How Should You Escalate Work From a Live AI Agent?

Escalate means the agent stops finishing the current item, packages what it saw, and routes that pack to a named human with a clear ask — instead of guessing the next sentence, inventing a field value, or burying a half-done draft in the same tray as routine Approves. Escalation is not pause, review, write-back, or stop. Pause is the trigger that a person must look. Review is Approve, Edit, or Hold on a playbook item. Write-back decides which fields may change after a decision. Stop parks the whole run. Escalate is for the moment the playbook runs out on this packet.

If your team cannot say what "escalate" attaches, who receives it, and when the agent is forbidden to improvise, the live agent is still guessing under a friendlier name. That is how money, promises, and VIP accounts get hurt while the dashboard still looks green.

Pro Logica demos on the AI agents solutions page show the same loop across Office, Field, Store, Law, CPA, Clinic, Dentist, Insurance, and Dealer: open the screen, follow the playbook, pause for a human. Escalation is the branch when the screen shows something the playbook never named. If the job needs a license, the agent stops. If the job is still inside the playbook, it belongs in review. Everything off-playbook needs an escalation pack — not a clever draft.

When is escalate the right move instead of finishing the draft?

Escalate when the agent hits a signal the playbook marked as out of bounds: a legal or compliance keyword, a VIP account flag, a dollar amount over threshold, a missing required field that cannot be inferred, a customer reply that changes scope, or any tool error that would force a guess to continue.

Do not escalate every missing comma. That recreates the old queue with louder labels. Do not finish the draft "for now" and hope review catches it. How Should You Review Work From a Live AI Agent? assumes the item was still inside the playbook. Escalation is the gate before a fake draft ever lands in that tray.

Also do not confuse escalate with When Should an AI Agent Pause for a Human? Pause is expected. Escalate is exceptional. A pause says "this class of step always needs a person." An escalate says "this specific packet left the map."

What belongs in an escalation pack?

Write the pack like a handoff a night-shift lead could act on without opening five tabs. Minimum contents: the trigger in one line, the source record id, screenshots or field snapshots the agent already had, what the agent attempted and explicitly did not do (no send, no price change, no delete), and the question for the human.

The question must be actionable: "Rewrite path A or B?" "Does counsel need this first?" "Is this VIP rule still in force?" Vague packs like "something weird" create delay and more guessing by humans.

Keep the pack short. If the agent dumps the entire CRM history, nobody reads it. If the agent attaches nothing, the human redoes the research. Aim for evidence the playbook already said was enough to decide.

This pairs with What Should an AI Agent Be Allowed to Write Back?: while escalated, write-back stays off except for a logged "escalated" status the allowlist already named. No silent field edits while waiting.

Who should receive the escalation — and how fast?

Route by rule, not by whoever happens to be online in Slack. Who Should Own an AI Agent After It Goes Live? names the accountable owner. Escalation routing is a table under that name: default ops owner, VIP path, money path, after-hours backup.

Speed belongs in the same table. High severity (legal language, money risk, regulated packet) pages or texts now. Medium can wait for the next review window. Low can sit as Hold with a due time. If everything is "urgent," nothing is.

Never route escalate to "the whole team" email alias as the only path. Shared inboxes bury severity. Named person, named backup, logged acknowledgment.

How is escalate different from Hold, Edit, or Stop?

Hold keeps a playbook item parked in the review tray because something is incomplete but still recognizable. Edit means a human will fix copy or an attachment on a draft that was mostly right. Escalate means the item should not be treated as a normal draft at all until a human chooses a path.

Stop is heavier. When Should You Stop a Live AI Agent? parks the whole run when Holds and escalations spike because the playbook or tools are wrong. One escalate is triage. A morning of escalations on the same trigger is a stop signal, then a playbook fix.

Do not invent a fourth casual outcome like "send anyway and note it." If you cannot name Escalate, Hold, Edit, Approve, or Stop, you are improvising.

What should the agent be forbidden to do while escalated?

No outbound send. No price, discount, identity, or delete writes. No "helpful" paraphrase of legal or medical language. No merging records to "clean up" the exception. The agent may refresh status to Escalated, attach the pack, and notify the route — nothing that creates a customer-facing promise.

That matches the first-job discipline in What Work Should an AI Agent Handle First?: smallest repeatable job, known steps, known gates. Escalation is a gate, not a creative mode.

If people keep asking the agent to "just take a shot" after escalate, the culture is the risk. Put the forbid list next to the playbook where the night shift can see it.

How do you measure whether escalations are healthy?

Track count, time-to-ack, time-to-resolution, repeat triggers, and how often escalations convert into playbook updates. How Do You Measure Whether an AI Agent Is Working? needs those numbers. A falling escalate rate with rising customer complaints means people stopped escalating and started guessing again.

Watch for theater: escalations that auto-close, packs with empty evidence, or owners who Always Approve from their phone without opening the pack. Theater here is as dangerous as fake review Approves.

Do not expand to a second job while escalate routing is still "whoever is free" and packs are screenshots with no question. Stable escalation is part of live, not a nice-to-have.

Human accountability for AI systems is a core theme in the NIST AI Risk Management Framework. For a shop that just went live, that maps to named routes, written severity, and a forbid list while escalated — not a slide that says "humans in the loop."

How should you stand up escalation this week?

Sit with the person who already handles weird cases. For one live job, write: the escalate triggers, the pack checklist, the routing table, severity SLAs, and the forbid list. Run it by hand for several days even if the agent only flags candidates.

Log each escalate with trigger and outcome. If the same trigger repeats, fix inputs or update the playbook. If humans keep asking for more evidence, add it to the pack once — do not make every owner reinvent the handoff.

AI agent development work at Pro Logica scopes structured execution, tool boundaries, and control loops so escalate is a named path — not a chat after the fact. For broader production patterns, see AI systems. If you need an engineer beside the team while those routes and packs take shape, that is closer to forward-deployed AI engineering.

Watch the nine-trade demos on What is an AI agent and when should a business build one if you need the loop in motion. The trade changes. The escalate rule does not: triggers written down, evidence pack, named route, forbid list, no guessing.

If you want help designing escalation for a live agent — triggers, packs, routing, and what must never ship while waiting — book a call. Bring one real off-playbook packet, the current owner name, and the list of times someone guessed anyway.

What should you read next if this issue sounds familiar?

If this topic matches what your team is dealing with, these pages are the best next step inside Prologica's site.

• How Should You Review Work From a Live AI Agent? for the tray that handles playbook items.
• When Should You Stop a Live AI Agent? for when escalations mean the whole run should park.
• Forward Deployed Engineer (FDE) Services for delivery context.