Dashboard Truths
Your average handle time is probably lying to you.
Most contact centre dashboards lead with averages. They are easy to calculate, easy to explain, and often wrong about what is actually happening on the floor. Dashboard Truths is a short series of free guides to the numbers on a contact centre wallboard: what each one measures, and what it hides. Parts 1 to 3 are out now.
The answer is not one better number
Swap the average for the median and the ten calls above look calmer: a typical call of 6 minutes. But the 48-minute call is probably the most important one on the list. Somewhere inside it is a confused customer, a broken process or a system that failed. The median hides it completely.
The same thing happens at scale. In the synthetic month used throughout the guide, a system slowdown stretched a block of calls for three hours on one day. The median barely moved. The average rose, mixed in with every other cause. Only the 90th percentile showed the incident clearly.
Flatten for the typical view, but never hide the tail. Every tidy figure on a dashboard should sit beside one that shows what was tidied away.
What Part 1 covers
Ten short chapters, one worked month of data.
Mean, median, minimum and maximum
What each one measures, why contact centre time data leans to the right, and why the shortest and longest calls usually belong on a data quality report rather than a performance one.
Percentiles in plain language
“Nine in ten calls finished within 12 minutes” means something to an agent, a team leader and a board. Why P90 and P95 show the calls where customer experience breaks.
Same data, different answers
Spreadsheets, databases and BI platforms do not all calculate percentiles the same way. On ten calls, three common methods give answers from 9 to 44 minutes.
Trimming and winsorising
Two ways to flatten the extremes, how they differ, and the trap in both: they turn down the part of the data where problems show up first.
Small numbers, big swings
Why a new starter with eight calls can look like the fastest or slowest agent in the building, and what minimum volume to set before anyone is ranked.
Honest tiles and a checklist
A before-and-after dashboard built from the same data, labelling rules, and ten questions to ask of your own reporting.
Every chart and figure uses synthetic data: one invented month for a generic inbound contact centre, 12,906 calls across 40 agents, generated to have the shape real contact centre data has. No customer or organisational data has been used. The patterns are real; the numbers are not.
What Part 2 covers
Voice and sentiment metrics that mean something.
Speech analytics made voice measurable. Every call now arrives with a sentiment score, a talk-to-listen ratio, silence, overtalk and an audio-quality figure. They look precise. Most of them are estimates built on other estimates.
There is no standard sentiment score
Each vendor chooses its own scale, model and weighting, so a score only means something within one system, over a stable period.
Where the call ended
A single overall score blurs the agent’s best work: the frustrated caller who was turned around. Start and end, side by side, show it.
Hold, silence, talk and overtalk
Hold barely matters until it gets long. There is no right talk-to-listen ratio. And overtalk means different things depending on who is talking over whom.
When bad audio looks like an angry customer
Sentiment is scored from transcripts, and transcripts depend on the audio. One failing carrier route can read as an unhappy afternoon and send managers to the wrong agents.
Transfers, surveys and QA
Why transferred calls are more often followed by a repeat contact, why survey replies are not a random sample, and how automated and human QA work best together.
Honest voice tiles and a checklist
The same month of data shown as comfortable averages and as useful counts, labelling rules, and ten questions for your own voice dashboard.
What Part 3 covers
Choosing the right chart.
The numbers can be right and the chart still wrong. Most chart mistakes on contact centre dashboards come from a handful of habits: the default chart the tool offers, a truncated axis, a gauge where a number would do, two measures forced onto one chart.
Start with the question
The question a manager is asking picks the chart, not the other way round, and one well-known finding about how people read charts explains why some work better than others.
What a rolling average hides
Smoothing calms a noisy line, and calms the incidents with it. The worst day of the month can end up looking like a slightly firm week.
Bars start at zero
A cut-off axis can make three queues within 1.5 points of each other look like a crisis. Most often, the BI tool chose that axis automatically.
Show the spread, and when
Box plots show what a bar of medians hides, and a heatmap lays out the whole week at a glance for workforce questions.
Scatter plots for two measures
The league table that ranks agents on handle time alone can reward the ones whose fast calls bring customers straight back.
Pies, gauges, dual axes and spaghetti
When each of these misleads, what to use instead, the difference between a wallboard and an analysis dashboard, and a ten-question checklist.
Where it fits
A misleading dashboard is a foundation problem.
A four-hour call that was really a stuck session. A hundred one-second records that were never conversations. A bad day that only shows up if you know where to look. Those are not reporting quirks. They are data integrity findings, and data integrity is one of the three foundations in Foundation First.
If an organisation cannot trust its own handle time figures, it is not ready to measure whether a new platform or an AI deployment has made anything better.
Free guide
Dashboard Truths, as PDFs.
Choose the parts you want. Each is a short A4 guide, written for the people who run contact centres and the people who build their reports, using the same synthetic month of data. No formulas to learn and no platform to buy. Tell us where to send them and they’re with you in a minute.
Coming next in the series
One more part, the same month of data.
Part 4: Analysing with AI
What AI analysis gets right and wrong, tested against a dataset with known answers.