Case Study · Research Initiative
Under Study Customer Experience · Contact Center

Project Threshold

Every contact center has an escalation policy. Almost none of them can prove the AI followed it.

The Problem

Deflection Metrics Reward the Wrong Ending

AI now fronts a large share of customer conversations, and the industry measures it by deflection — how many contacts were resolved without a human. That number rewards the AI for keeping the conversation, which is exactly the wrong incentive at the moment a customer needs a person.

So the failure is quiet. A frustrated, confused, or vulnerable customer stays in an automated loop that is scored as a success, because the one thing the system optimizes for is not handing them off. Nothing in the stack is accountable for the opposite obligation: recognizing when the AI must stop and yield to a human, and being able to show afterward that it did.

Our Approach

Make the Handoff a Governed Event

Project Threshold applies Corvion’s emotional-governance research to the escalation decision. It governs the AI’s obligation to yield — treating the handoff to a human as a governed, recorded event rather than a failure of containment.

Served, or Just Handled

Asks whether a customer was actually served, rather than whether the contact was contained.

Deflection Bounds

Limits that stop the system from optimizing containment past the point of a customer’s benefit.

Human-Required Conditions

Defined conditions under which the AI must yield the conversation to a person.

Consistency Testing

Studying whether the yield behaves the same for customers who express the same need in different ways.

Governance, not customer scoring

Project Threshold governs when an AI system must hand a conversation to a human. It does not score, profile, or assess customers; it does not evaluate agent performance; and it does not make eligibility, credit, or entitlement decisions of any kind.

Methodology — What We’re Studying

The Open Research Questions

  • Which signals reliably indicate that a conversation has passed the point where automation should continue.
  • How to define a human-required condition precisely enough to be governed, without escalating everything and erasing the value of automation.
  • Whether the yield decision behaves consistently across accent, phrasing, disability, and emotional expression.
  • What a handoff record must contain to satisfy a conduct or complaints review months later.
Open, By Design

What We’re Working On

This is a live research direction at Corvion, run in the open by design. The work centers on a single question: when an AI must yield a conversation to a human, against every metric that rewards it for holding on.

Our approach is to define that question rigorously — the criteria, the failure modes, and what would count as an answer — before building anything to serve it. We publish findings here as the work develops, including the results that don’t hold up.

Get Involved

Partner on This Research

We’re looking for contact-center operators, CX platform teams, financial-services conduct and complaints functions, and accessibility researchers willing to test where the yield line belongs.

A downloadable research brief will live here as the study matures.