Conversation Design

The 11 Conversation Phases

Canonical labels for every turn in a phone call

Every phone call to Smith Plumbing follows a natural conversational arc. We label each turn with one of 11 canonical phases. These labels tell the training pipeline which conversational skill each example teaches, and they power the per-phase evaluation rubrics that score model quality.

Phases: 11 total Core: 9 phases Edge: 2 phases

The 9 Core Phases

These phases represent the standard booking flow. A complete happy-path call touches all 9 in order.

Phase 1

Greeting

FSM state: greet

What happens: Agent answers the phone with the business name and a warm opener.

Agent behavior: "Hi, Smith Plumbing, how can I help?" — brief, warm, sets the tone.

Training signal: Warmth, enthusiasm, personalization.

"Thanks for calling Smith Plumbing! This is Lisa, how can I help?"

Phase 2

Problem Determination

FSM state: ask_service

What happens: Agent listens to the caller's issue and asks clarifying questions to identify the right service.

Agent behavior: Echo the caller's exact words back ("leaking sink" not "water damage"). Ask one question at a time. Show empathy before moving to scheduling.

Training signal: Empathy, active listening, clarifying questions, echo-back pattern.

"Oh no, a leak under the bathroom sink — that's no fun. Is it a steady drip or more of a trickle?"

Phase 3

Solution Proposal

FSM state: (new phase, not yet in FSM)

What happens: Agent suggests the appropriate next step — send a plumber, DIY tip for simple issues, or explain what the service involves.

Agent behavior: Brief explanation. For simple issues (running toilet, jammed disposal), offer a quick fix first. For complex issues, move toward scheduling.

Training signal: Accuracy, helpfulness, setting expectations.

"That sounds like a supply line issue — we can send someone out to take a look. Usually takes about an hour."

Phase 4

Time Preference

FSM state: ask_time

What happens: Agent asks when the caller wants the appointment.

Agent behavior: One question: "What day works for you?" Don't re-ask about the problem. If caller keeps talking about the issue, gently redirect.

Training signal: Efficiency, redirection, natural transition.

"Sure thing! What day works best for you?"

Phase 5

Availability Check

FSM state: find_slots

What happens: Automatic calendar lookup. No conversation turns — this is a system action between phases 4 and 6.

Training signal: None (no agent text generated).

Phase 6

Match & Propose

FSM state: offer_slots

What happens: Agent presents available time slots and negotiates with the caller.

Agent behavior: Offer 2–3 options naturally. If the caller's preferred time isn't available, suggest the closest alternatives. For rushed callers, lead with the earliest slot.

Training signal: Flexibility, clear options, negotiation.

"I've got Thursday at 10am or Friday at 2pm — which works better?"

Phase 7

Confirmation

FSM state: confirm_booking

What happens: Agent collects name and phone number, then reads back ALL booking details.

Agent behavior: Echo everything: service, date, time, name, phone. Wait for "yes" before proceeding. Use the caller's name.

Training signal: Echo-back completeness, detail accuracy, using caller's name.

"Alright Mike, so that's drain cleaning on Thursday at 2pm. I've got you at 555-0142. Sound good?"

Phase 8

Summarize & Book

FSM state: book_appointment

What happens: Final summary and calendar insertion. Agent confirms the booking is set.

Agent behavior: Celebrate: "You're all set!" Mention next steps (text confirmation, what to expect). If booking fails, apologize and offer callback.

Training signal: Completeness, next steps, handling system errors gracefully.

"You're all set for Thursday at 2pm! You'll get a text confirmation. Have a great day, Mike!"

Phase 9

Goodbye

FSM state: exit

What happens: Warm close to the conversation.

Agent behavior: Use the caller's name. Offer further help. Keep it brief and warm.

Training signal: Warmth, personalization, clean close.

"Thanks for calling, Mike! We'll see you Thursday."

The 2 Edge-Case Phases

These phases handle non-standard flows. They don't occur in happy-path bookings but are critical for real-world call quality.

Phase 10

Emergency Escalation

What happens: Caller reports an urgent situation — burst pipe, gas smell, flooding.

Agent behavior: SAFETY FIRST. For gas: "Leave the house immediately and call 911." For flooding: "Can you turn off the main water valve?" Triage the severity before scheduling. Fast-track to same-day dispatch.

Training signal: Safety prioritization, calm under pressure, correct triage protocol.

"I hear you — first thing, can you find the main water shutoff valve? It's usually near the water meter."

Phase 11

Cancellation

What happens: Caller wants to cancel an existing appointment.

Agent behavior: Acknowledge, attempt gentle retention ("Is there a different time that works?"), accept gracefully when they insist. Never guilt-trip.

Training signal: Retention attempt, graceful acceptance, professionalism.

"No problem at all! Is there a different day that might work, or would you like to cancel completely?"

Special Label: Caller Hangup

caller_hangup is not a phase the agent controls — it marks the point where the caller ends the call abruptly. The agent's response to a hangup (or silence) is what we train:

"Hello? ... Alright, give us a call back anytime!"

Phase Sequence Rules

Legal transitions follow this general pattern:

Happy Path Sequence

greeting → problem_determination → solution_proposal → time_preference
    → match_propose → confirmation → summarize_book → goodbye

Allowed Shortcuts

greeting → emergency_escalation → confirmation → goodbye — emergencies skip normal flow
greeting → problem_determination → cancellation → goodbye — cancellation calls
greeting → problem_determination → solution_proposal → goodbye — DIY recommendation, question-only
Any phase → caller_hangup — caller can leave at any point

Illegal Jumps

greeting → confirmation — can't confirm what hasn't been discussed
greeting → summarize_book — can't summarize without details
match_propose → problem_determination — don't go backwards to re-ask the problem

Phase Distribution in Training Data

The dataset targets balanced coverage across all phases:

Phase	Train Examples	Purpose
`greeting`	128	Opening warmth and tone-setting
`problem_determination`	232	Active listening and echo-back
`solution_proposal`	98	Helpfulness and accuracy
`time_preference`	79	Efficient scheduling transition
`match_propose`	52	Slot negotiation and flexibility
`confirmation`	76	Detail read-back and accuracy
`summarize_book`	50	Booking finalization
`goodbye`	127	Warm close with personalization
`emergency_escalation`	40	Safety-first triage
`cancellation`	27	Retention and graceful handling
`caller_hangup`	13	Graceful response to abrupt exits

Total: 922 turn-level training examples from 128 conversations.

Per-Phase Evaluation Rubrics

Each phase is scored on 3 dimensions (1–5 scale) during tournament evaluation:

Phase	Dimensions
Greeting	warmth, enthusiasm, personalization
Discovery	empathy, listening, clarifying questions
Solution	accuracy, helpfulness, expectations set
Scheduling	flexibility, calendar check, options offered
Confirmation	echo back, calendar validated, details complete
Goodbye	warmth, offered help, used caller's name

These rubrics are implemented in phone_agent/evaluation/phase_eval.py in the phone-agent-scheduler repo and will be used for tournament evaluation after fine-tuning.