Where callers hang up on AI calls

The short answer

Many hang-up reports count every call that ended early as a failure. This is wrong in two ways. First, a short call can be completed: the caller heard the delivery date and left after 40 seconds. Second, the stage that loses the most callers is often just the stage that every call passes through.

So sort every caller hang-up into one of two classes. A hang-up is completed when the caller got what they called for, then hung up. It is abandoned when the caller hung up before that. Then place each abandoned call at the one stage where it happened: Opening, Reason, Identity check, Lookup wait, Action or Wrap-up.

Finally, compare stage rates, not shares. The stage rate is the number of calls abandoned at a stage divided by the calls that reached that stage. The share is the same number divided by all abandoned calls. The share shows which stage has the most abandoned calls. The stage rate shows which stage loses the highest proportion of the callers who reach it.

This is a method for your own calls. Every figure below is an example written for this article, not DRING data and not industry data.

Short does not mean failed

Duration is not an outcome. It tells you how long a call took, not whether the caller got what they called for. Here are two example calls, written for this article:

• Call A, 40 seconds: the caller gives an order number, hears "Thursday, between nine and twelve" and hangs up. Completed.
• Call B, 3 minutes 10 seconds: the agent asks for the order number twice, then goes silent while it checks the order. The caller says "Hello? Is anyone there?" and hangs up. Abandoned.

A report that flags every call under one minute as a failure gets both calls wrong. It flags Call A as a failure and does not flag Call B at all.

Average duration misleads in the same way. Suppose you make the system check faster, so the agent reads the date sooner. Average duration falls and more calls end in under a minute. A report that reads short calls as failures shows both changes as warning signs, while the line actually got better.

Duration is still useful inside a stage: the length of the silence during Lookup wait is worth measuring.

The six stages of a call

A stage is a part of the conversation with a clear start and end. Place each hang-up at the stage in which it happened. If the caller hung up during a silence, use the stage the silence belongs to. For example, the pause after the caller's last answer in Identity check is Lookup wait.

• Opening: from pickup until the conversation reaches its topic. On an outbound call, the person hears in the first sentences who is calling and why, and decides whether to stay. On an inbound call, the caller chose to call, so fewer callers abandon here than on outbound calls. Inbound abandoned calls at Opening often mean a wrong number, or a caller who expected a person or a menu.
• Reason: the caller explains the need and the agent confirms it. An abandoned call here usually means the agent asked again or misunderstood the request.
• Identity check: the agent asks for the details that find the caller or the order. Abandoned calls here point to too many questions. They can also point to a request for something the caller cannot easily find, such as an order number while driving.
• Lookup wait: the agent checks the system and the caller waits. Abandoned calls here usually follow a silence, when the caller cannot tell whether anyone is still on the line.
• Action: the agent gives the answer or makes the change. Completed hang-ups start here: a caller who hears the date and hangs up before Wrap-up is completed. An abandoned call here often means the change failed, or the agent promised something and then could not do it.
• Wrap-up: the summary, "anything else?" and the goodbye. Most hang-ups in Wrap-up are completed: the caller has the answer and does not wait for the goodbye. Abandoned calls are rare here. One example is a summary that gets the new address wrong.

Four rules keep the map clean:

• Write down what completed means for each call type. For a status call, it means the caller heard the date. For an address change, it means the agent confirmed the new address. Two reviewers who use the same rule mark the same call the same way.
• Count a handover as an exit, not a hang-up. A caller who hangs up while waiting for the transfer is abandoned, at the stage where the handover started. Our guide to designing human handover covers what the transfer should carry.
• Keep only the caller's decisions. A call cut off by the line, or ended too early by the agent, can look like a caller hang-up in the log. Check a sample and count those calls separately.
• Build one map per line and direction. Inbound and outbound Openings are not comparable. For outbound, count only calls that a person answered. Voicemail and unanswered calls are not part of the map.

The hang-up map: an example

Example, not real data: 1,000 inbound calls to a delivery line in one month. The agent handles two call types, delivery status and address changes, and both use all six stages. Other requests, such as refunds, are handed over to the team at the Reason stage, by design.

Every call ends exactly once: as a completed hang-up, an abandoned hang-up, a handover or a normal end after the goodbye. Callers who hung up while waiting for a transfer are in the Abandoned column. Calls cut off by the line, or ended too early by the agent, were set aside before counting. The calls that reach the next stage are those that reached this stage, minus those that ended at it.

The other 262 calls ended normally, after the goodbye. Together, 216 completed hang-ups, 200 abandoned hang-ups, 322 handovers and 262 normal ends make 1,000 calls.

Three readings of the same table

Count every hang-up, and Action comes first with 128, followed by Wrap-up with 98. But 120 of the hang-ups at Action and 96 of those at Wrap-up are completed. This reading sends you to fix the two stages where calls succeed.

Rank abandoned calls by share, and Opening comes first with 30%, just ahead of Lookup wait with 29%. It looks like a problem with the greeting. But every call passes through Opening, so even a moderate rate there produces a large count.

Compare stage rates, and the order changes. Lookup wait loses 10.0% of the callers who reach it, one in ten. Identity check loses 8.0%. Opening loses 6.0%. Start with Lookup wait and Identity check.

Shares have one more weakness: they move when another stage changes. Suppose a fix cuts abandoned calls at Lookup wait from 58 to 35, and the other stages keep their counts. Opening's share rises from 30% to 34%, with the same 60 calls. Its stage rate stays at 6.0%. Raw counts can mislead as well. When fewer callers leave early, more reach the later stages, so the counts there can rise even if nothing got worse.

What happened before the hang-up

The map shows where callers leave. The recording shows what happened just before. For each abandoned call at the stage you are looking at, listen to the last minute before the hang-up. Note the last agent turn, the caller's last sentence, the silence before the hang-up and any question asked twice.

Write what you find on an abandon review card, one card per call. The example column shows the card for Call B.

Read about 20 abandoned calls at each stage with a high stage rate before you change anything. Twenty calls do not give you a rate, but they show whether one pattern repeats. For example, if 14 of 20 Lookup wait cards show a long silence before the hang-up, you have a pattern worth testing. If every card looks different, read more calls first. Do this in addition to your routine call quality sampling.

The stage tells you where callers leave, not which part of the system caused it. A long silence during Lookup wait can come from a slow system, or from an agent that says nothing while it waits. To find the layer that broke, use the first-break method in our article on good demos and bad live calls. The same article covers calls that the agent ends before the caller says goodbye.

A stage is not a cause

A caller who hangs up during Lookup wait may have left because of the silence. They may also have been driving and needed to end the call. Both calls look the same on the map.

So test the likely trigger from the review cards in three steps:

1. Change one thing at the stage, for example a spoken update during the system check. If you change two things, you cannot say which one moved the rate.
2. Compare stage rates before and after on comparable calls: the same call type, the same hours and the same days of the week. Morning callers in cars and evening callers at home can leave for different reasons.
3. Check the difference against chance. On 580 calls, a stage rate of 10% carries a 95% margin of about ±2.4 points. The difference between two periods of about 580 calls each has a wider margin, about ±3.5 points. So a drop to 9% can be chance. A drop to 6% is unlikely to be chance alone. Our article on voice AI pilot success criteria shows how to compute the margin.

Fixes by stage

Once the review cards show a pattern, start with the first fix that matches it. Then test it as described above.

Where DRING fits

A hang-up map needs two labels on each call: the stage where it ended, and how it ended. On DRING, after-call analysis can return additional fields that you define, through the API, your CRM, the dashboard and reports. Both labels can be among them. Reviewers still read the abandoned calls, and the fields show them which calls to read. Our call analytics page shows the dashboard and reports.

Human handover is part of the shared foundation of every DRING package. So "offer a person" is a fix you can test at the stage where callers leave. Every DRING agent runs 1,000 to 10,000 simulated conversations, built for your company, before the first real call. Simulated callers test the flow, but they do not show where real callers give up. The hang-up map comes from live calls.

If one stage on your line loses too many callers, request a callback. Have a sample of abandoned calls ready, with recordings where your policy allows. Our team can read them with you, stage by stage.