Last updated: September 12, 2026
How average handle time is calculated
Average handle time is total talk time plus total hold time plus total after call work, divided by the calls handled in the same window. After call work is the piece most often left out, and dropping it can understate the real figure by about a fifth. What comes out is a capacity number rather than a quality one.
Operations watches it because payroll is the largest line in a contact center, and handle time is what turns headcount into capacity. Quality distrusts it for a plain reason. The fastest way to bring the number down is to stop doing the things a scorecard gives points for.
Take the callback number read back digit by digit. On an answering service scorecard that read back is worth 8 points out of 100 and costs about twelve seconds. An agent under handle time pressure says the number once at speed, keeps the points still within reach, and moves the average the way their supervisor asked. Two dashboards then disagree about the same call and both are right.
The formula, and the piece that goes missing
Add the talk seconds, the hold seconds, and the after call work seconds for a period, then divide by the calls handled in that same period. Whatever your platform calls wrap, disposition, or post call processing belongs in the third bucket.
| Piece | What it covers | How it goes missing |
|---|---|---|
| Talk time | Live conversation, from answer to disconnect | Rarely. Every phone platform reports it |
| Hold time | Seconds the caller spends parked, consult holds included | Platforms reporting talk time only, and holds parked on a second leg |
| After call work | Message typing, dispatch to the on-call tech, CRM notes, disposition codes | Agents who never enter wrap, notes typed during the next call's ring, a batch at end of shift |
Say a queue handled 4,200 calls last month against 312 hours of talk, 41 hours of hold, and 78 hours of after call work. That is 431 hours over 4,200 calls, or about 6 minutes 9 seconds each. Leave the wrap out and the same queue reports 5 minutes 3 seconds, so roughly 18% of the real number disappears. Eighteen percent of a contact center's cost is not a rounding error.
Answering services call the same idea average work time, and the metrics they run on treat it that way, since dispatch and the message both land after the caller hangs up.
Per agent and per queue are two different numbers
The formula does not change. What changes is what the answer is good for.
A queue level number is the planning number. Erlang staffing math wants handle time for one queue answering one kind of call, and blending a billing queue with an outage queue produces a figure that describes neither.
An agent level number is a coaching input and a bad ranking. Routing decides what lands on whom, so whoever takes the escalations and the second language calls carries a longer number by design. Read an agent's handle time against the same call type, in the same period, at similar volume, or leave it alone.
What a target should be anchored to
Take your own baseline, per queue, per call type, measured over a stretch long enough to include a bad week, then set the target against that.
It reads as a dodge, and part of it is. The reason it holds anyway is that no published average handle time benchmark met the evidence bar when we went looking on 2026-09-12, and the figures circulating do not say which call types they cover. A number quoted without the call type behind it is decoration.
Your second anchor is the script. A call whose account instructions require a four question triage sequence cannot come in at thirty seconds, and asking for that produces a floor that skips the triage. Changing the target without changing the script mostly changes what gets skipped.
Comparing handle time across call types tells you nothing
An answering service operator picks up in a client's business name, takes a message under that account's script, and is done in under a minute. Most of those calls land closer to 45 seconds.
A Medicare enrollment call is a different object. The TPMO disclaimer has to be delivered before any discussion of plan benefits, a Scope of Appointment goes on the record, and the agent reads plan detail the caller has to understand well enough to choose. Ten minutes there is not waste.
A home services CSR gets judged on whether the job got booked, where the four minute call that books Thursday beats the ninety second call that did not. Blending those three accounts into one handle time is the same mistake as one blended QA average across them, which the QA metrics that describe your program covers from the quality side. Split the number the way you split the work.
How handle time gets gamed
Nobody has to decide to cheat. Each of these is a rational response to being measured.
| The move | What it does to the number | Where the cost lands |
|---|---|---|
| Parking the caller on hold where the platform reports talk time only | Removes the hold seconds | The caller's clock, which nobody is measuring |
| Transferring instead of resolving | Ends the first agent's handle time at the transfer | The second agent's number, and repeat contacts |
| Skipping wrap and typing notes during the next call's ring | Removes after call work entirely | Message accuracy on the call being answered |
| Rushing the close | Cuts the recap and the read back | The QA score, and a callback number nobody can dial |
| A hard call that drops at 40 seconds and comes back in | Turns one long call into two short ones | The caller, who often lands on a different agent |
| Promising a callback that never happens | Moves the work out of this window | Next week's volume |
Every one of those shows up plainly in the recordings, and they persist anyway because most floors review 1 to 3% of calls by hand, a range Verint, Gistly, and ViciStack all report. The calls a supervisor picks are rarely the ones where somebody cut the close short.
The honest case for watching it anyway
Handle time is how a schedule gets built. Seats and the answer time you promised a client both come out of volume multiplied by handle time, so refusing to measure it means guessing at payroll.
Movement in it is the useful part. A queue that gains 40 seconds in a week usually has a system that got slow, a script that changed on Monday, or a new account whose triage nobody has learned. None of those are agent problems, and all of them surface here first.
Set against repeat contacts, it also tells you whether your short calls are finished calls. What it will never tell you is whether the call was any good. A scorecard cannot fix a bad script, and handle time cannot tell you which of the two you have.
When handle time and quality fall together
A handle time drop is worth celebrating for about a day, until somebody checks whether the scores came with it.
Pull the shortest calls from the week that moved, the bottom slice of the distribution rather than a random sample, and score them against your existing scorecard. Real efficiency looks like short calls scoring about the same as everything else, while rushing looks like the score falling as the duration falls, with the points coming off capture and close. Repeat contacts from the same caller inside 48 hours are the second check, since work that got skipped tends to arrive again.
The pattern in the audio is specific. An agent resolving faster finishes the close and lets the caller answer. Somebody rushing it starts the closing line on top of the caller's last question, delivers the callback number once at speed instead of digit by digit, and hangs up without a confirmation, so the last twenty seconds carry overlapping speech and no read back. That overlap at the end is the most reliable tell in the recording.
Criteria like that one score reliably because they are rule based, since whether the number was read back is a yes or a no. The read on whether somebody sounded rushed holds up less well, which Aircall and StackAI both note, and any criterion resting on who was speaking needs audio where the two voices can be separated. A call review scores a batch of your own recent recordings against your own scorecard, so the quality side of that week sits next to the handle time drop.
Frequently asked questions
What does average handle time mean?
The average total time an agent spends on one contact, counting the conversation, any hold, and the work done after the caller hangs up. It measures what a call costs to handle and says nothing about how well it went.
How do you calculate average handle time?
Add total talk time, total hold time, and total after call work for a period, then divide by the calls handled in that same period. Run it per queue or per call type, because a blended figure across different kinds of calls describes none of them.
Does average handle time include after call work?
Yes. Wrap, disposition coding, message typing, and dispatch all belong inside the number, and leaving them out hides about 18% of the real figure on the worked example above.
What is a good average handle time?
No published benchmark met the evidence bar on 2026-09-12, and a figure quoted without its call type is not usable. Build the target from your own baseline per queue, then check it against what the account script requires an agent to do.
Should average handle time be on a QA scorecard?
Keep it off. Handle time penalizes the agent who reads a callback number back digit by digit, which is the behavior the capture criterion exists to produce, so report both numbers side by side and let them disagree in public.
Sources
- Verint, Gistly, and ViciStack on manual QA reaching 1 to 3% of calls. Third party, unverified, setup only.
- Aircall and StackAI on rule based criteria scoring more reliably than tone. Third party, unverified.
- CMS TPMO disclaimer timing and Scope of Appointment requirements. Verified from search on 2026-09-08, recheck against the current CFR text before publish.
- Sub 60 second answering service calls and the 8 point read back criterion. Trinsic call notes and the CallSniper scorecard set, 2026-08-27 and 2026-09-03. Verified internally.
- No published average handle time benchmark met the evidence bar on 2026-09-12, so none is quoted here.
- The 4,200 call worked example, the gaming table, and the target anchoring guidance. Editorial judgment.