Voice-driven AI service agents
We are studying how technicians and managers interact with pools when voice becomes the primary interface. Our goal: let AI handle documentation, diagnostics, and scheduling so humans can focus on customer experience.
Field observations
Documentation steals 25% of a technician’s day
Ride-alongs with 14 service companies showed techs spend nearly a quarter of the day typing notes, logging chemistry, or updating CRMs. Voice frees their hands and brains, but only if the AI understands pool jargon and matches the job flow.
- Lexicon challenge: Terms like “SWG”, “backwash”, or “shock to breakpoint” must be detected reliably.
- Context fusion: Voice alone isn’t enough; we fuse telemetry, route schedules, and CRM records before answering.
- Trust: Agents narrate what they understood and ask for confirmation to avoid incorrect actions.
Surfacing hidden risk in conversation
Voice captures nuance typed notes miss. When a tech says “I’m seeing moisture around the equipment pad” or “return jets are weak again,” the agent flags leak, circulation, and safety intents instantly, pairing the transcript with telemetry so managers know whether to escalate.
Escalation detection
- • Detects phrases like “leaking”, “tripped breaker”, “gas smell”—triggers high-priority workflows.
- • Cross-checks audio sentiment and telemetry anomalies before notifying managers.
- • Creates follow-up tasks with suggested scripts and required parts.
How the agent thinks
Capture, reason, act
Our pipeline blends on-device speech models with cloud reasoning. The agent summarizes each visit, proposes actions, and enriches the service database without human typing.
Capture
Technician speaks into the Clyr mobile app or wearable; audio is streamed and transcribed with speaker diarization.
Reason
Hybrid large language models combine telemetry context (chemistry, flow, weather) with domain-specific ontologies to classify intents and detect anomalies.
Act
The agent drafts work orders, schedules follow-ups, and alerts managers—routing only high-risk items for human approval.
Pilot results
Voice AI changes the service rhythm
Over 9,400 logged voice sessions across Arizona, Florida, and Texas show that AI helpers reduce paperwork, help managers stay ahead of issues, and keep clients informed faster.
Average voice log
41 sec
Down from 3.2 min typed entries
Intent accuracy
94%
Across 28 service intent classes
Resolution speed
-36%
Time to close high-priority tasks
Manager queries
+5.4x
More proactive route reviews
Manager copilots
Ask anything about the route
Managers can ask natural language questions—“Which pools ran hot last week?”—and receive a structured answer with supporting telemetry. We pair LLM reasoning with deterministic checks to avoid hallucinations.
Prompt
“Which work orders should I follow up on before Friday?”
AI answer
- • Riverbend HOA leak remains open; valve shipped, ETA Thursday.
- • Garcia residence heater replacement waiting homeowner approval.
- • Lagoon Villas acid wash scheduled Friday 9 AM—confirm crew.
Suggested actions
Creates task queue for dispatcher, escalates any item aging beyond SLA.
What’s next
- On-device wake words: Let techs summon the agent without tapping their phone.
- Contextual memory: Recall last visit actions instantly for every pool.
- Fleet scheduling: Agent suggests optimal route adjustments based on real-time traffic and chemical demand.
Collaborate with us
Help shape the future of service ops
We’re inviting select partners to co-develop voice workflows, provide anonymized training data, and validate outcomes in real routes. Participants get early access to features and direct influence over the roadmap.