Average Handle Time — The Metric That Gets Worse When Automation Works (2026)

Quick answer: AHT is talk time plus hold time plus after-call work, divided by the number of contacts handled. It is a long-standing operational number in the contact center and it is one you will often see quoted with no source attached. Two things on this page are worth more than the definition. The first is that we followed one widely repeated "industry standard AHT" citation back through its chain and it dead-ended in an uncited vendor blog post, while the only figure we found with a named measurement population attached is roughly double the number the blogs call standard. The second is arithmetic: when a bot successfully absorbs your simplest contacts, AHT goes up, because the short contacts are the ones that left. In the worked example below, total human handling time falls by about a quarter while AHT rises 27.6 percent and no category performs worse than it did.

The formula, and the four things it hides

The standard form is uncontroversial:

AHT = (total talk time + total hold time + total after-call work) ÷ contacts handled

After-call work is the wrap-up: notes, disposition codes, the refund actually being keyed in. It is invisible to the customer and it is routinely a large and under-measured component of the number, which is why the first choice below matters as much as it does.

What varies between two organizations quoting AHT at each other is not the formula. It is four choices underneath it, none of which travels with the figure.

The practical rule that follows is the same one our customer effort score entry reaches about a different metric: a figure with no instrument attached is not a figure. Before comparing your AHT to anyone's, write down your four answers. A reporting dashboard has already answered them on your behalf, and in our experience it will not volunteer which way.

One more distinction, because it is the one readers most often collapse. AHT is not first response time, which is how long the customer waited before anyone replied, and it is not first contact resolution, which is whether the problem went away. AHT is occupancy. A contact can be fast, answered instantly and completely unresolved, and AHT will look excellent.

We tried to find the industry standard and could not

Search for an AHT benchmark and you will land on a figure around six minutes, usually phrased as the industry standard. We picked one chain that recurs often in those results and followed it on 21 August 2026.

The starting point was Scorebuddy's guide to AHT, first posted 10 February 2021 and marked updated 30 July 2025. Scorebuddy sells contact-center quality-assurance software. The page states a global industry benchmark of six minutes and three seconds and attributes it, in the sentence, to Call Centre Magazine. The words "Call Centre Magazine" are a hyperlink. The hyperlink goes to a Zendesk blog post.

We read that Zendesk post the same day. It carries a published date of 5 June 2026 and a machine-readable modified date of 17 June 2026, and it does not contain the figure six minutes and three seconds, does not mention Call Centre Magazine, and does not cite any source at all. What it does say is that "a good average handle time is around six minutes", unattributed, followed by an eight-row table of per-industry figures introduced as "estimates of standard industry benchmarks" with no methodology, sample or date. The byline is a Senior Director of Product Marketing, and every link in the body goes to another Zendesk page.

There is a detail on both pages that is funnier and more useful than the missing citation. Each one publishes a worked example, and neither example lands anywhere near the benchmark the same page endorses: Scorebuddy's arithmetic produces 15 minutes, Zendesk's produces 28. Both pages then carry on recommending six. Scorebuddy's example is about two and a half times the benchmark on its own page; Zendesk's is nearly five times its own. If the illustrations the authors chose themselves miss the standard they print by that much, the standard is not doing any work.

We are not accusing Scorebuddy of misquoting anyone. The Zendesk post is dated five years after that guide was first published, so the target of the link has plainly been rewritten since it was placed. That is the point. This is what happens to an AHT citation when you try to run it down: the destination changes underneath the claim and the claim stays put.

The same guide's second citation is a real one, and it is worth looking at properly, because it is the source most often meant when a page says "according to Cornell." It links The Global Call Center Report: International Perspectives on Management and Employment, by David Holman, Rosemary Batt and Ursula Holtgrewe, copyright 2007, an academic study of employment and management practices covering almost 2,500 centers in 17 countries and 475,000 employees, with the research conducted between 2003 and 2006. The per-sector figures quoted from it in the wild run from 282 to 528 seconds.

We read the report. Two things in its own prose are hard to reconcile with the way it gets cited. It states that "the typical worksite in this report has an average call handling time of 190 seconds, or 3 minutes and 10 seconds," and adds that there is "surprisingly little variation in this number across the wide range of countries in this study." Its executive summary says calls "typically last from 3-4 minutes," and separately reports subcontractor and in-house medians three seconds apart, at 3 minutes 17 and 3 minutes 20. That second pair is a different axis from the cross-country claim, and we cite it only as a second place the report lands close to three minutes. We should be precise about the limit of our own check: we read the report's extracted text and did not locate the per-sector figures in it, so they may sit in a table our extraction dropped rather than being absent. What we can say is that the headline number the report itself prints is 190 seconds, barely over half the six-minute standard it is used to support, and that it describes a voice-only sector, surveyed before the smartphone. The report notes that the overwhelming majority of centers in it operated as voice-only rather than multi-channel.

That last row is the one we would actually use, and it comes from a source already cited on this site. Our first contact resolution entry carries SQM Group's benchmark reporting, and SQM's 2024 report puts average handle time at 697 seconds, which SQM describes as an 18 percent increase on the prior year. Its 2022 report prints 589 seconds; 697 against 589 is a rise of 18.3 percent, so the two reconcile, but SQM does not itself pair them and we are not asserting a two-year series from it. We are carrying both numbers from our own entry's source citation rather than re-reading SQM today, and that entry also warns that the two reports' data years may not sit where their labels suggest. With those caveats stated, note what the figure does to the story: the one number here with a named measurement population is roughly double the one the blogs call standard. The gap is measurement population, not performance. SQM measures inbound voice at large North American call centers. The six-minute figure measures nothing anyone has named.

Why a working bot pushes AHT up

Here is the part that is specific to automation. It is our own arithmetic, and it is worked in full so you can check it rather than take it.

Take a small support operation handling 1,000 contacts a month across three request types.

Now deploy a bot that does what bots are good at. It absorbs 70 percent of the lookups and 20 percent of the general inquiries. It touches none of the complaints, which is the boundary our guide on when not to use a chatbot argues for at length. Every remaining contact is handled exactly as well as before.

Three results, and they point in different directions:

• Contacts fall 41 percent (1,000 to 590).
• Total human handling time falls 24.7 percent (5,700 to 4,290 minutes, a saving of 1,410 minutes or 23.5 hours a month).
• Blended AHT rises 27.6 percent (5.70 to 7.27 minutes), with no category performing worse than before.

Cost behaves the same way. Put a fully loaded agent cost of $36 an hour on those minutes — an illustrative round number, not a benchmark, and not drawn from any dataset we hold. The before picture is $3,420 a month, or $3.42 per contact. The after picture is $2,574 a month, or $4.36 per contact. Total agent labor cost is down about 25 percent and cost per contact is up 27.6 percent, from one deployment, at the same moment.

Two things are deliberately missing from that. The bot has its own cost, on the other side of the ledger, and none of these figures contain it. And we are holding agent occupancy at 100 percent to keep the arithmetic legible; our bot versus human cost calculator divides by occupancy instead, which raises both per-contact figures without changing the direction of either.

The mechanism is dull and inescapable: automation removes the short contacts first, so what is left is a harder mix. This is the mirror image of the effect our first contact resolution entry works out, where a bot absorbing high-FCR categories drags blended FCR down. Same cause, two metrics, both moving the way that looks like failure. Neither is failure.

Three practical consequences, and the first is the one that costs money:

• Value automation in total handling minutes, never in AHT. Our ROI guide and quick math both price a ticket as hourly rate times AHT. That is correct for a snapshot and wrong for a comparison: multiply the new AHT by the old contact volume and you will manufacture a saving that is not there, or compare cost per contact before and after and conclude the bot made every ticket more expensive. Multiply minutes.
• Do not set an AHT target on a queue you are automating. If the person running support is measured on AHT and the same person is asked to roll out a bot, you have paid them to sabotage it. Change the target before the rollout, not after the first bad month.
• Re-baseline the mix, not just the number. Keep AHT broken out by request type. The blended figure is now describing a different population than it was last quarter, and only the per-type rows are comparable across the change.

What AHT does not tell you, and what to read it against

AHT is an occupancy metric. On its own it is compatible with almost any customer experience, good or bad, which is why it is so easily gamed: an agent who ends calls early and an agent who solves problems in one pass produce the same improvement.

The two layers with the cleanest handle-time case are both invisible to the customer: after-call work automation, which writes the wrap-up, and agent assist, which puts the answer in front of the human while they work. Our contact center automation guide sets out the full seven-layer stack these sit in. Our agent assist guide covers the baseline to record before switching either on.

A third layer does touch the customer, briefly, and is worth naming because it is the one with a published estimate attached. Gartner's August 2022 release put partial containment — capturing the caller's name, policy number and reason for calling before a human joins — at up to a third of the interaction time that would typically be supported by a human agent. That is a 2022 projection with an "up to" in front of it, not a measurement, and we have not tested it. If it holds, the appeal is that nothing is deflected and nobody is refused a person; the call simply starts with the agent already knowing who is on it.

Where it breaks

Comparing your AHT to a published benchmark. Of the four figures in the table above, three could not be tied to a stated measurement of the thing they claim to measure, and the one that could is measuring inbound voice at large North American call centers. We have not audited the whole field, but a benchmark that cannot tell you who it measured is not a benchmark for you. Your own trend on an unchanged instrument is the comparison that means something.

Blending channels. Voice, chat and email have different physics. An agent runs three chats at once and cannot run three calls. Blend them and you get a number that moves when your channel mix moves and is otherwise inert.

Treating a bot's session duration as AHT. It is not the same measurement. A bot session includes the user's thinking time, has no after-call work, and can sit open until a timeout that you configured. There is no agent occupancy in it, which is the entire thing AHT exists to measure. Our chatbot metrics guide covers what to measure on the bot side instead.

Setting it as an agent-level target. AHT is a capacity-planning input. Pointed at individuals it rewards ending contacts rather than resolving them, and the repeat contact that follows costs more than the minutes saved. Our escalation playbook is built to catch that after the fact; it is much cheaper not to cause it.

Reading a post-automation drop as success without asking why. If AHT falls after a self-service rollout, mix shift is not the explanation, so something else moved. Sometimes that is good: after-call work automation or agent assist landing in the same quarter, or better capture before the agent joins. Sometimes it is not: the bot quietly answering complaints, or escalating its hardest cases so abruptly that the agent inherits a short call and a bad problem. The direction alone tells you nothing. Find the cause.

Related terms

• First contact resolution — the outcome metric that moves the same counterintuitive way for the same reason.
• Response time SLA calculator — where AHT does its capacity-planning job: the input that moves the Erlang C service level most.
• Chatbot containment rate — the volume metric that tells you how big the mix shift is.
• Chatbot first response time — the queue-side number AHT is most often confused with.
• Agent assist — the automation layer with the cleanest handle-time case.
• Interactive voice response — where partial containment happens on the voice channel.
• Chatbot ROI — the calculation AHT is an input to, and the one the mix shift distorts.
• Customer effort score — the entry whose keyword note declined this term, and the metric that catches AHT bought at the customer's expense.

FAQ

What is average handle time?

It is the mean time an agent spends on one contact from picking it up to being free for the next: talk time plus hold time plus after-call work, divided by contacts handled. It measures how long the agent was occupied, not how long the customer waited and not whether the issue was resolved.

What is the average handle time formula?

AHT = (total talk time + total hold time + total after-call work) ÷ number of contacts handled. The formula is not the part that varies between organizations. What varies is whether after-call work is included, whether the denominator counts contacts or calls or tickets, whether queue time is folded in, and which channels are blended together. None of those choices travels with the number, so ask about all four before comparing.

What is a good average handle time?

We publish no benchmark and we would treat the ones in circulation with care. The widely quoted figure of about six minutes appears on vendor pages without a source; the academic study most often cited for per-sector figures reports a typical call handling time of 190 seconds in its own prose and its fieldwork ran from 2003 to 2006. The only figure we found with a named measurement population, carried from our own first contact resolution entry, is SQM Group's 697 seconds, measured at North American call centers on inbound voice, which SQM describes as an 18 percent increase on the prior year. Compare yourself to yourself, on an unchanged instrument, broken out by request type.

What is AHT in a call center?

The same metric, under its usual initials. AHT is the term of art in voice operations, where after-call work is a formal agent state and the number feeds staffing models directly. In digital support the same calculation is often called handle time or handling time and after-call work is less consistently captured, which is the main reason voice and chat figures should not be blended.

Why did my average handle time go up after we launched a chatbot?

Most likely because the bot is working. Automation absorbs the shortest, simplest contacts first, so the contacts left for humans are a harder mix and the blended average rises even when no category got slower. In the worked example on this page, total human handling time falls almost 25 percent while AHT rises 27.6 percent. Check total handling minutes and per-request-type AHT before concluding anything went wrong.

How do I reduce average handle time?

The durable levers are the ones that remove work rather than hurry it: after-call work automation, better agent-side retrieval so nobody hunts for an answer mid-call, and partial containment that collects identity and intent before a human joins. Gartner's 2022 release estimated the last of those at up to a third of the interaction time, which is a projection rather than a measured result. Pushing agents to talk faster produces shorter contacts and more of them.

Is average handle time the same as average handling time?

Yes. Ahrefs treats them as separate parent topics and we serve both from this page rather than publishing twice, on the same reasoning our first contact resolution entry sets out for its own naming split. The instrument is one instrument.

Should I measure AHT for my chatbot?

Not as AHT. A bot session has no agent occupancy, no after-call work and includes the user's thinking time, so the number is not comparable to a human figure and should not be averaged into one. Measure the bot on containment, escalation and outcome, and measure AHT on the humans downstream of it.

Sources

• Scorebuddy (ScorebuddyCX). On-page headline Contact Center Average Handle Time (AHT): Is it Important?, posted 10 February 2021, marked updated 30 July 2025, read 21 August 2026 at the URL below, which redirects from the older scorebuddyqa.com domain and carries a slug from an earlier headline — the starting point of this entry's citation trail. Specifically: the AHT formula given as (Talk + Hold + Delay + Follow-Up) ÷ number of calls; the worked example producing 15 minutes; the statement of a global industry benchmark of "6 minutes and 3 seconds" attributed in the sentence to Call Centre Magazine with the attribution hyperlinked to a Zendesk blog post; the per-sector figures quoted from the Cornell report as telecommunications 528 seconds, retail 324 seconds, business and IT services 282 seconds and financial services 282 seconds; and the description of AHT as a "lazy metric". Scorebuddy sells contact-center quality-assurance software; we have no relationship with the company and cite the page for what it says, not as authority. scorebuddycx.com
• Zendesk. What is average handle time (AHT) and how do you calculate it?, read in full 21 August 2026 — the destination of the attribution above, cited here for what it does and does not contain. Specifically: the statement that "a good average handle time is around six minutes", carrying no source; the eight-row per-industry table introduced as "estimates of standard industry benchmarks", carrying no source, methodology, sample or date; the call formula given as (talk + hold + follow-up times) ÷ total number of calls; the worked example arriving at 28 minutes; the article-level published date of 5 June 2026 and modified date of 17 June 2026 against a visible "Last updated June 16, 2026"; the product marketing byline; and the absence of any external citation anywhere in the body. The page contains no occurrence of "6 minutes and 3 seconds", of Call Centre Magazine, or of Cornell. Zendesk is a customer service software vendor. zendesk.com
• Holman, David; Batt, Rosemary; and Holtgrewe, Ursula. The Global Call Center Report: International Perspectives on Management and Employment, Report of the Global Call Center Network, copyright 2007 (US format ISBN 978-0-9795036-0-3; UK format ISBN 978-0-9795036-1-0), read 21 August 2026 — the study most often meant when an AHT page says "according to Cornell". Specifically: that it covers almost 2,500 centers in 17 countries and 475,000 employees, with research conducted 2003-2006; the statement that "the typical worksite in this report has an average call handling time of 190 seconds, or 3 minutes and 10 seconds" with "surprisingly little variation in this number across the wide range of countries in this study"; the executive summary statement that the overwhelming majority of centers operate as voice-only rather than multi-channel contact centers and that calls "typically last from 3-4 minutes"; and the median call times of 3 minutes 17 seconds at subcontractors against 3 minutes 20 seconds in-house. Limit of our check, stated because it bounds the claim: we read the report's extracted text and did not locate the per-sector figures that circulate under its name. They may sit in a table our extraction dropped. We do not assert they are absent from the report. The report states it was not funded by private corporations or companies operating in the call center sector. ecommons.cornell.edu
• SQM Group, via Chatbotscape. Call Center FCR Benchmark 2024 Results by Industry (published 6 February 2025) and What is a Good First Call Resolution Rate? (published 11 May 2022) — the source of the 697-second and 589-second figures and of SQM's own description of the first as an 18 percent increase on the prior year. These are carried from the source citation on our own first contact resolution entry, where both reports were read on 19 August 2026. They were not re-read at SQM for this entry. Two limits travel with them. The measurement population is set out on that page — 500-plus leading North American call centers, inbound voice only — but the post-call telephone survey described there is SQM's instrument for first contact resolution, and neither report as we have it states how the AHT figure specifically is derived; we do not assume it is survey-derived. That page also warns that the two reports' data years may not sit where their labels suggest, which is why this entry pairs the numbers only to show they reconcile rather than treating them as a trend series. SQM Group sells contact-center benchmarking and quality-assurance software.
• Gartner, Inc. Gartner Predicts Conversational AI Will Reduce Contact Center Agent Labor Costs by $80 Billion in 2026, press release, Stamford, Connecticut, dated 31 August 2022, read in full 21 August 2026 — cited on this page for one claim only, the partial-containment statement attributed in the release to Daniel O'Connell, VP analyst, that automating the identification of a customer's name, policy number and reason for calling "could reduce up to a third of the interaction time that would typically be supported by a human agent". This page paraphrases rather than quotes it, and treats it as the 2022 projection it is rather than as a measured outcome. The $80 billion forecast this release is named for is examined in the companion guide to contact center automation rather than here, and neither page asserts whether that forecast held. gartner.com
• Chatbotscape mix-shift model — the two tables in the automation section, the derived changes (contacts down 41.0 percent, handling minutes down 24.7 percent, blended AHT up 27.6 percent) and the cost figures at a fully loaded $36 an hour are our own arithmetic on illustrative volumes, published so they can be checked rather than as a measurement of any real operation. The $36 rate is illustrative and is not a benchmark. The model counts agent labor only, excludes the bot's own cost, and holds occupancy at 100 percent, all of which are stated where the figures appear. The structural result generalizes; the specific percentages depend entirely on the mix assumed.
• Site-wide string counts, run 21 August 2026 before this entry was saved. Search string grep -ril "average handle time\|average handling time" --include="*.md" ./glossary ./academy ./best ./news ./sample-reviews matched seven files before this publishing run: glossary/first-contact-resolution.md, where the phrase appears once in the running body and twice more across that page's two SQM source notes, and glossary/customer-effort-score.md, where it appears in the keyword note and once again in that page's own Sources block; and academy/chatbot-roi-guide.md, academy/chatbot-roi-quick-math.md, academy/agent-assist-guide.md, academy/chatbot-best-practices.md and academy/how-to-build-chatbot.md, where it is an operational input to a calculation or an instruction to measure. No file defined the term. Two disclosures so the command reproduces: after this run the same command matches nine, adding this entry and its same-day companion contact center automation guide; and running it unscoped from the repository root doubles every count, because web/.content/ holds a build mirror of the content directories. Published as a transparency statement about a gap in our own coverage.
• Ahrefs Keywords Explorer, US overview and volume-by-country, queried 21 August 2026 — the search-demand, difficulty, parent-topic and country-split figures in this entry's keyword note, including the checks behind declining 'voice of the customer', 'chunking', 'conversational analytics', 'ticket deflection' and 'cost per contact'.
• Chatbotscape evaluation methodology. /methodology (continuously updated).