ambiguity, architecture trade-offs, sensitive data, security exceptions, irreversible migrations, external side effects, or risk acceptance. Routine deterministic operations should follow reviewed policy.
How do you avoid approval fatigue?
Reduce prompt volume, bundle related decisions, route by risk, and give the reviewer a compact packet. Measure whether approvals contain reasons instead of counting clicks.
Can the AI recommend an option?
Yes, if the recommendation includes evidence, uncertainty, and alternatives. The recommendation should not hide who owns the final decision or prevent that person from stopping the work.
What next?
Apply the human-decision packet to one current feature. If the repository cannot answer the questions it exposes, install paqad-ai and test a governed workflow with local evidence and explicit human authority.