Use cases
Built for the calls that go wrong most.
Four kinds of calls where the audio, not the model, decides whether the agent sounds competent. Each one lists what breaks, what Tenms changes, and what to measure in a pilot.
01 / 04
Customer support
Callers ring in from cars, kitchens and open offices, usually on speaker. The agent talks for a while, walking them through steps, and they acknowledge as they go.
What breaks today
- The agent's own voice comes back through the speakerphone and the agent stops itself mid-step.
- “Okay” and “right” are treated as interruptions, so instructions arrive in fragments.
- When a call fails, nobody can tell whether it was the line, the transcript or the prompt.
What Tenms changes
- FocusRemoves the agent's echo using its own playback as a reference, before your STT sees the audio.
- TurnLabels each interruption. Acknowledgements are ignored; only a real barge-in stops the agent.
- PulseAttributes each failed call to audio, transcription or the LLM, with the reasons.
- AgentHold the button on the back of the router for ten
- Callerokay
- AgentSorry, go ahead.
- CallerI didn't say anything. Ten what?
- AgentHold the button on the back of the router for ten seconds,
- Callerokay
- interruption.backchannelignored, agent keeps talking
- Agentuntil the light starts blinking. Tell me when it does.
Measure in a pilot
turns.backchannels- acknowledgements that no longer stop the agent
quality.echo_ratio- how much of the call was agent echo
failure.cause- where failed calls actually broke
02 / 04
Appointment scheduling
Short, structured calls: confirm, reschedule, remind. Callers agree constantly while the agent reads details, and change their mind halfway through a sentence.
What breaks today
- “Mm-hm” halts the agent while it is reading out a date and time.
- A real correction, “actually, Friday”, arrives while the agent is still talking and gets talked over.
- Every answer waits for a fixed silence timeout before the LLM is even called.
What Tenms changes
- TurnIgnores acknowledgements and stops for a genuine barge-in.
- TurnFires turn_prediction before turn_end, so you can start the LLM early.
- FocusKeeps the agent's own playback out of the caller's audio.
- AgentYour appointment is confirmed for Thursday at
- Callermm-hmm
- AgentSorry, did you want to change something?
- AgentYour appointment is confirmed for Thursday at three p.m.
- interruption.backchannel“mm-hmm” ignored
- CallerActually, can we move it to Friday?
- interruption.barge_inagent stops
- turn_end360 ms after the last word
- AgentOf course. Friday at three works.
Measure in a pilot
turns.eot_delay_ms_p50- time from the caller's last word to turn_end (360 ms live)
turns.backchannels- acknowledgements handled without stopping
turns.barge_ins- real corrections that stopped the agent
03 / 04
Drive-through and ordering
An outdoor speaker post, an idling engine and wind. The agent's voice comes out of the post and straight back into the microphone, and passengers talk while the driver orders.
What breaks today
- The agent hears its own menu prompt through the post and interrupts itself.
- Engine and road noise push up word errors on every item.
- A passenger's comment is taken as the customer cutting in.
What Tenms changes
- FocusRemoves the agent's echo and suppresses steady noise such as engines, fans and road hum.
- TurnLabels noise and other_speaker interruptions so they do not stop the agent.
- PulseReports SNR and ASR risk every second, so you can see which lanes or hours are hard to hear.
- AgentWould you like to make that a large
- ( engine idling, passenger talking )
- AgentSorry, I didn't catch that. What would you like?
- AgentWould you like to make that a large for one dollar more?
- Passengerget me a cookie too
- interruption.other_speakerignored
- CustomerYes, large, and add a cookie.
Measure in a pilot
quality.snr_db- signal-to-noise per call
quality.asr_risk- likelihood the audio was mis-transcribed
turns.ignored_other_speaker- voices that were not the customer
Known limit. Crowd chatter is barely reduced today (about +0.5 dB), and speaker lock only separates clearly different voices. Steady noise is where Focus is strongest.
04 / 04
Sales and lead qualification
Fast back-and-forth. Prospects answer from cars and streets, cut in to object, and give short answers that need a quick reply.
What breaks today
- The agent keeps pitching over an objection.
- Short answers are followed by a long pause, and the call starts to feel scripted.
- A lost call gets blamed on the pitch when the line was the problem.
What Tenms changes
- TurnStops the agent on a barge-in: about 355 ms with interim transcripts, about 1 s from audio alone.
- Turnturn_prediction lets the LLM start before the caller's turn has formally ended.
- PulseSeparates calls lost to a bad line from calls lost in the conversation.
- AgentMost teams our size save around a fifth on
- CallerSorry, we already have a vendor.
- Agenttheir annual spend, and onboarding takes a week.
- AgentMost teams our size save around a fifth on
- CallerSorry, we already have a vendor.
- interruption.barge_inagent stops
- AgentUnderstood. Who are you using today?
Measure in a pilot
turns.barge_ins- objections the agent stopped for
turns.eot_delay_ms_p95- the slowest replies, not just the median
failure.cause- line problems versus conversation problems
Bring a call from your own traffic.
Pick the use case closest to yours and send a recording. We run it through Tenms with your engineers and compare the events and summaries with what your pipeline does today.