Maybe let Kevin check it before we embarrass ourselves.

The AI says: "I have reviewed 12,487 documents and reached a conclusion."

Kevin says: "you read the wrong PDF."

That entire exchange is Human-in-the-Loop. Fancy term. Boring practice. Saves your company.

The version of AI that ships in production looks nothing like the demo. The demo is a confident model with a microphone. Production is the model, plus Kevin, plus three layers of review, plus a logging table called flagged_for_review_pls_god. The companies you assume are pure-AI are almost always paying humans to clean up behind the scenes — and the ones that don't, eventually end up in court.

1. The "AI product" is almost always a human supervising a model.

Google doesn't run search on pure ranking math. They employ roughly 16,000 search quality raters through contractors like RaterLabs and Appen, scoring results against a 168-page document called the Search Quality Rater Guidelines ↗. In September 2025 they added the first concrete examples specifically for evaluating AI Overviews. The most successful search company on Earth still believes the right number of humans in the loop is sixteen thousand.

Amazon's "Just Walk Out" grocery tech — the one that was supposed to let you grab items and leave — relied on ~1,000 workers in India manually reviewing transactions. The target was 50 reviewed per 1,000 sales. Reality was closer to 700 per 1,000. The story broke in April 2024. Amazon killed the product shortly after.

2. Skip the human, meet the lawyer.

Air Canada, February 14, 2024. A bereavement-fare chatbot invented a refund policy that didn't exist. A customer relied on it, bought a ticket, and asked for the refund. Air Canada refused. The customer took them to BC's Civil Resolution Tribunal. Air Canada's defense — and this is real — was that the chatbot was "a separate legal entity responsible for its own actions." The tribunal disagreed. Air Canada paid C$812.02 and a credibility tax much larger than that.

Microsoft Tay, March 23, 2016. Shipped at 8am. Shut down 16 hours later after tweeting over 95,000 messages including Holocaust denial. The mechanism wasn't even sophisticated — trolls discovered Tay had a "repeat after me" function. There was no human moderation in the loop. Just a model and an open mic.

These are not edge cases. These are what happens when the loop is closed too tightly around the model and Kevin is on PTO.

// "Look how clean this is"
async function reply(msg: string) {
const out = await llm.generate(msg);
await db.respond(out);
return out; // ship it
}

// Failure mode: model invents
// a refund policy. You get sued.

// What's actually in production
async function reply(msg: string) {
const out = await llm.generate(msg);

if (isHighStakes(msg) ||
    confidence(out) < 0.85) {
  await queue.forReview(msg, out);
  return TEMPLATES.willGetBack;
}

await db.respond(out);
return out;
}

// Failure mode: Kevin is slow.
// Acceptable.

The math nobody puts on the deck

Before you decide HITL is "too expensive," do the actual math. Labeling vendors are public about pricing now.

3. Pick which decisions humans actually own.

The mistake teams make is binary: either "humans review everything" (expensive, slow, demoralizing) or "model decides everything" (Tay). The real design is gradient. Decisions get routed by what the model is sure of and what it costs to be wrong.

• Low stakes + high confidence → ship it. (Categorize support ticket.)
• Low stakes + low confidence → ship it but log. (Suggest a product.)
• High stakes + high confidence → human reviews after. (Auto-refund under $50.)
• High stakes + low confidence → human reviews before. (Refund over $500. Anything that creates a legal obligation.)

That's the entire framework. Two axes, four quadrants, one Kevin.

So what's the real shape?

Human-in-the-loop is not a fallback. It is not a temporary scaffolding you remove when the model "gets good enough." It is the architecture pattern that makes shipping a probabilistic system to humans legally and ethically tractable.

The teams who internalize this build slower, ship more, and end up on stage at conferences. The teams who don't end up as case studies in legal blogs.

The AI says it reviewed 12,487 documents. Kevin says it read the wrong PDF. Kevin is right. Pay Kevin.