Callers are hanging up in the first eight seconds

The agent is being judged in the first turn, on evidence that has nothing to do with whether it can book a job. Nobody has heard the scheduling logic yet. They have heard how long the phone rang, how the line sounded when it picked up, how many words arrived before anyone asked what was wrong, and whether there was a gap afterwards that felt like a dead line.

The tell is a set of numbers that look contradictory: answer rate at or near 100 percent, booking rate well below what the same volume used to produce, and a pile of recordings 6 to 12 seconds long. That is not a booking problem wearing a disguise. It is a first-impression problem, and it is measurable to the second.

Build the drop-off histogram before you change a word

You need one chart: for every inbound call that was answered and did not reach a booking, the number of seconds between answer and hang-up, bucketed by second. Your telephony provider's call records carry both timestamps, whether the platform is Twilio, RingCentral or a carrier's own portal.

1. Pull 30 days of answered inbound calls with answer time, end time and end reason. Exclude calls the agent transferred or ended itself, so what remains is caller-initiated abandonment.
2. Bucket by second from answer to hang-up, from 0 to 30, then one bucket for everything longer.
3. Segment by first-time versus repeat caller, matched on the inbound number. These two curves differ sharply and averaging them hides both.
4. Segment again by hour of day. Evening and overnight callers arrive with a different expectation and a different tolerance from a caller at 10:00 on a Tuesday.
5. Add ring duration before answer as a second series. If a meaningful share of hang-ups happened before the agent said anything, the fault is upstream of the script entirely.
6. Listen to 20 calls from the largest bucket, in order, without reading the transcripts first. Transcripts hide exactly the things that cause this: the pause, the tone of the pickup, the beep.

The five first-turn faults, ranked

Ranked by how often each turns out to be the dominant cause, and each cheap to rule out.

1. Too many rings before pickup

An automated line has no excuse for ringing. If the deployment answers on the third or fourth ring because a human queue is tried first, a share of callers is lost before the agent exists. Target an answer inside 2 rings, and measure the ring duration rather than trusting the configuration — overflow rules and forwarding hops add time that nobody set deliberately.

2. A greeting that lists options before asking what is wrong

A homeowner standing in water does not want to know about the parts department. Any structure of the form 'for service press 1, for accounts press 2' tells that caller the line cannot help them, and they will ring the next firm rather than navigate. The first turn has one job: identify the business briefly, then ask an open question about the problem.

3. The silence after the disclosure

This is the one that surprises people. The disclosure sentence itself rarely loses the call; the gap after it does. If the agent finishes speaking and then takes over a second to begin listening or to respond, the caller hears a dead line and hangs up into it. Keep the gap between the agent finishing and the caller being able to speak under roughly 800 milliseconds, and treat anything past 1.5 seconds as a defect with a ticket rather than a tuning preference.

4. It sounds like voicemail

A beep at connect, a bar of hold music, or a flat recorded-sounding opening line all read as an answering machine, and people hang up on answering machines without thinking about it. Listen to the first two seconds of 20 calls with the transcript closed. If you cannot tell within that window that a conversation is starting, neither can a caller who is annoyed and holding a phone in a plant room.

5. The caller expected a specific person

Small trades numbers are often a mobile that has been answered by the owner for years. A caller who has been told to 'ask for Dave' and gets an automated line hangs up to try again later. This shows up as a hang-up spike among repeat callers only, and it is a communication problem rather than a design one. It also appears when the number was published somewhere that sets a different expectation — a caller who came through a shared lead marketplace is expecting a person to ring them back, which is one of the differences drawn out in an owned channel versus a shared lead marketplace.

Nobody abandons a call because the scheduling logic is weak. They abandon because eight seconds passed and nobody asked what was wrong.

Rewriting the first turn, with a number to hit

The target is time-to-first-question: seconds from answer until the caller is asked something they can answer. Under 8 seconds is achievable on any stack, and most failing first turns are at 15 to 25 because of what has been packed in ahead of the question.

One boundary worth being explicit about: everything above concerns the turn before the caller has spoken. Once they are talking, the failure modes change entirely and belong to how interruption handling decides whether callers trust the agent. And the lookups that make the agent useful — customer history, real availability — must not be what creates the pause, which is why the read-hold-confirm sequencing in an intake agent that only offers slots it can hold matters as much for latency as for correctness.

Recovering the ones who already hung up

While the first turn is being fixed, the abandoned calls are still jobs. A caller who hung up at second 6 is holding a phone and is very likely still looking, so the recovery window is minutes rather than hours. Whether that recovery should be an immediate text or a human callback depends on the tier and on who is holding the phone, which is the split argued in a callback queue versus missed-call text-back.

One caution: do not run a recovery campaign and a greeting rewrite in the same week. Both move booking rate, and if you change them together you will not know which one worked, or whether one of them made things worse.

What to watch afterwards

• Time-to-first-question, as a p50 and a p90. The p90 is where routing hops and cold-start delays hide.
• Hang-ups before second 10, as a share of answered calls. This is the number the whole exercise is moving.
• Inter-turn gap after every agent utterance, alerted rather than reviewed. A regression here arrives with a model or platform change and is otherwise invisible.
• Booking rate split by first-time and repeat caller, so an improvement among strangers is not masked by regulars who were always going to book.
• Abandonment by hour, checked after any change to overnight handling, because the evening curve moves independently of the daytime one.

None of this is a large build: instrumentation on the call path, a rewritten first turn, and a threshold that alerts. It is the kind of narrowly scoped work we run as MVP and product builds, and it belongs before any expansion of what the agent is allowed to do. The rest of this silo — what a booking must capture, how tiers are set, what the agent may promise — sits in intake and call capture, within our field service and trades work.

• intake
• voice agents
• call quality
• measurement