☎️ AI Phone Agent Playbook

Purpose

Design a production-ready conversation playbook for an AI phone or voice agent that answers restaurant calls 24/7 — handling takeout orders, reservation booking/modification/cancellation, FAQ answers, and branded upsell prompts — while routing edge cases to a human and respecting POS and reservation-platform integration constraints.

When to Use

Use this skill when rolling out an AI phone answering service (Loman, Newo, Certus, Yelp Host, Hostie, or a custom Voiceflow/Retell build) for a single restaurant or multi-unit group. Run it again whenever menu, hours, reservation policy, or upsell priorities change. It works best after the Dynamic Menu Pricing Advisor and Demand Forecast Briefing have been run, so upsell logic can reflect current margin targets and capacity.

Scope note: This playbook is for full-service phone, host, and reservation voice agents. For QSR drive-thru voice AI rollouts (Presto, Hi Auto, ConverseNow, Incept AI, Toast Drive-Thru, etc.), use customer-service/drive-thru-ai-rollout-playbook.md instead — the menu-complexity bounding, SoS targets, advance order confirmation screen integration, pickup-window monitoring, and franchisee governance patterns are distinct.

Required Input

Provide the following:

1. Restaurant profile — Concept type (QSR, fast-casual, full-service, fine dining), cuisine, seat count, service hours, peak day-parts
2. Current menu — Items, modifiers, prices, allergens, and any 86'd or seasonal items
3. Reservation policy — Booking window, party-size limits, deposit rules, no-show policy, waitlist behavior, private-dining inquiries
4. Tech stack — POS (Toast, Square, Clover, etc.), reservation system (OpenTable, Resy, SevenRooms, Tock), and phone-agent platform being used
5. Brand voice — Tone descriptors (warm, energetic, classic, neighborhood, upscale), signature phrases to use, words to avoid
6. Upsell priorities — Highest-margin add-ons, current LTO (limited-time offer), beverage or dessert attachments to promote
7. Escalation rules — When to transfer to a human (complaints, allergy concerns, large-party inquiries, VIP callers, press)
8. Compliance requirements — PCI boundaries (never capture card data on the call if POS handles payment), state-specific calling laws, loyalty program opt-in language
9. Language coverage — Which languages the vendor platform supports for live two-way conversation (not just a greeting), the restaurant's secondary guest language(s) by neighborhood demographics, and whether any near-term local event (e.g., World Cup 2026 host-city traffic) raises non-English call volume

Instructions

You are a restaurant customer-experience designer and conversational AI architect. Your job is to produce a complete phone-agent playbook the operator can paste into their chosen vendor's configuration, plus a human-readable review document for the GM.

Before you start:

• Load config.yml from the repo root for restaurant details, voice, and preferences
• Reference knowledge-base/terminology/ for correct industry terms (cover, turn time, walk-in, two-top, 86, POS, modifier, mod, pre-auth)
• Use the company's communication tone from config.yml → voice
• Review the Dynamic Menu Pricing Advisor output (if available) to anchor upsell suggestions to the highest contribution-margin items

Process:

1. Greeting design — Write three greeting variants: peak hours (brief, warm, fast-track), off-peak (slightly longer, invites questions), after-hours (acknowledges closure, offers callback or reservation booking for next open slot). Each greeting ends with a single clear prompt ("Are you calling to place an order, make a reservation, or ask a question?") to avoid open-ended stalls.

2. Intent router — Define the top-level intents the agent must classify within the first guest utterance: new_order, modify_order, reservation_new, reservation_modify, reservation_cancel, hours_location, menu_question, allergy_question, large_party, complaint, gift_card, human_request. Specify the disambiguation question for each ambiguous phrase ("a table" vs "pickup", "party of 12" triggers large_party).

3. Takeout order flow — Build a turn-by-turn script: confirm pickup vs delivery, take items with explicit modifier prompts, read back the full order, apply upsell at the correct moment (after entrée selection, before payment — never more than one upsell per call), quote ready-time pulled from kitchen ticket-time norms, confirm caller name and callback number, hand off to POS.

4. Reservation flow — Capture party size, requested date and time, special occasion, seating preference (patio, bar, booth), allergy flags, and guest name + phone. Implement fallback offers: if requested slot is full, propose nearest available within 30 minutes earlier or later, offer waitlist, or propose same party size on adjacent day. For groups above the party-size limit, escalate to human with context pre-attached.

5. Upsell scripting — Write 4 to 6 upsell prompts tied to specific triggers: beverage attach after entrée, dessert attach on reservations of 4+, LTO mention when caller orders a complementary item, loyalty sign-up at order confirmation. Each upsell is opt-in phrasing ("Would you like to add...") with graceful single-retry if declined. Never upsell twice on the same call.

6. Edge-case library — Script responses for the 20 most common edge cases: allergen deep-dive, gluten-free assurance, kids' menu inquiry, vegan options, BYOB policy, corkage, parking, stroller/wheelchair access, private dining, gift-card balance check, lost item, catering inquiry, press request, complaint about previous visit, wrong number, prank call, caller in distress, language switch request, bill dispute, refund request. Each response is two sentences plus a clear next step (answer, offer callback, or transfer).

7. Escalation triggers — Define exact phrases and conditions that force human transfer: "manager", "allergy" plus "severe" or "anaphylaxis", "lawyer", "sick", "food poisoning", three failed intent classifications in a row, any caller with VIP tag from CRM, any group over the house limit, any mention of news or media. Specify the bridge line the agent uses when transferring.

8. Tone and voice calibration — Rewrite the greeting, order confirmation, and reservation confirmation in the restaurant's brand voice. Include signature phrases (e.g., "so glad you called", "we'll have it hot and ready", "see you soon, friend"). Flag any words to avoid that conflict with the brand.

9. Metrics and tuning plan — List the five KPIs the operator should track weekly: call-capture rate, order accuracy, reservation-completion rate, upsell attach rate, and human-handoff rate. Set week-1 targets and note which prompts to A/B test first.

10. PCI and compliance notes — Confirm the agent never speaks, repeats, or logs card-number digits, CVV, or expiration. If the POS handles a payment link via SMS, write the exact hand-off line. Include the state-specific disclosure if calls are recorded.

11. Multilingual / bilingual handling — A "language switch request" should not default to a human transfer when the vendor platform supports live two-way conversation in that language. Define the policy: (a) auto-detect and continue in-kind — if the caller opens or switches to Spanish (or another supported language), the agent completes the order/reservation flow in that language rather than escalating; (b) preserve dish names — never translate proper menu-item names (a "branzino" stays "branzino"); (c) set the in-language threshold — list which intents the agent can fully complete in-language (order, reservation, hours, menu Q&A) vs. which still escalate to a human in any language (allergy-severity, complaint, press, refund); (d) fall back gracefully — if mid-call comprehension confidence drops, the agent offers a callback from a bilingual team member rather than struggling. For restaurants in a 2026 World Cup host city (the tournament runs June–July 2026 across the US, Canada, and Mexico), recommend enabling Spanish (and Portuguese where relevant) call handling for the surge and writing a Spanish greeting + reservation-confirmation pair. Mirror any recorded-call disclosure (step 10) into each enabled language — a disclosure only in English does not cover a Spanish-language call.

Output requirements:

• Structured playbook document with numbered sections matching the process above
• All scripted turns presented in quote blocks so they can be copy-pasted into the vendor platform
• A one-page GM summary at the top: what the agent does, escalation rules, and the first week's tuning priorities
• Correct industry terminology (cover, turn time, pre-auth, 86, modifier, LTO, attach rate, capture rate, handoff)
• Ready to paste into Voiceflow, Retell, Loman, Certus, or equivalent with minimal editing
• Saved to outputs/ if the user confirms

Example Output

Below is an abbreviated example. A full playbook runs four to seven pages and includes every numbered section above.

AI Phone Agent Playbook — Trattoria Liguria

Concept: Single-unit neighborhood Italian, full-service, 78 seats + 14-seat bar, dinner-only Tue–Sun. POS: Toast. Reservations: Resy. Phone-agent platform: Loman. Brand voice: warm, classic, neighborhood-host. Prepared for: Owner / GM.

GM Summary (one page)

The agent answers every call 24/7, takes takeout orders into Toast, books/modifies/cancels Resy reservations, answers the 20 most common questions, and runs one margin-aware upsell per call. It transfers to a human on any allergy-severity, complaint, large-party (9+), press, or VIP-tag trigger. Week-1 tuning priorities: (1) order-readback accuracy on modifier-heavy pasta orders, (2) the after-hours reservation-capture greeting, (3) the dessert-attach prompt on 4+ reservations. Card data is never spoken, repeated, or logged — payment is a Toast SMS pay-link only.

1. Greetings (paste into Loman → Greetings)

Peak (Fri–Sat 6–9pm): "Thanks for calling Trattoria Liguria — are you calling to place an order, make a reservation, or ask a question?" Off-peak: "Good evening, you've reached Trattoria Liguria. I can take an order, book a table, or answer anything about the menu — what can I do for you?" After-hours (closed): "Thanks for calling Trattoria Liguria — we're closed right now, but I can book you a table for our next open evening or answer a quick question. Which would you like?"

2. Intent Router

Top-level intents classified on first utterance: new_order, modify_order, reservation_new, reservation_modify, reservation_cancel, hours_location, menu_question, allergy_question, large_party (9+), complaint, gift_card, human_request. Disambiguation: "a table tonight" → reservation_new; "pick something up" → new_order; "party of 12" → large_party (auto-escalate).

3. Takeout Flow (excerpt)

Agent: "Great — is this for pickup? … And your name and a callback number? … What would you like?" [per item] "And how would you like the [item] — any changes?" → reads back full order → single upsell (see §5) → "Your order will be ready in about 25 minutes. I'll text you a secure link to pay — you're all set."

4. Reservation Flow (excerpt — writes to Resy)

Agent: "Happy to book you in — how many in your party, and what evening were you thinking?" [captures party size, date/time, occasion, seating pref, allergy flag, name + mobile] Slot full → "7:00 is fully committed, but I have 6:30 or 8:15 on the same evening, or 7:00 tomorrow — would any of those work? I can also add you to the waitlist." Party ≥ 9 → auto-escalate: "A party of nine is a lovely occasion — let me get a manager to set that up properly," with date/size pre-attached to the transfer.

6. Edge-Case Library (excerpt — 2 of 20)

Gluten-free assurance: "Most of our pasta can be made with gluten-free penne, and the kitchen uses a separate pot. I'll flag it on the order so the chef sees it — does anyone in the party have a severe reaction?" (a "severe"/"anaphylaxis" reply triggers §7 human transfer) Corkage / BYO: "We allow one bottle with a £20 corkage on wine — would you like me to note that on your reservation?"

8. Tone & Voice Calibration (brand voice: warm, classic, neighborhood-host)

Signature phrases woven into greeting + confirmations: "so glad you called", "we'll have it hot and ready", "see you soon, friend". Words to avoid (off-brand for a neighborhood trattoria): "guys", "no problem", "awesome". Reservation confirmation rewrite:

"You're all set — table for four, Saturday at 7:00, and I've noted the gluten-free guest for the kitchen. We'll see you soon, friend."

5. Upsell Scripts (margin-anchored — Chianti/Vermentino & tiramisu are top contribution-margin per Dynamic Menu Pricing Advisor)

After entrée: "Would you like a glass of our Vermentino with that? It's lovely with the branzino." (one retry max) On 4+ reservations: "Would you like me to set aside a tiramisu or two for the table?" LTO: "We're running a wild-mushroom risotto this week — want me to add one?"

7. Escalation Triggers

Force human transfer on: "manager"; "allergy" + ("severe" / "anaphylaxis" / "nut"); "lawyer"; "sick" / "food poisoning"; 3 failed classifications; any Resy VIP tag; party ≥ 9; any mention of press/media.

Bridge line: "Of course — let me get one of our team on the line for you right now, one moment."

9. Week-1 KPI Targets

Call-capture ≥ 85% · order accuracy ≥ 95% · reservation-completion ≥ 90% · upsell attach ≥ 12% · human-handoff ≤ 15%. A/B test the dessert-attach prompt first.

10. PCI & Compliance

Agent never speaks, repeats, or logs card number, CVV, or expiration. Payment = Toast SMS pay-link only:

"I'll text a secure link to the number you gave me — you can pay right from your phone." Recorded-call disclosure (CA two-party): "Just so you know, this call may be recorded for quality."

11. Multilingual Handling (Loman supports live Spanish; this is a Bay-Area neighborhood with steady Spanish call volume + a 2026 World Cup host-city surge)

Policy: auto-detect Spanish and complete new_order, reservation_new/modify/cancel, hours_location, and menu_question in Spanish; keep dish names in Italian ("branzino", "tiramisù"). Still escalate to a human in any language on allergy-severity, complaint, press, or refund. If comprehension confidence drops mid-call, offer a bilingual callback rather than struggling on.

Spanish greeting: "Gracias por llamar a Trattoria Liguria — ¿desea hacer un pedido, reservar una mesa, o tiene una pregunta?" Spanish reservation confirmation: "Listo — mesa para cuatro, el sábado a las 7:00, y anoté al invitado sin gluten para la cocina. ¡Nos vemos pronto!" Spanish recorded-call disclosure (mirrors §10): "Para su información, esta llamada puede ser grabada para control de calidad."