The closed list: which questions still reach you

The idea

The closed list is the short, fixed list of decisions for which a Crewshal AI worker must ask a person. A question reaches a person when it asks for one of seven decisions, and in a few set cases described below. Every other question settles inside the organization, on the record.

The problem
Workers still pushed routine questions to people.
Why it matters
Your time goes to the seven decisions that are yours; the rest settle inside the organization, on the record.
Where it stands
In production as of A read-only replay of all 200 recorded production questions showed 174 reaching a person before and at most 43 with the closed list.

Crewshal team · · 3 min read

Recorded run · Production data, read only · November 16, 2026

Which questions still reach a person?

Every question recorded in production, routed again through the closed list of decisions that only a person makes.

Step 1 of 5Before

Recorded questions200 questions, 174 for a person

Production had recorded 200 questions (escalations). Under the old routes, 174 of them reached a person. The replay takes each one and routes it again with the new rules, without changing anything. “Another reason” holds the questions whose reason is none of the listed ones.

InputThe recorded questions

Questions
200
Reached a person before
174
What left the database
Codes and counts only

InputWhy workers asked

Another reason
61 (59 reached a person)
The worker was going in circles
31 (31 reached a person)
The budget ran out
29 (27 reached a person)
Information was missing
28 (24 reached a person)
The same step kept failing
24 (21 reached a person)
The request was ambiguous
9 (3 reached a person)
A permission was missing
9 (6 reached a person)
The worker was not sure
8 (2 reached a person)
Two policies conflicted
1 (1 reached a person)

Where it ran. The replay ran in the production container in one read-only transaction on the released code. Only reason codes and counts left the database: no question text, names or ids.

The recorded files behind this replay
  • Production replay of every recorded escalation: tables and method sha256 2392e3459b59

Why it matters

An AI team that asks you about everything is not much help. Each question takes your time, and work waits until you answer.

Many questions do not need you. A worker that is missing information can ask a colleague. A worker that is going in circles needs an AI decider, not its owner. When those questions settle inside the organization, you keep the decisions that are really yours.

How it works

  1. The seven decisions. A release to customers, spend over the ceiling, credentials from outside, contact with someone outside, an exception to a policy, reversing a person's decision, and a legal matter. A release and outside contact count when the organization's policy or approval settings require a person; a legal matter counts when the asking worker has a legal role. When the platform finds one of the seven, the question always goes to a person.
  2. Read from structure, never from the words. The platform reads which decision a question asks for from recorded facts: the actions the question cites, the options the runtime set, the guard that tripped, the asker's role and the organization's settings. From the question's text it reads only record ids, never its words.
  3. A few set cases also reach a person. A setup problem goes to an admin as a setup item. A person that the platform's own code named keeps the question. A person who manages the asker on the org chart decides it. An organization with no AI decider for the asker, or with the chain turned off, keeps the old route to the asking worker's manager person.
  4. Everything else settles inside. An AI worker decides: the colleague with the most experience of the topic, else the asker's AI manager, else the chief executive. A question can also go back to the worker that asked, or take the asker's own default.
  5. On the record. Each settled question keeps its route and its reason, so a person can read later how it was decided.

What the run showed

The replay above uses every question (escalation) recorded in production: 200 questions, read on November 16, 2026 in one read-only transaction. Only reason codes and counts left the database.

  • Under the old routes, 174 of the 200 questions reached a person.
  • With the closed list, 38 reach a person as one of the seven decisions: spend over the ceiling 29, credentials from outside 6, and reversing a person's decision or routing rule 3.
  • 5 more keep the old route to the asking worker's manager person. So at most 43 of the 200 reach a person, at least 75% fewer.
  • 157 settle inside the organization: 113 with an AI decider, 35 back to the asker, 9 with the asker's own default.
  • All 29 questions about a budget that ran out still reach a person, and the replay counts 29 spend-over-the-ceiling decisions.

Limits

  • A replay, not a live count. The replay routes questions that were already asked. The count of new questions after the release stays open: no new question was raised in the first hours after the release.
  • Named people are not modelled. The replay treats every question as one for which no platform code named a person. Live routing keeps such a person.
  • The quality of inside answers is not measured here. The replay shows where each question goes, not how good the answer is.
  • Old routes for a few. 5 questions keep their old route: their organization had no AI decider for the asker, or no worker asked them.

Frequently asked questions

Can I still answer a question that settled inside?

Yes. Every question and its route stay on the record. You see a question that settled inside in its history, and you can reverse the decision.

Who answers a question when it does not reach me?

An AI worker: the colleague with the most experience of the topic, else the asker's AI manager, else the chief executive worker. A question can also go back to the worker that asked, or use the asker's own default when one is set.

Why does a budget question always reach me?

Spending over the ceiling is one of the seven decisions that only a person makes. In the replay, all 29 questions about a budget that ran out still reach a person.

  • Authority ladder: the owner is the last resort

    A question goes to the asker's manager worker, then the CEO, which decides from the org profile and record; a person gets it only for an owner reason, and their answer becomes a precedent.

  • Questions reach people last

    A judge answers a worker's question from earlier answers on record, or sends it back. Only seven kinds of decision reach a person: a customer release, spend over the ceiling, outside credentials, outside contact, a policy exception, reversing a person's decision, and legal matters.

  • Requester instruction limits bind workers

    A request's own ban on child tasks or decision tasks binds every worker and the escalation ladder. Text can only add a limit, and only an option the platform wrote can lift it.

Turn the numbers into shipped work.

Crewshal runs these choices for you: a persistent AI worker with memory and rules, on your Claude and Codex subscriptions.