Approval fatigue sets in when an agentic system generates more approval requests than the people reviewing them can meaningfully evaluate, and the reviewers respond the only way anyone can under that kind of volume: they start approving quickly, without real scrutiny, because slowing down for every request isn't sustainable. This is a specific and somewhat predictable failure of human-in-the-loop design, and it's more common with AI agents than with older systems, because an agent can generate approval requests at a pace and volume no human process was designed to absorb, especially once an organization scales up the number of agents or the range of actions they're allowed to attempt.
Why agents produce this problem faster than humans do
A human employee submitting requests for approval is naturally rate-limited by how fast a person can think through and initiate a request. An agent has no such natural limit. If an approval gate is placed too broadly, covering routine, low-impact actions alongside genuinely sensitive ones, the volume of requests can climb quickly as the agent takes on more tasks, and the reviewer's attention doesn't scale to match it.
What fatigue actually looks like in practice
The visible symptom isn't reviewers complaining, it's approval rates climbing toward the ceiling and denial rates dropping toward zero, even for requests that would have raised questions if anyone looked closely. Reviewers start recognizing patterns and approving based on the pattern rather than the specific details of each request, which works fine until a request that superficially resembles a normal pattern is actually something different underneath.
Narrowing the approval surface deliberately
The most direct fix is reducing what actually requires approval, not adding more reviewers to keep up with volume. Reserve approval gates for actions that are genuinely high-impact or irreversible, and let lower-risk, reversible actions proceed under monitoring instead. This isn't about approving less carefully, it's about making sure every request that does require approval is one that deserves the attention it's supposed to get.
Batching and summarizing similar requests
Where many requests share a similar shape, refunds under a certain amount, routine data lookups, batching them into a single summarized review, rather than presenting each one individually, can reduce fatigue without eliminating oversight entirely. The reviewer sees the pattern and can flag anomalies within the batch, rather than clicking approve on the same shape of request dozens of times in a row.
Rotating and pacing reviewer load
Concentrating all approval responsibility on one person guarantees fatigue sets in faster. Spreading approval responsibility across a team, and making sure any individual reviewer's queue stays within a volume they can realistically evaluate carefully, keeps quality of review higher than dumping the entire load on whoever happens to be on call.
Measuring whether approvals still mean anything
Track approval and denial rates over time, and treat a denial rate that's dropped near zero as a signal worth investigating, not a sign everything is working smoothly. A healthy approval gate should produce some denials; if it never does, either the gate is scoped to actions that are never actually risky, or the review has stopped being real.
Distinguishing genuine fatigue from a poorly scoped gate
Not every high approval rate is fatigue. Sometimes it reflects a gate that was scoped correctly and the agent genuinely, consistently proposes legitimate actions, in which case a high approval rate is a sign the system is working, not a red flag. The distinction matters because the fix is completely different: fatigue calls for narrowing what requires approval or spreading reviewer load, while a well-calibrated gate with a naturally high approval rate might just need occasional spot-checks rather than a redesign.
Telling the two apart usually requires looking past the raw approval rate and into how the approvals happen: quick, uniform response times regardless of request complexity, no variation in how long a reviewer spends on a large request versus a small one, and no record of reviewers asking follow-up questions are all signs pointing toward fatigue rather than a genuinely low-risk request stream. A healthy gate shows some variability in review behavior that tracks the complexity of what's being reviewed.
Where TELEON fits
TELEON's audit trail records every approval decision it enforces at the tool-call boundary, which gives teams the data needed to notice approval fatigue setting in, denial rates trending toward zero, approval times shrinking suspiciously, before it becomes a serious blind spot. Deciding which actions actually need approval, and redesigning workflows to reduce unnecessary volume, is a judgment each organization has to make based on its own risk tolerance.
A short checklist
- Reserve approval gates for genuinely high-impact or irreversible actions, not routine ones.
- Monitor approval and denial rates as an ongoing signal, not a one-time design check.
- Investigate a denial rate that trends toward zero rather than treating it as good news.
- Batch or summarize similar low-risk requests instead of presenting them one by one.
- Spread approval responsibility across a team rather than one overloaded reviewer.
- Periodically audit whether approvers can articulate why they approved specific requests.
- Reassess and narrow the approval surface as agent task volume grows over time.
Approval fatigue is what happens when a control designed for careful human judgment gets fed more volume than judgment can actually process. The fix is almost never "ask people to pay closer attention"; it's reducing what reaches them in the first place so the requests that do arrive are ones that genuinely warrant the scrutiny the workflow assumes they're getting.