Michael & Son Call copilot Figures as of November 16, 2026

We read 916 real calls. 13% of call time was the caller waiting while the CSR looked something up.

13%

of all call time was waiting on a lookup

    40 seconds
    of waiting in the average call
    656 of 916
    calls had a wait of 10 seconds or more
    11 hours
    of waiting in every 1,000 calls

    Calls QA chose to grade, August 18 to October 5, 2026. A wait is 10 seconds or more with the CSR silent.

    Three in four calls are about a visit, and most end with one booked.

    What the call was about calls
    How the call ended calls

    109 calls ended with nothing settled: the caller declined, would call back, or it was not resolved.

    The longest wait is most often for the schedule. The CSR had it at once less than half the time.

    What the longest wait on each call was for calls
    How often the CSR had what they needed at once times needed

    The schedule came at once 397 of 860 times it was needed. Another team's answer came at once only 14 of 149 times.

    About three in four calls QA marks down include a miss a panel could prevent.

    8,866
    calls graded by QA, July 13 to November 1, 2026, each with the evaluator's note
    4,650
    of those calls lost points
    About 3,430 of 4,650
    of those include a miss a panel could prevent
    The most common misses calls

    The most common are routine steps forgotten. If every such miss went away, the average QA score would rise from 94.2 to about 98 at most.

    The copilot sits beside the CSR and does the lookups and the required steps.

      On 20 calls nobody had opened, about 1 panel row in 3 was wrong. The target is 1 in 10.

      Wrong, on the 20 held-back calls
      6 of the 20 calls had no wrong row

      Tested by replaying real recorded calls the copilot was never built on. Two separate reviewers mark every row and step right or wrong.

      It is not ready for a CSR's screen yet. It is being rebuilt, and three fresh sets of 20 unopened calls are set aside for the next tests.

      Running it costs about 48 cents a call.

      30¢
      a call, for the panel
      18¢
      a call, to turn the audio into text
      ½¢
      a call, for the notes

      For an average call. The audio step is priced on the cloud service used in tests; running it on the company's own server is planned.

      Every call has about 40 seconds of waiting. Here is what saving part of it adds up to.

      40 seconds of waiting in the average call (611 minutes over 916 calls)

      10 of 40 seconds saved on each call

      3.5 hours
      of waiting saved a day
      104 hours
      of waiting saved in a 30-day month
      $150
      a day to run it on those calls, at about 48 cents a call

      Set both sliders to your own numbers. The share starts near 62%: on 410 of the 662 calls with a wait, the longest one was for something the copilot looks up itself. The saving has not been measured on a live call yet.

      A later step · the readers' judgment

      Later, a voice agent could take many calls whole. Readers judged 626 of 916 could have been.

      Calls a voice agent could have taken whole, by kind of the calls of that kind

      A judgment from reading each call, not a test. 78 more could have been taken with a hand-off to a person at the end, 199 partly, and 13 not at all.

      Three things are needed before the copilot can sit beside a CSR on a live call.

      1. Live call audio out of NICE CXoneNobody has looked into it yet.
      2. Answers to 38 open questionsFor customer service, about fees, job types and the required steps.
      3. A fresh test near 1 wrong row in 10On calls nobody has opened.

        How these figures were made

        • The 916 calls are calls QA chose to grade, not all calls. 5 more could not be read.
        • Each was read by an AI reader from its transcript, with every spoken number masked. No names or call details are on this page; only totals.
        • A wait is a gap of 10 seconds or more in the CSR's speech. What the CSR was looking for is the reader's judgment.
        • QA's misses were counted from the evaluators' notes and checked against two hand-read samples of 420 and 160 notes.
        On 12 calls read twice, readers agreed on

        Readers agree well on kinds and outcomes, and much less on how much a copilot would help.