Skip to content
Welzin
Case study · Voice AI

A voice agent that answers, books, and closes the loop.

A national home-services brand was losing calls to hold times and after-hours gaps. Welzin designed and shipped a production voice agent that talks naturally, takes real action, and writes every outcome back to the systems the business already runs on.

RoleDiscovery → design → build → run
Timeline14 weeks to launch
StackLLM + streaming STT/TTS, telephony, RAG, evals
01 / The challenge

Every missed call is a job that goes elsewhere.

Home-services demand spikes are unpredictable and the phone is still the front door. When callers hit hold music, an empty after-hours line, or an answer that contradicts the last one they got, they hang up and dial a competitor. The brand needed coverage that never tired and never went off-script.

01

Long hold times

Peak-hour queues pushed callers past their patience, and abandoned calls turned into bookings for someone else.

callers hang up
02

After-hours gaps

Emergencies and evening research happen off the clock, when no agent is on the line to capture the lead.

03

Inconsistent answers

Pricing, coverage areas and policies drifted between reps, eroding trust and creating costly follow-up calls.

04

Cost per call

Staffing for the busiest hour meant paying for idle capacity the rest of the day, with overtime stacked on top.

rising spend
02 / What we built

A voice agent that listens like a person and acts like a system.

Six capabilities work together so the agent holds a natural conversation, grounds every answer in the client's own knowledge, and follows through with real actions, never an empty promise.

[ 01 ]

Low-latency turn-taking

Streaming STT and TTS with predictive endpointing keep replies fast and the back-and-forth feeling human, not transactional.

[ 02 ]

Barge-in handling

The caller can interrupt mid-sentence and the agent stops, listens, and adjusts, just like a good rep would.

[ 03 ]

RAG-grounded answers

Responses are retrieved from the client's knowledge base - pricing, coverage, policies - so callers get one consistent source of truth.

[ 04 ]

Action via tool calls

The agent books appointments, updates the CRM, and takes payments through tool calls against the brand's live systems.

[ 05 ]

Guardrails & evals

Policy guardrails and an eval suite stop the agent from inventing prices, dates, or commitments it cannot honor.

[ 06 ]

Full observability

Every call ships transcripts, latency traces, and QA scoring, so the team can see, measure, and improve what happens on the line.

03 / How a call flows

Two paths, one disciplined conversation design.

Inbound and outbound calls follow purpose-built scripts with clear states. The agent always knows where it is, what it still needs, and where the conversation should land.

Inbound

Caller reaches the brand and wants something resolved now.

  1. GreetWarm, on-brand opening that sets expectations in seconds.
  2. IntentListen and classify what the caller actually needs.
  3. VerifyConfirm identity and account against the CRM.
  4. Resolve or bookAnswer from the knowledge base, or schedule a visit live.
  5. Wrap-up + CRM writeRecap, confirm, and write the outcome back automatically.

Outbound

Agent follows up on a lead or a scheduled callback.

  1. ReachPlace the call at a sensible time and confirm who is on the line.
  2. QualifySurface the need, budget, and timing before going further.
  3. ScheduleOffer real open slots from the connected calendar and lock one in.
  4. HandoffPass qualified, context-rich detail to a human when it matters.
04 / Integrations

Wired into the tools the business already runs.

The agent is only as useful as the systems it can reach. We connected it to telephony, CRM, calendars, and data so it can both know and do, in the same call.

Telephony

Twilio Telnyx SIP trunking

CRM

HubSpot Salesforce

Calendars

Google Calendar Outlook

Data & actions

Webhooks REST APIs

Observability

Transcripts Latency traces QA scoring
05 / The impact

Calls answered, jobs booked, cost taken out.

After launch the agent took the routine volume off the queue and handed the team cleaner, qualified work. Figures below reflect observed outcomes in production and are directional, not guarantees.

0%
of calls auto-resolved end to end, with no human needed.
<0ms
median response latency, fast enough to feel like a real conversation.
0/7
coverage, so after-hours and overflow calls are captured instead of lost.
0%
lower cost per call versus the prior staffing-for-peak model.
It picks up on the first ring at midnight, books the slot, and the notes are already in our CRM before the next caller. It feels like we hired a whole shift.
- Operations lead, national home-services brand (client confidential)

Want an outcome like this one?

Talk to the senior pod that shipped it - we answer within one business day.

Talk to us

Prefer email? Write to us directly.