Research-backed calculator guide

Dead Air Cost in Voice AI: Measure Silence, Repeat Prompts and Abandoned Calls

Learn how to measure voice AI dead air cost by separating silent seconds, repeat prompts, telephony time, and abandoned-call exposure with a worked monthly model.

What counts as dead air in a voice AI call?

Use “dead air” for a measurable period during a connected interaction when the caller receives no useful audible response and the system has not clearly signaled what it is doing. That can happen after the caller finishes speaking, before the first TTS byte arrives, during a tool call, while a transfer is being prepared, or after a speech-recognition error. It is different from a caller’s intentional pause, a disclosed hold message, or a short natural gap that helps turn-taking.

The measurement boundary matters. A call can feel silent to a caller while the LLM is working, yet only one component may be billing. Conversely, an agent can say a filler such as “one moment” and still accumulate connected minutes while a tool call runs. Record both perceived silence and the meter that was running at that moment. This is why a single “cost per call” average is too blunt for diagnosing voice bot abandoned call cost.

A useful event taxonomy has four rows:

1. Response dead air: time from the caller’s end of speech to the first useful agent audio.
2. Listening dead air: caller silence while the agent remains actively listening and the call stays connected.
3. Processing or transfer dead air: time spent waiting for a tool, human handoff, or remote service.
4. Outcome exposure: the cost of abandonment, repeat contact, or human recovery that may not be billed by the voice provider.

Vapi’s documentation makes a related operational distinction: a silence timeout ends a call after configured inactivity, while maximum duration ends a call after total elapsed time. It documents a dashboard silence-timeout range of 5 to 3,600 seconds. That setting is a control, not proof that a shorter timeout is financially better; an aggressive value can cut off a caller and create a repeat contact. ^[4]

The meters behind voice AI dead air cost

Start by listing every meter in the deployed stack. Official pages show why rates cannot be copied from one vendor to another. Twilio’s US Voice page lists pay-as-you-go call rates by number type and separate charges for phone numbers, recording, transcription, Media Streams, and Conversation Relay; the page says its pricing was current as of August 2026. ^[3] Deepgram lists Voice Agent rates per minute of WebSocket connection time and says its speech-to-text audio is billed per second without block rounding for normal requests. ^[5] In another stack, an orchestration platform may combine speech-to-text, LLM, and TTS into one minute rate while telephony stays separate.

Build a meter map with these columns:

• Event: for example, 12 seconds of silence after a tool call.
• Observed duration: the timestamp difference from your trace or call recording.
• Provider meter: connected minute, audio second, WebSocket second, token, character, or fixed fee.
• Rate and unit: such as dollars per minute or dollars per 1,000 characters.
• Included or excluded: whether the event is already inside a bundled rate.
• Confidence: measured, vendor-documented, or planning estimate.

This prevents double counting. If a bundled Voice Agent line already includes STT and TTS, do not add those layers again. If your carrier continues billing after the AI drops from a transfer, place those post-transfer minutes in the telephony row only. If an open WebSocket bills while no audio arrives, include it only when the provider’s terms or invoice confirms that rule.

Use the same discipline when comparing options in the model pricing comparison. A low speech rate is not automatically a low conversation rate if the agent has longer pauses, more retries, or another platform fee.

A practical measurement model

For each call, capture at least:

• connected start and end timestamps;
• caller speech end and first useful agent-audio timestamps;
• total silence seconds, split by cause;
• number of repeat prompts and their duration;
• transfer start, connection, and post-transfer duration;
• abandonment event and whether a retry occurred;
• provider usage IDs and invoice dimensions;
• resolution or completion outcome.

Then calculate three outputs instead of one headline number.

Direct silent-time cost is silent minutes multiplied by the meter rate that was active. Prompt-retry cost is repeat-prompt seconds divided by 60, multiplied by the active rate. Outcome exposure is abandoned calls multiplied by an estimated retry or recovery cost. Keep the third number visibly labeled as an estimate because an abandonment is not the same as a lost sale, and a retry may be free, self-service, or handled by a human.

A simple per-call version is:

direct waste per call = (silent seconds + repeat-prompt seconds) / 60 × active rate

A monthly version is:

monthly direct waste = sum of direct waste per call + measured extra telephony or platform charges

For outcome exposure:

retry exposure = abandoned calls × retry rate × retry minutes × applicable rate

If you have a reliable contribution margin or human recovery cost, add a separate business-impact scenario. Do not present that scenario as an invoice amount.

Worked cost example

Assume an English-US inbound support line with these planning inputs. The blended rate is a deliberately rounded internal estimate that combines the relevant AI and telephony meters; replace it with reconciled provider rates after a pilot.

The first two dollar rows are direct modeled usage. The retry row is an expected exposure, not a guaranteed loss. If the 280 retry calls would have happened anyway, the attribution should be reduced. If 50 of those callers instead require a human callback, add that callback cost in a separate scenario. If the platform bills per second and the carrier bills by a different increment, calculate each layer independently before summing.

This example also shows why reducing the silence rate can beat shaving a fraction of a cent from a model. Cutting the average avoidable silence from 15 to 10 seconds saves 1,666.7 minutes, or $300 per month at the assumed rate. That is a model result, not a universal ROI promise. Validate it against outcome quality: a rushed turn detector can create interruptions, repeat prompts, and more abandonment.

For invoice reconciliation, join the call-level event ID to usage exports rather than relying on a monthly average. The AI API invoice reconciliation calculator is the natural next step when the operational ledger and vendor invoice disagree.

How to reduce dead air without creating new cost

First, set a baseline. Sample calls across busy and quiet periods, then report median and p95 response silence. Averages hide the calls that callers perceive as broken. Split silence by cause: end-of-turn threshold, model time, tool latency, TTS startup, transfer setup, or an actual disconnect.

Second, make the first audible response useful. A short acknowledgment can protect caller confidence, but it is not free if it triggers extra TTS or extends the call. Measure whether it reduces abandonment enough to justify its added audio and connected time.

Third, cap unproductive waits. A silence timeout can end truly inactive calls, but use a warning or recovery prompt first where appropriate. Log whether the caller was speaking, whether speech recognition failed, and whether the timeout was expected. Vapi’s timeout distinction is a useful schema for separating inactivity from maximum-duration endings. ^[4]

Fourth, instrument tools and transfers. A tool call that takes two seconds in a test may take ten seconds under load. Record the time at tool start, result arrival, first spoken byte, and call end. For transfer flows, separate AI time from post-transfer carrier time. Retell’s published guide illustrates this separation and notes that transfers can leave telephony running after the AI fee stops. ^[1]

Finally, reconcile the telemetry to a bill. Deepgram’s per-second and WebSocket examples show why the trace duration and invoice unit must be stored together. Twilio’s page likewise separates carrier minutes from optional recording, transcription, and AI-related lines. ^{[3] [5]} A cost dashboard that shows only “minutes” is not enough to explain a variance.

What to exclude from the first model

Exclude caller-intent pauses that are part of a successful conversation unless your question is total connected-time efficiency. Exclude human labor and lost revenue from direct provider cost. Keep them in an outcome scenario with documented assumptions.

Do not assume every silence second is billed. Confirm whether the provider bills connected duration, processed audio, WebSocket connection, tokens, characters, or a rounded block. Do not transfer one vendor’s pricing date to another vendor. The cited Twilio page is explicitly dated August 2026, and the AssemblyAI guide dates its public-price checks to October 2026; rates can change, so record the checked date beside every input. ^{[2] [3]}

Do not count a repeated prompt twice if the original prompt already includes it in a bundled per-minute rate. Do not call a retry a lost customer without a measured conversion or recovery baseline. Finally, do not infer ROI from dead-air reduction alone: a shorter call can be worse if it clips speech or lowers resolution.

FAQ

Is dead air billed in voice AI?

Often it can be, but there is no universal rule. Connected-time platforms may keep billing while a call is silent, while other services meter processed audio, WebSocket time, tokens, or characters. Retell’s guide says silence and hold are billed in its described model, and AssemblyAI says per-minute models generally bill connected time; Deepgram documents its own Voice Agent WebSocket-minute and STT per-second meters. Treat those as product-specific evidence and verify your contract and invoice. ^{[1] [2] [5]}

How do I calculate the cost of silence in phone AI?

Measure avoidable silent seconds per call, multiply by monthly connected calls, divide by 60, and multiply by the rate for the meter that was active. If more than one meter runs during the gap, calculate each layer separately. Keep natural caller pauses, outcome exposure, and human recovery costs in separate rows.

What is the difference between dead air and a silence timeout?

Dead air describes the caller-visible quiet period. A silence timeout is a configuration that ends an inactive call after a threshold. A timeout may reduce open-call usage, but an overly short setting can interrupt a caller and create a repeat contact. Vapi documents silence timeout and maximum duration as different controls. ^[4]

Should repeat prompts be included in voice bot abandoned call cost?

Yes, but report them in two ways. The repeat prompt itself creates measurable audio and connected-time usage. The caller’s later abandonment is an outcome event that may create retry or human-follow-up exposure. Combining both into one cost hides the mechanism and makes it hard to test an intervention.

What rate should I use for a dead-air cost estimate?

Use the rate actually applicable to your stack, with the unit and pricing date. If you are pre-launch, use a clearly labeled planning estimate and show a low, base, and high case. Do not copy a bundled headline rate into a layered stack without checking telephony, TTS, LLM, concurrency, and add-on charges. The AI observability ROI calculator can help frame the measurement and payback side once you have baseline events.

How can I prove that a latency fix saved money?

Run a before-and-after comparison with the same call mix where possible. Track p50 and p95 response silence, repeat prompts, total connected minutes, abandonment, retries, resolution, and invoice cost. Attribute savings only to the changes supported by the data. A lower p95 delay with a higher repeat-prompt rate is not automatically a financial improvement.