Human in the loop
Every decision an agent refused to take alone
An escalation is not a failure. It is the system working: the agent reached the edge of its authority, stopped, wrote down what it recommends and why, and handed a named human a decision with the options already costed. What matters is how long the human takes and whether the work went back to the agent afterwards.
Health
What has been handed to a person
Escalations by tier and status, with how long the oldest have been waiting.
Open now
24
Acknowledged
11
Resolved
10
Returned to agent
11
human decided, agent resumed
Past SLA
35
the human is the bottleneck
Units behind them
6,548
blast radius if nothing moves
Litigation & Disputes: LG7A-1004 held at counsel instruct
Matter at counsel instruct valued at $13k stopped on a novel pattern. The lane cleared 54 percent of 5697 units this week without a human; this one did not clear because the evidence does not carry the decision on its own.
waiting on a person
10h 40m
SLA 8h 18m · breached
What the agent recommends
VP Litigation recommends rejecting and returning to the requester with the gap named.
Confidence 52 percent. The comparable population is 670 units; 84 percent of them cleared on the same pattern. What separates this one is a jurisdiction the policy version does not cover.
Options put in front of the human, already costed
Employment Legal: LG8B-1006 held at jurisdiction test
Employment matter at jurisdiction test valued at $15k stopped on a novel pattern. The lane cleared 56 percent of 4773 units this week without a human; this one did not clear because the evidence does not carry the decision on its own.
waiting on a person
7h 29m
SLA 44m · breached
What the agent recommends
Termination Risk Agent recommends pausing the lane until the upstream data defect is fixed.
Confidence 94 percent. The comparable population is 549 units; 86 percent of them cleared on the same pattern. What separates this one is a jurisdiction the policy version does not cover.
Options put in front of the human, already costed
Network Design: SC10A-1006 held at draft position
Network study at draft position valued at $15.7m stopped on a novel pattern. The lane cleared 24 percent of 3 units this week without a human; this one did not clear because the evidence does not carry the decision on its own.
waiting on a person
8h 8m
SLA 1d 11h
What the agent recommends
Carbon Modeling Agent recommends approving a partial and holding the remainder.
Confidence 88 percent. The comparable population is 806 units; 92 percent of them cleared on the same pattern. What separates this one is a conflicting record in the master data.
Options put in front of the human, already costed
Returned with a corrected policy version. The lane re-ran it without a human.
Investor Relations Support: FI9C-1005 held at frame
Disclosure item at frame valued at $3m stopped on a segregation of duties. The lane cleared 20 percent of 4 units this week without a human; this one did not clear because the evidence does not carry the decision on its own.
waiting on a person
13h 6m
SLA 23h 42m
What the agent recommends
Consensus Tracking Agent recommends pausing the lane until the upstream data defect is fixed.
Confidence 80 percent. The comparable population is 102 units; 62 percent of them cleared on the same pattern. What separates this one is a jurisdiction the policy version does not cover.
Options put in front of the human, already costed
Approved with a compensating control after 166 minutes. Rationale and clause recorded on the unit.
Environment Health & Safety: AD8C-1001 held at challenge
Safety case at challenge valued at $11.4m stopped on a novel pattern. The lane cleared 11 percent of 1 units this week without a human; this one did not clear because the evidence does not carry the decision on its own.
waiting on a person
20h 28m
SLA 1d 16h
What the agent recommends
Regulatory Reporting Agent recommends approving with the compensating control attached.
Confidence 97 percent. The comparable population is 281 units; 79 percent of them cleared on the same pattern. What separates this one is a value above the envelope.
Options put in front of the human, already costed
Returned with a corrected policy version. The lane re-ran it without a human.
The number worth watching on this page is not how many escalations there are. It is how many are past SLA. A queue of well-formed escalations sitting on a human's desk is the most common way an agentic operating model quietly stops delivering: the agents got faster, the approval layer did not, and the cycle time moved from the process to the person.
Actions
What is waiting on a person
An escalation is an agent saying it will not decide this. Nothing on this list moves until a person answers it.
Pick up an unacknowledged escalation
Raised, nobody has claimed it.
Answer a tier-three escalation
The highest tier on the estate, still unresolved.
Conclude an acknowledged escalation
Somebody owns it. It has not been closed out.
Clear an escalation older than a day
Age measured from the moment the agent stopped.
Operations
What this desk is allowed to start
A surface that only reports is not operable. This is the work this page can set in motion, and the bound it runs into.
Trigger and bound
This desk can start an escalation, route it to a tier, and record the answer it receives. It cannot decide the underlying question, extend an SLA, or close an escalation on the raising agent’s behalf. Returning one to the agent is itself a recorded decision with a named author.
Live observability
What the record shows right now
Where every escalation ever raised currently sits.
Current distribution
56 escalations
Is policy and strategy coming to fruition
Whether the written intent is holding here
10 of 56 escalations have been answered. 35 are still waiting on a person.
Not holding on the record
The design position is that an agent stopping is a healthy outcome, provided a person answers quickly. Measured: 17.9% have been answered, 9 have been open longer than a day, and 22 sit at the highest tier unresolved. The queue is the honest reading of human capacity on this estate — escalation volume is not the problem, answering time is.