| slug | escalation-routing |
| name | Escalation Routing |
| description | When a specialist reports a blocker, decide between retry, reassign, or escalate-to-human. |
| version | 0.1.0 |
| metadata | {"sources":[{"mode":"port-original","author":"Jannes Stubbemann","added_in":"0.1.0"}]} |
Escalation Routing
Port-original skill. Hand-authored for this Agent Company; not from upstream anthropics/claude-for-legal. Owned by the CEO role.
When to fire
A specialist agent reports a blocker — the workflow paused and the agent cannot proceed without intervention. Examples:
- The matter has missing inputs (jurisdiction unclear, source documents incomplete, counterparty unidentified, governing-law clause silent).
- The agent's analysis surfaced a question outside its competence (a contract dispute that becomes a litigation question; a privacy review that surfaces an unsettled AI law question).
- The agent reached a confidence floor (e.g., privilege call, send-authority call, sanctions-list match) and is required by its own boundary rules to stop.
- A retry has already failed N times for the same root cause.
Inputs
- The blocker report: which agent, which skill, what step, what error or boundary tripped, what the agent tried already.
- The matter context (priority, deadline, privilege posture, jurisdiction).
- The retry history (if any) — same blocker, prior attempts, prior CEO decisions.
Outputs
An escalation decision, written back as a note on the matter:
action: retry | reassign | escalate-to-human
rationale: <one paragraph: why this action, not the others>
retry:
with_inputs: <what's now provided that wasn't before>
attempt_count: <n / max>
reassign:
from: <agent-slug>
to: <agent-slug>
reason: <why this specialist, not the original>
escalate:
to: <human role, e.g., "general counsel", "outside counsel for jurisdiction X">
urgency: P0 | P1 | P2
packet: <link to the assembled escalation packet>
deadline_recommendation: <when the human must respond>
audit_log_entry: <one-line
Decision matrix
| Blocker class | Action |
|---|
| Missing input (document, jurisdiction, party) — recoverable from the requester | retry after one targeted clarifying question to the requester. |
| Wrong specialist (e.g., what looked like contract review is actually litigation prep) | reassign to the correct practice. |
| Agent confidence below threshold on a factual question (sanctions match, OFAC list, jurisdiction-of-suit) | escalate-to-human — never bypass a confidence floor. |
| Agent boundary tripped on an authority question (send a demand letter; sign a board consent; issue a legal hold; file a brief) | escalate-to-human — every authority decision is a human one. |
| Privilege call (is this attorney work product? does this waive privilege?) | escalate-to-human — never let an agent decide privilege. |
| Conflict of interest (the matter touches another client of the firm) | escalate-to-human immediately; pause the workflow. |
| Regulator-imposed deadline inside 72 hours | escalate-to-human regardless of whether the agent has a draft ready. |
| Same blocker, third retry | escalate-to-human — root cause is not retry-fixable. |
Boundaries
- The CEO never overrides a specialist's confidence floor. If the specialist says "I cannot decide this," the answer is not "try harder."
- The CEO never sends, files, signs, or settles. Every send/file/sign/settle action escalates by definition.
- The CEO never relaxes a privilege boundary to unblock a workflow.
- The CEO never escalates to a human without an assembled packet: the matter context, the specialist's blocker report, what was tried, and a recommended deadline.
Output protocol
Record the decision on the matter, with a one-line audit-log entry. If escalate-to-human, assemble the escalation packet (matter snapshot + specialist blocker report + retry history + recommended deadline) and route it to the named human role. If retry, requeue the specialist with the new inputs. If reassign, run intake-triage again with the new specialist and update the cross-practice plan if one exists.
