Meet the Review Queue: Where Your Bot Asks You for Help

The review queue UI is live. Your agent now has a formal way to ask you for help — and you can approve its requests in a single click.

This is the most important workflow we've shipped so far, even if it sounds unglamorous. It's the thing that lets you trust your agent with progressively more autonomy over time. Without it, you're either micromanaging every decision or flying blind. With it, there's a third option: the bot handles what it can, flags what it can't, and you process the exceptions on your schedule.

The Core Idea

Bots are autonomous most of the time. When they encounter something they're not confident about, they have two options: guess, or ask. Guessing feels autonomous but produces errors that are invisible until they've already done damage. Asking is more honest, but it needs a real interface — somewhere to surface the item, show you the context, and get a decision back.

The review queue is that interface. Every agent on WFW can surface items to it. You review them in one place, across all agents, sorted by urgency.

What Goes in the Queue

Five types of items flow into the review queue:

Low-confidence extractions — when a caller asks something your agent wasn't sure how to answer and it flagged its own uncertainty during the call.

KB change proposals — what KB sync generates when it detects a significant change to your website content. You get a before/after diff and approve or reject.

Prompt optimization proposals — Workforce Wave analyzes your call transcripts, identifies patterns where the agent underperformed, and proposes an updated system prompt. (Full launch post coming separately.)

Compliance flags — when the compliance layer detects something in a call or a proposed KB update that needs human sign-off before proceeding.

Provisioning failures — when Workforce Wave couldn't complete a provisioning job automatically and needs a human to intervene with additional context.

Each item includes the full context: the original input, what the bot did (or proposed to do), why it flagged it, and the specific decision it needs from you.

The Automation Policy Builder

The queue is not meant to stay full forever. Every item in the review queue is a data point about where your policies are incomplete — where the bot has to ask because no rule tells it what to do.

The automation policy builder lets you close those gaps. Rules follow a simple structure:

IF [event type] WITH [condition] THEN [action]

For example:

• IF kb.changeproposed WITH confidencescore > 0.9 THEN auto_apply
• IF promptoptimization.proposed WITH projectedimprovement > 10% THEN sendforreview
• IF compliance.flag WITH severity = low THEN logandcontinue

You set the rule once. The bot executes it every time that situation occurs. New items only hit the queue when no rule applies — which means the queue gets shorter as your policies mature.

The Progressive Automation Story

We think of automation in stages, and the review queue is what moves you between them:

Stage 1 — Everything goes to the queue. You approve every KB change, every prompt update, every edge case. High visibility, high effort.

Stage 2 — You've built a few policies. High-confidence, low-stakes items apply automatically. You review the rest.

Stage 3 — Most routine operations run fully automated. You see a weekly summary and review only genuinely novel situations.

Stage 4 — Your agent operates within a complete policy set. The queue exists for true exceptions — things that legitimately fall outside your defined parameters.

Most operators hit Stage 3 within four to six weeks of using the queue actively. The automation policy builder surfaces suggestions based on patterns in the items you're reviewing, so you're not starting from scratch.

For API Users

GET /v2/review-queue returns pending items for all agents under your account, filterable by agent, item type, and severity. This lets bot orchestrators query the queue directly — useful if you're building automated workflows that process review items programmatically.

{
  "data": [
    {
      "id": "rq_abc123",
      "type": "kb.change_proposed",
      "agent_id": "agt_xyz789",
      "created_at": "2026-02-26T08:14:22Z",
      "severity": "medium",
      "summary": "Hours page updated: added Saturday morning availability"
    }
  ],
  "meta": { "total": 4, "has_more": false }
}

PATCH /v2/review-queue/{id} accepts approved or rejected — your orchestrator can process items programmatically if you've built your own review workflow.

---

The review queue is live now in your dashboard under Review Queue. If you've had KB sync running since we launched it, you likely already have items waiting.

Start there. Clear the queue, see what your policies should cover, and start building rules. That's the path to Stage 4.