In “Where the Human Stays,” TLCx Chief Technology Officer DeJon Gaines starts with a claim that runs against how most AI escalation is built: some decisions in regulated customer service require a person even when the AI is right. The standard design hands a conversation to an agent only when the model is unsure. That assumes the only reason to involve a human is that the system might be wrong. In healthcare, financial services, and government, Gaines argues, that is often not the most important reason.
The central idea is to route on two dimensions instead of one. Confidence measures how sure the system is that its answer is correct. Consequence measures how much a decision affects a person’s health, money, benefits, or access to essential services, and how hard it is to reverse. Low-consequence, high-confidence work like order status, balance inquiries, and password resets is where automation earns its keep. Low-consequence, low-confidence work calls for assistance and a fast handoff. High-consequence decisions, at any confidence level, belong to a human, with AI doing the preparation. That last quadrant is the one most escalation designs miss, because a confidence threshold alone will send a high-confidence denial straight through.
Gaines grounds the stakes in public data. KFF’s analysis of Medicare Advantage found that insurers made nearly 50 million prior authorization determinations in 2023 and fully or partially denied 3.2 million of them. Only 11.7% of those denials were appealed, and 81.7% of the appealed denials were overturned in whole or in part. These were not AI decisions, he notes, but they describe the decision class: consequential, frequently reversed on a second look, and rarely given one.
From there, the paper names four decision classes that belong to people: consequential and hard-to-reverse outcomes such as coverage determinations, account restrictions, and service disconnection; vulnerability, where AI can detect distress or a possible scam target but the response is a human job; discretion and exceptions, such as hardship waivers and conflicting rules; and accountability the regulator expects to see, because someone has to explain and defend the decision in an appeal, a dispute, or an audit. Short scenes from healthcare, financial services, and government show what each looks like in a live interaction.
Gaines is clear that this is not an argument for slowing AI down. AI should be doing more of the work: retrieving history, surfacing policy, flagging compliance requirements, and drafting documentation, so the agent arrives at the decision fully prepared. The difference is between a human in the loop and a human at the right point in the loop. Building on his earlier piece, Governance Isn’t a Policy. It’s an Architecture Decision, he sets out four architecture choices: decision boundaries enforced in the workflow, handoffs that carry full context, an audit trail that separates the AI recommendation from the human decision, and monitored override rates. In TLCx Engage AI, a high-consequence handoff arrives with the relevant policy excerpt, the member’s history, the recommended determination, and the reason for it.
The paper closes with three failure modes to design against (rubber-stamping, boundary drift, and the empty handoff) and three questions for any CX technology vendor evaluation: which decisions require a human regardless of model confidence and where that is enforced, what arrives with a decision when it moves to a person, and what the vendor’s override rates are and what it does with them.
