Last updated: August 6, 2026
Average handle time is the metric support leaders quote most and understand least. It is easy to pull, easy to chart, and easy to weaponize — which is exactly why so many teams have driven it down while their customer satisfaction quietly collapsed underneath them.
The number itself is not the problem. AHT is a legitimate signal about process friction, tooling gaps and knowledge quality. The problem is treating it as a performance target for individual agents rather than a diagnostic for the operation. This guide covers what AHT actually measures, what a realistic benchmark looks like for your call type, and the specific levers that shorten handle time by removing work instead of rushing people.
Average handle time is the mean total duration of a customer interaction from the moment an agent picks it up to the moment they finish everything the interaction requires. That last part is where most teams get the definition wrong.
AHT is the sum of three components divided by the number of interactions handled:
AHT = (Total talk time + Total hold time + Total after-call work) ÷ Number of interactions
After-call work — the wrap-up, the CRM notes, the ticket tagging, the internal escalation email — is the component teams most often leave out, and it is frequently the fastest-growing part of the total. If your AHT looks suspiciously stable while agents complain about being underwater, check whether wrap-up is being counted at all.
AHT measures agent-side effort, not customer-side experience. It excludes queue wait time, time spent in self-service before reaching an agent, and any delay between a customer's first attempt and eventual resolution. A customer who waited eleven minutes in queue and then had a brisk four-minute call contributes a very good AHT and a very bad experience.
This is why AHT should never travel alone. Pair it with first contact resolution, which tells you whether the interaction actually ended the problem, and with customer effort score, which tells you what the interaction cost the customer.
Published benchmarks put the call center industry standard somewhere between roughly seven and ten minutes depending on whose methodology you use, according to SQM Group's analysis of top call center KPIs. Cross-industry aggregations from Kayako's AHT benchmark data land in a similar range, with wide variance by sector.
Those numbers are close to useless applied at the top level. A blended AHT across a mixed queue averages a two-minute password reset with a twenty-minute billing dispute and produces a figure that describes neither.
| Interaction type | Typical AHT range | What drives the time |
|---|---|---|
| Status and order lookups | Under 3 minutes | System lookup speed |
| Account and profile changes | 3–6 minutes | Verification and data entry |
| Billing and dispute handling | 8–15 minutes | Policy judgement, approvals |
| Technical troubleshooting | 10+ minutes | Diagnosis and escalation |
Key takeaway: Segment AHT by contact reason before you set any target. A single portfolio-level goal will always push agents to rush the complex work where rushing is most expensive, because that is the only place where meaningful seconds can be shaved.
When AHT becomes an agent-level target, agents optimise for the target. That optimisation has a predictable shape: shorter calls, more transfers, more premature closures, and more customers who call back tomorrow about the same thing. Handle time drops. Contact volume rises. Total cost goes up.
The failure mode is arithmetic. Two five-minute contacts to resolve one issue produce a better AHT than one nine-minute contact that resolves it — and a worse outcome for everyone. If you are not watching repeat-contact rate alongside AHT, you cannot tell the two situations apart.
Customers notice, and they are less tolerant than they used to be. Zendesk's CX Trends 2026 research, based on responses from more than 11,000 people worldwide, found that 88% of customers expect faster response times than they did a year earlier, and that 74% are frustrated when they have to repeat information they have already given. A transfer-heavy strategy for reducing handle time collides with both of those expectations at once.
Never move AHT without watching first contact resolution and repeat-contact rate in the same view. A genuine improvement shows AHT falling while FCR holds or rises. A false improvement shows AHT falling while FCR falls with it. The same logic applies to automation: containment rate only means something if the contained conversations were actually resolved rather than abandoned.
Every durable AHT reduction comes from the same place: removing work from the interaction rather than compressing it. Here are the levers that do that, roughly in order of return on effort.
A meaningful share of every interaction is spent establishing who the customer is, what they bought, and what happened last time. None of that is judgement work. Surfacing the account record, recent orders, prior tickets and open issues at the moment the interaction opens removes a chunk of talk time without touching the quality of the conversation. This is also the fix with the best side effect: it directly addresses the repeat-yourself frustration that drags down satisfaction scores.
After-call work is pure overhead from the customer's point of view and it is often the single largest recoverable block in the AHT formula. Auto-generated summaries, auto-populated disposition codes and auto-drafted follow-up emails hand the agent something to approve rather than something to compose.
Knowledge base search moves the problem rather than solving it: the agent still has to read, interpret and translate. Retrieval that returns a drafted response grounded in your own policy documents removes the translation step entirely. Research on generative AI deployments in customer support has found the effect is largest for less-experienced agents, effectively narrowing the gap between new hires and veterans — one widely cited study measured a 14% average productivity increase, concentrated among newer staff.
The fastest handle time is the one that never happens. Order status, shipping windows, password resets, invoice copies and appointment changes are deterministic requests that resolve fully in an automated channel. McKinsey's analysis of generative AI in service operations suggests the technology could reduce human-serviced contacts by up to 50% in sectors like banking, telecommunications and utilities. Note what this does to your reported AHT, though: pulling out the short, simple contacts raises the average of what remains. That is a good outcome that looks like a bad one on the dashboard, which is why segmenting matters.
If ticket volume rather than duration is your binding constraint, the mechanics are covered in more depth in our guide to cost per resolution, which reframes the whole question in unit-economics terms.
Agents put customers on hold for a small, knowable set of reasons: looking something up, waiting for a supervisor approval, or checking with another department. Each is a process defect with a specific fix — better search, delegated approval thresholds, a shared channel with the other team. Attacking hold time as a number just teaches agents to keep customers on the line in awkward silence instead.
Most routing logic asks who is free. Better routing asks who is fastest at this specific problem. Matching contact reason to demonstrated agent strength shortens the interaction and improves the outcome simultaneously, and it makes the eventual AI-to-human handoff land with someone equipped to close it on the first pass.
This is the layer where Darwin AI's Eva tends to earn her keep: she handles the repetitive, deterministic contacts end to end in the customer's channel of choice, assembles full context for the ones that need a person, and drafts the wrap-up so the agent's remaining work is review rather than reconstruction.
Run the evaluation as a before-and-after on a segmented cohort, not a portfolio average. For each contact reason, track four numbers together over the same period:
A real improvement moves AHT down while the other three hold steady or improve. Anything else is displacement. McKinsey's field work on gen AI in customer care documented one deployment across roughly 5,000 agents that produced a 14% increase in issues resolved per hour alongside a 9% reduction in handling time — the useful detail being that both numbers moved in the right direction together, which is the pattern you are looking for.
Publish an acceptable band per contact reason rather than a ceiling. A band signals that unusually short interactions are as much of a red flag as unusually long ones, which is true and which agent-level ceilings never communicate.
Cut handle time by removing the work, not rushing the conversation.
Darwin's AI workers resolve routine contacts end to end and hand agents full context plus a drafted wrap-up on everything else.
See how Eva works →There is no single good number. Published cross-industry benchmarks generally sit between roughly seven and ten minutes for voice, per SQM Group, but that figure blends contact types that have nothing in common. A useful target is set per contact reason and expressed as a range: under three minutes for simple lookups, eight minutes or more for billing disputes and technical diagnosis.
Yes. Wrap-up is real agent capacity consumed by the interaction, and excluding it hides the fastest-growing component of handle time in many operations. If your current measurement leaves it out, expect your AHT to jump when you add it — that jump is a correction, not a regression.
It does when the reduction comes from rushing agents, and it does not when the reduction comes from removing steps. The distinguishing test is first contact resolution: if FCR and repeat-contact rate hold steady while AHT falls, the work genuinely got shorter. If FCR falls with AHT, you have moved cost from this contact to the next one.
Published results vary widely by baseline and use case. McKinsey's work on gen AI in service operations documented a deployment across around 5,000 agents that saw a 9% reduction in handling time alongside a 14% increase in issues resolved per hour, and a separate study of a large support organisation measured a 14% average productivity gain concentrated among less-experienced agents. Treat single-digit to low-double-digit percentage improvements as a realistic planning assumption rather than the headline numbers in vendor case studies.
Almost certainly because automation absorbed the short contacts. When two-minute status checks stop reaching agents, the remaining human queue is composed entirely of harder work, so the average rises even though total cost fell. Segment by contact reason and compare like with like, and check total resolution cost rather than duration alone.