Background Noise and Average Handle Time

Noise converts to handle time through four channels, all of them visible in call recordings:

• Repeat-request loops. "Sorry, could you say that again?" — in either direction. Each loop costs the question, the answer, and usually a confirmation. Inbound noise (the caller's side) drives more of these than most teams expect; that's the case for two-way noise cancellation.
• Verification re-asks. Names, digits, security answers heard through noise get read back and corrected. Failed verification is the expensive tail.
• Defensive holds. Agents park callers to escape a loud moment — a lawnmower outside a WFH window, a burst on the floor. The hold is noise-attributable AHT wearing a disguise.
• After-call rework. A misheard address caught later becomes a callback: a second call that exists only because the first one was noisy.

Reference floor: 200 seats × 40 handled calls per agent-day = 8,000 calls/day, fully loaded agent cost $18/hour, 260 production days a year. The one number you should argue with is the noise-attributable seconds per call, so we model three, and you should replace them with your own QA data (the A7 defect code in our metrics guide counts exactly this):

These figures were computed, not typed: the model is a 40-line Python script that ships with this site (tools/aht-model.py), and the table above is its verbatim output. Change the seat count, the wage, or the seconds, and it recomputes your floor.

The comparison line: the published list price for the dedicated software layer — Krisp CC Core at $15 per agent/month billed annually, verified 20 September 2026 — works out to $180 per seat per year. Against $312–$1,040 per seat of modeled noise cost, the software doesn't need to eliminate the problem to pay for itself; it needs to remove a little over half of the light scenario's cost ($180 of $312), and under a third of the moderate one ($180 of $624). That arithmetic is exactly the kind of claim you should verify on your own floor rather than take from us — which is why the pilot design in the ops guide comes before any annual contract.

Three ways this model can mislead you if used carelessly:

• The seconds are the whole argument. If your real noise-attributable time is 2 seconds per call, the light scenario overstates your problem threefold. Measure first: two weeks of A7/A2 tagging gives you a defensible per-call figure.
• Saved seconds aren't automatically saved dollars. Freed capacity becomes money only when erlang math turns it into fewer required seats or absorbed growth. On a small queue, 13 agent-hours a day may just soften service levels — valuable, but a different sentence in the business case.
• Noise isn't your only AHT problem. If systems are slow and scripts are bloated, noise might be the third-biggest tax on your handle time. The model prices one tax; it doesn't rank them.