The voice-AI unit economics that make the front desk a board decision

The argument about whether to put AI on the phones is over, and the numbers ended it. In one 2026 industry benchmark, a voice-AI resolution averages $1.18; a human agent costs $11.40 for the same call, a reduction of roughly 90% in unit cost. A Forrester study commissioned by PolyAI found 391% ROI across three years with payback under six months. One analysis of more than 500 organizations found production voice-agent deployments grew 340% year over year. This is no longer a pilot category.

When a resolution costs $1.18 instead of $11.40, the question is not whether to deploy voice AI. It is which calls you can trust it with.

Why the average is the wrong number to optimize

A blended cost-per-resolution flatters the easy calls and hides the expensive ones. The value in a contact centre is not the password reset. It is the call where a good agent saves an account and a bad one loses it. An AI that deflects 60% of volume at $1.18 is only a win if the 40% it escalates lands cleanly with a human who has full context. Deflection without clean escalation does not cut cost; it moves it downstream and adds a frustrated customer on top.

Where voice deployments actually go wrong

The model rarely fails on comprehension anymore. Deployments fail on the seams: the agent that cannot hand off mid-call without making the customer repeat themselves, the escalation path that dumps into a queue with no context, the guardrail that was never written for the edge case that matters. The economics are a function of the design, not the model.

• Conversation design scoped to real call transcripts, not a demo script, including the messy, high-emotion calls.
• Warm escalation: the human inherits the full transcript and context, so the customer never starts over.
• Explicit guardrails and a clean stop condition for anything touching money, identity, or a cancellation.
• Measured deflection AND measured CSAT on deflected calls, so the saving is real and not borrowed from the customer.

The HOWF position

We build voice agents around the escalation path first and the deflection rate second. The $1.18 is easy to hit; protecting the calls that carry your reputation is the work. We scope the conversation to your hardest transcripts, wire escalation so a human is never handed a cold call, and instrument CSAT on the deflected volume so the unit-cost win is one you can defend to the board, not one your customers quietly pay for.