The AI escalation layer that saved us $3M a year

In my last post — the one about why chatbots fail compliance review — I made a claim in passing and then moved on: that our escalation layer "cut support costs by roughly $3M annually while keeping us audit-clean."

A few people wrote back and asked the obvious question: how? So here's the longer version, including the part I'm not proud of — the first design, which was obvious, and wrong.

The obvious version (that quietly bleeds money)

When you build an AI support system, the first instinct for escalation is simple: if the bot isn't sure, hand it to a human. Low confidence, route to a live agent. It feels responsible. It feels safe.

I shipped exactly that. And it's wrong in two directions at once.

It's wrong on cost, because "the bot is unsure" fires constantly. A human types a little differently, asks two things in one message, uses a word the model hasn't seen much — confidence dips, and you've just paid a person to handle something the bot could have answered with a citation. At a million interactions a month, a few percentage points of unnecessary escalations is a payroll line, not a rounding error. Cheap escalations are how AI products bleed money.

And it's wrong on trust, because confidence is not the same as stakes. A model can be extremely confident about something it has no business doing autonomously — moving money, changing a beneficiary, surfacing a medical record. Wrong-headed confidence is how AI products bleed trust. Routing on confidence alone gets both failure modes: it escalates the cheap stuff and lets through the dangerous stuff.

So the real problem isn't "is the bot sure?" It's "what is this request, who is asking, and how certain are we — and what does the combination of those three demand?"

The reframe: escalation is a routing decision, not a fallback

Once I stopped treating escalation as a safety net and started treating it as a routing decision on three independent axes, everything got clearer. Every incoming request gets scored on:

1. Intent risk — is this informational ("what are your hours?") or transactional ("move $4,000 to my other account")? Informational is cheap and reversible. Transactional is neither.
2. Identity strength — is this an anonymous visitor, an authenticated member, or a CIP-verified identity? You are allowed to do very different things for each.
3. Confidence — how sure is the model for this specific answer, with abstention as a first-class action. "I don't know, let me get someone" is a valid, encouraged output — not a failure.

The routing rules fall out of the combination, not any single axis:

• Low-stakes informational queries the bot answers itself, with a citation. No human, no matter how "unsure" the confidence score looks.
• Medium-stakes account questions go to the bot only when identity is verified and an audit trail is captured — otherwise they escalate.
• High-stakes actions — money movement, beneficiary changes, anything touching PHI — never run autonomously, regardless of how confident the model is. A verified human is in the loop by design, not by exception.

The magic isn't any one rule. It's that the same request can route three different ways depending on identity and stakes, so you stop paying humans for safe work and stop letting the model near dangerous work.

Field note: the single most valuable change was making abstention a first-class action instead of a failure state. The moment "I'm not sure, let me get a person" became a legitimate, well-instrumented output, the whole system got both cheaper and safer — because it no longer had to choose between guessing and escalating everything.

Where the $3M actually came from

The savings weren't one big lever. They were the compounding of a few:

• Fewer needless escalations. Confident, low-stakes answers stopped going to humans. That's the biggest line.
• Faster, cheaper good escalations. When something did reach a person, it arrived with full context and an audit trail, so the human spent seconds orienting instead of minutes.
• No expensive incidents. The high-stakes gate means the failure that costs you a HIPAA finding or an SEC reportable event — the one that dwarfs any payroll saving — simply can't happen through the bot.

Across a system handling more than a million interactions a month, that combination netted out to roughly $3M a year — while staying audit-clean, which for our customers is the whole point.

If you're building this

You don't need our stack to steal the idea. The shift is small and it's mostly a mindset:

1. Stop routing on confidence alone. Score intent risk and identity strength as separate, explicit signals.
2. Make abstention a real action the model is rewarded for, not punished for.
3. Gate high-stakes actions on a human, always — independent of confidence. This is the line you do not cross to save money.
4. Log the routing decision, not just the answer. Six months later, someone will ask why a specific request went where it did. Have the answer.

The naive "escalate when unsure" design will always be the first thing you reach for. It's intuitive, and it's a trap. Escalation isn't a fallback — it's the most important routing decision your AI makes, and it's where both your margin and your trust live.

Building something in this space? I lead product for AI platforms in regulated environments. Always happy to compare notes on escalation, identity, and the messy real-world version of "just add a human."