Voice-to-Estimate (Universal)

Service trade businesses with two pain points: (1) leads call after hours, leave voicemails, and never get a response before they call the next contractor on Google; (2) the owner/lead tech goes on-site, diagnoses the problem in 20 minutes, then spends 2-3 hours that night typing up a formal estimate that should have taken 5 minutes. This tool fixes both. Particularly strong fit for: HVAC, plumbing, electrical, roofing, appliance repair, garage door, landscape, cleaning, auto repair.

The pitch lands hardest with one-truck operators who do their own estimating (their time IS the bottleneck) and small shops doing $500k-3M annually where the owner can't afford a $60k/year estimator.

The system accepts three types of voice input and turns each into a customer-ready estimate within minutes: PATH A, Customer voicemail (most common trigger): customer calls the business after hours, leaves a voicemail describing their problem. The system's Twilio number forwards voicemails to a transcription endpoint. The AI extracts the customer's name, contact info, address, problem description, equipment details, urgency.

It then queries the vertical-specific Pricing Knowledge Base to generate a ranged estimate (e.g., "$185-$340 likely"). The owner sees the draft estimate in their phone within 5 minutes, taps APPROVE or EDIT, and the customer receives a professional SMS quote before they've finished calling competitors. PATH B, Contractor voice memo: after a site visit, the contractor opens the admin app on their phone, taps "Record Estimate", and speaks: "Sarah Johnson, garbage disposal replacement, recommend Insinkerator Badger 5, plus new mounting hardware, 90-minute job." The AI structures this into a complete estimate with line items, total range, and customer-facing language.

Owner reviews and sends. PATH C, AI Receptionist call transcript: if the client is also using AI Receptionist (Tool #8), call transcripts flow directly into this tool's queue. No voicemail transcription step needed, the AI Receptionist already structured the data.

Every estimate goes through an owner approval gate before the customer sees it (legal/liability firewall). Once approved and sent, the estimate auto-flows into Quote Follow-Up Sequencer (Tool #10) if the client also uses that tool. The customer journey becomes: customer calls → voicemail transcribed → estimate generated → owner approves → SMS sent within 10 min → follow-up sequence runs over 21 days → closed deal.

This tool has more moving parts than most: voicemail recording, audio transcription, LLM estimate generation, owner approval queue, SMS delivery, and optional cross-tool integration. Expect Base44 to build in 20-30 minutes.

Likely follow-up prompts:

If Whisper transcription quality is poor: "Whisper is returning low-quality transcriptions on voicemails. This is often due to audio quality from Twilio. Verify: (1) Twilio recording configuration uses adequate sample rate (16kHz minimum), (2) we're using whisper-1 model, (3) we're providing the language='en' parameter, (4) we're providing prompt context like 'This is a voicemail from a customer describing a home service issue' to bias the model."

If estimate ranges are wildly off: "AI estimates are coming back with ranges that don't match the Pricing Knowledge Base. Audit the LLM prompt, the Pricing Knowledge Base should be loaded as system context, not just as one-shot examples. The LLM must reference specific PricingKnowledgeBase entries when generating the estimate.

Add to the prompt: 'You MUST select your estimate range from the entries in the Pricing Knowledge Base above. Do not invent prices outside this database. If the problem doesn't match any entry, return confidence=low and recommend a diagnostic visit.'"

If owner approval queue isn't updating in real-time: "Owner approval queue needs to push notifications immediately when new estimates are ready. Send {{ownerName}}'s personal cell an SMS the moment a new estimate lands: '📞 New estimate ready: [customer name] - [problem summary] - [estimate range]. Tap to approve: [link]'."

If cross-tool integration with AI Receptionist isn't working: "The importFromReceptionist function isn't pulling new call transcripts. Verify: (1) the linked_ai_receptionist_workspace_id is set in Settings, (2) the AI Receptionist's CallLog entity is readable from this workspace, (3) the scheduled function is running every 15 minutes. The integration only works if both tools share a Base44 workspace OR if cross-workspace API access is configured."

If estimates are being sent before owner approval: "CRITICAL BUG: estimates are auto-sending without owner approval. Audit the approval_status field, by default it MUST be 'pending' and the SMS send function MUST check approval_status='approved' or 'edited_approved' before firing. Default auto_send_threshold_confidence in Settings MUST be OFF."

The biggest gotcha: audio quality. Twilio voicemail recordings can have variable quality (background noise, mumbling, cell coverage). Set realistic expectations, Whisper handles ~85% of voicemails cleanly; the other 15% will have transcription errors that the owner sees and corrects in the approval queue.

The second biggest gotcha: estimate confidence calibration. The AI will sometimes over-confidently quote a $200 capacitor on what's actually a $4,000 compressor issue. The Pricing Knowledge Base ambiguity_flags help prevent this, but the owner approval gate is the real safety net.

NEVER skip the approval step.