Handle-Time Measurement System in a Claims Contact Center

Two metrics, two different verdicts - and the difference matters

• %GRR of 28% of tolerance sits in the marginal band (10–30%). Marginal, not acceptable.
• %study variation of 31% is above the 30% threshold and is normally judged unacceptable.

The %tolerance figure only looks better because the tolerance is generous relative to the process variation. But note the scale of the problem: the project target is 52 seconds, and the tolerance is 90 seconds. Measurement error of 28% of 90 seconds is about 25 seconds - half your intended improvement. You cannot reliably detect a 52-second shift with a gauge that scatters ±25 seconds.

So: do not proceed to baseline capability yet. But first identify which system the 28% belongs to. The Gage R&R was run on supervisors with a stopwatch, not on the ACD. The ACD is the system of record and its error is a data-integrity problem, not a gauge-repeatability problem.

Sequence

1. Write the operational definition of AHT: exact start event (call answered / agent connected) and stop event (after-call work submitted or disposition saved). The 30-second gap between audit and ACD is almost certainly a definition difference - hold time, after-call work, or transfer handling.
2. Reconcile 20 calls side by side, second by second, until the gap is fully explained.
3. Retire the stopwatch audit as a decision metric, or retrain supervisors against the agreed definition and re-run the study.
4. Then quantify ACD data integrity: missing wrap-up codes (3%) and double-counted transfers (5%). Approximately 8% of records are affected; that is not random noise at the level you are trying to detect.
5. Only after 1–4, collect baseline and compute capability against the 360-second target.

Stratification Site (2), team (4), call type or claim complexity band, shift, day of week, and agent tenure band. Handle time is almost always multimodal across call types - mixing a 90-second address change with a 900-second injury claim produces a meaningless mean and an inflated standard deviation. Until you stratify by call type, capability indices are not interpretable.

Sample size From the ACD, the standard deviation of AHT is roughly 110 seconds. For a 95% confidence interval on the mean with a half-width of 10 seconds: n = (1.96 × 110 / 10)² ≈ 465 calls. Round to ~500. That is a sample for estimating the mean. If you also need to detect a 52-second shift in a later hypothesis test, the required n per group is larger - roughly n = 2 × (1.96 + 0.84)² × 110² / 52² ≈ 70 per group at 80% power. Confirm with a power calculation once the baseline SD is known.

Rational subgrouping Sample in time order within team and shift - for example 5 consecutive calls per team per hour - so the chart reflects within-shift variation rather than mixing day and night. Twenty-five subgroups of 5 gives a usable baseline.

Transfers and missing codes

• Transfers (5%): decide whether AHT is first-contact handle time or total handle time across queues. Pick one, document it, and either de-duplicate by call ID or report transferred calls as a separate stratum. Do not silently average them in.
• Missing wrap-up codes (3%): test whether the missingness is random. If codes are missing more often on short calls or on one team, deleting those records biases the mean. Investigate the mechanism - likely a system timeout or an agent behaviour - and report the impact on the baseline mean before deciding how to handle them.

Coaching people against a number they do not trust is a predictable failure mode: they will optimize the metric rather than the process, and the project loses credibility before it starts.

Recommendation

1. Publish the definition before publishing the number. Circulate the start/stop events agreed in the reconciliation, with two worked examples of a call that includes hold and after-call work.
2. Show the reconciliation openly to team leads. Take 20 calls and walk through where the 30 seconds comes from. If the ACD is right and the stopwatch is wrong, the leads will accept it once they see it. If part of the gap is ACD behaviour (for example, wrap-up timer starting late), fix that and say so.
3. Give leads a data-quality view. Show that ~8% of records are affected by transfers and missing codes and describe how those are handled. Transparency converts a dispute into a joint problem.
4. Delay individual coaching on AHT until the measurement system is settled - but do not delay coaching on the behaviours the definition reveals, such as lengthy holds or unclosed dispositions.

Impact on Improve Your pilot comparison will be made on this metric. If measurement error is 25 seconds and the pilot is expected to deliver 20 seconds, the pilot will be inconclusive regardless of whether the change worked. Fixing the measurement system now determines whether you can detect the improvement at all. Also budget for re-running the Gage R&R after any change to the ACD configuration or the disposition screen - a system change in Improve can invalidate the measurement system you just validated.