How to route borderline AI outputs to the right human (without burning out your editorial team)

A 200-location franchisor running an AI marketing swarm produces 12,000-25,000 candidate outputs per month at steady state. If even 20% land in a human review queue, that is 2,400-5,000 items requiring review — roughly 80-170 per day. No three-person marketing team can sustain that without rubber-stamping. The editorial queue becomes a graveyard. Either drift escapes to production or the team rebels and shuts the swarm off.

The fix is not "buy a better queue tool." The fix is queue design — a four-tier routing decision for every borderline output, role-based assignment, an explicit escalation tree, and a 24-hour SLA with safety-valve defaults. This guide walks through how to build that, including the volume math nobody runs before they buy and the bottleneck rules that keep the queue functional.

The volume math you need to run before designing anything

Start with the math. The queue volume that actually reaches a human determines whether your editorial team can adjudicate well or starts rubber-stamping. Volume targets:

• Auto-publish (Tier 1): 92-97% of all candidate outputs
• Light-touch human review (Tier 2): 3-7% of candidates
• Specialist routing (Tier 3): 1-3% of candidates
• Escalation (Tier 4): <0.5% of candidates

For a 200-location franchisor producing ~15,000 candidate outputs/month, that translates to: ~14,250 auto-publish (no human action), ~750 Tier 2 (single editorial coordinator, batched daily review, ~30-60 min/day), ~225 Tier 3 (specialist routing, 24-48h SLA), ~50 Tier 4 (executive decision, 4-24h SLA).

These targets are reachable for a 2-3 person editorial team. Anything above 10% of candidates routed to humans is a failure of the queue design, not a feature of conscientious oversight.

The four-tier queue, named

Tier 1 — Auto-publish queue (no human action required)

Routing criteria: aggregate brand-voice gate score ≥ surface-specific auto-publish threshold (typically 0.88-0.95); no anomaly flags from telemetry layer; no NAP-canonical changes (those are always Tier 3+); no outreach drafts (always at least Tier 2 — outreach has different volume + trust dynamics).

Action: publish after surface-specific delay window (5-30 minutes) — this is the human-intervention catch window where someone monitoring the queue in real time can override before publish. Telemetry logs the publish event. Volume target: 92-97% of all candidate outputs.

Tier 2 — Light-touch queue (single approver, batched review)

Routing criteria: aggregate gate score 0.75-0.90 (between auto-publish and regenerate thresholds); one borderline dimension flagged by the gate; outreach drafts within volume cap (always require approval); local content pieces with a new claim type the gate hasn't seen before.

Action: routed to a single editorial approver in a batched daily review interface. The reviewer sees a list, not a real-time stream. Three buttons per item: Approve | Edit + approve | Reject (send back). Volume target: 3-7% of candidate outputs. Daily review window: 30-60 min for a 200-location operation.

Tier 3 — Specialist routing (role-based)

Routing criteria: compliance-relevant content (healthcare, financial, cannabis disclaimers); NAP-canonical changes (any change to name/address/phone that propagates across 60-180 directories); brand-spec-deviation requests (producer wants to use a phrase the spec forbids; humans decide if the spec needs updating or the output needs rejecting); region-specific content where local market knowledge matters.

Action: routed to the specialist who owns that domain — not to a general queue. The compliance officer for category content. Operations lead for NAP. Brand director for spec-deviation. Regional marketing manager for local content where applicable. Volume target: 1-3% of candidate outputs. SLA: 24-48 hours.

Tier 4 — Escalation (executive decision)

Routing criteria: repeated regeneration failures (output failed gate twice — signals brand-spec ambiguity or producer config issue); high-stakes anomalies from telemetry (review-volume spike, sentiment cliff, GBP suspension warning); crisis-mode review responses (1-star with anger keywords); any output the Tier 3 specialist explicitly bumps up.

Action: routed to the franchisor's marketing director or VP. Decision required within 4-24 hours depending on severity. Decision artifacts feed back into the brand spec or producer config so the same escalation does not repeat. Volume target: <0.5% of candidate outputs. Aim for <3 escalations/day at 200-location scale.

The role-based routing matrix

Routing is not "everything to one queue" or "everyone sees everything." It is a per-role inbox driven by an assignment matrix the franchisor owns. The editorial coordinator owns the default Tier 2 queue for non-specialist content. The compliance officer owns Tier 3 for healthcare/cannabis/financial disclaimers + claims, and Tier 4 for compliance crises (e.g., FDA inquiry). The operations lead owns Tier 3 for NAP-canonical changes + vendor-relationship escalations, and Tier 4 for at-risk vendor accounts. The brand director owns Tier 3 for brand-spec deviation requests + new claim type approvals, and Tier 4 for brand crises (negative-trend press). The regional marketing manager (where applicable) owns Tier 3 for region-specific local content. The marketing director or VP owns all Tier 4 escalations.

The matrix lives in version control alongside the brand spec. Changing routing rules is an explicit, auditable action. Adding a new content category (e.g., a finance vertical from an acquisition) means adding a row to the matrix, not redesigning the queue.

The "approve, edit, reject" interaction model

Tier 2 reviewers see a batched daily interface, not a real-time stream. Per-item, the interface shows: the output rendered as it would publish; the gate's per-dimension scores + justifications; the borderline dimension(s) highlighted; the producer's prompt + context (collapsible — most reviewers won't need it); three primary buttons (Approve | Edit + approve | Reject); and an optional secondary action ("Tag for spec update") that flags the output as evidence the brand spec needs revision.

The "tag for spec update" path is the second learning loop. When a reviewer notices the gate fired on a perfectly fine output (false positive), tagging queues a brand-spec review meeting where the team decides whether the spec needs adjustment. This is how the spec evolves without ad-hoc edits that introduce drift.

The escalation tree — explicit, not vibes-based

Tier 4 escalations follow a fixed tree:

Anomaly / repeated failure detected
         ↓
Tier 3 specialist evaluates within their SLA
         ↓
   ┌─────┴─────┐
Resolved   Specialist bumps to Tier 4
   ↓             ↓
Logged      Marketing director / VP decision within 4-24h
   ↓             ↓
            Decision logged + artifact captured
                 ↓
   ┌─────────────┼─────────────┐
Spec update   Producer config   Process change
needed        needed            needed
   ↓             ↓                  ↓
Brand team    Engineering        Operations
PR + review   PR + review        playbook update

The 24-hour SLA + safety-valve defaults

Defaults per tier: Tier 2 (light-touch) auto-approves with audit flag — items here scored above 0.75 so they're plausible; aging beyond 24h means the queue is overloaded and the cost of holding them exceeds the cost of publishing with an audit trail. Tier 3 (specialist) holds + escalates to Tier 4 — specialist domains are too high-stakes for auto-approve. Tier 4 (executive) holds indefinitely + alerts weekly — executive decisions cannot be automated away; a weekly digest surfaces every aging Tier 4 item.

The 24-hour SLA forces the operations team to right-size the queue volume. If Tier 2 items keep auto-approving by default (because the team can't keep up), the gate thresholds need tightening — that's the signal. The default action is not the goal; it's the safety valve that makes the goal observable.

What this layer does NOT do

Scope clarity matters. Four things the editorial governance layer is explicitly not for:

• Replace editorial judgment. The humans in the queue are doing real work. The layer routes the right items to the right humans at the right cadence so judgment is applied where it actually matters.
• Eliminate human review. Aiming for 0% human review is the failure mode. 3-8% combined Tier 2 + Tier 3 is the target. Below 3%, the franchisor is flying blind on quality drift.
• Make the brand spec autonomous. Spec updates are human decisions, captured from the queue's "tag for spec update" signal but executed by humans through PR review.
• Handle legal sign-off for net-new claims. The compliance officer in Tier 3 holds the line; new claim types may require external legal review beyond what the agent layer surfaces. The governance layer routes; it does not adjudicate.

Validation: is the queue working?

Three signals to monitor weekly for the first 60 days:

1. Tier-distribution percentages. Plot the per-tier volume against targets. Tier 1 (auto-publish) should be 92-97%. If Tier 2 keeps creeping above 8%, gate thresholds are too tight; loosen. If Tier 1 is 99%+, gate thresholds are too loose; tighten.
2. 24-hour-default trigger rate. How often does the safety-valve fire? Should be <5% of Tier 2 items. Higher means the team can't keep up; either right-size the team OR tighten the gate to reduce Tier 2 inflow.
3. Tier 4 escalation rate. Should be <0.5% of candidates. Higher means upstream is broken (gate config OR producer config OR brand spec ambiguity); fix upstream before adding executive review capacity.

Cost expectations

For a 200-location operation with the volume targets above: editorial coordinator for Tier 2 batched review (1 FTE, 30-60 min/day) is an existing role with no incremental cost; specialist routing for Tier 3 uses existing roles (compliance officer, operations lead, brand director), each spending 2-4 hours/week on AI-output review; executive escalation for Tier 4 takes the marketing director/VP 1-3 hours/week reviewing escalations + decision artifacts.

What this gets you

A queue that scales with operation size without scaling reviewer count proportionally. An audit trail that logs every routing decision + every queue action against every output. Two learning loops (edit-capture + tag-for-spec-update) that compound the producer + spec quality over time. A 24-hour SLA + safety-valve defaults that prevent the queue from becoming a graveyard. Per-role inboxes that route the right items to the right humans without flooding any one role.