The shape of a real call flow
Most useful flows have 4–7 top-level branches. More than that and the AI gets confused; fewer and you're missing call types. The canonical shape:
- 01Greeting + recording disclosure (one line, on-brand)
- 02Urgency check: emergency or routine?
- 03Customer status: new or existing? (CRM lookup if integrated)
- 04Intent capture: what does the caller need? (qualification or FAQ)
- 05Resolution: book the appointment / answer the question / escalate to human / end with promised follow-up
- 06Always-on fallback: 'speak to a person' works in every state
Each branch has its own sub-flow with the prompts, qualification questions, and escalation rules specific to that path. The AI should be able to switch between branches mid-call (a customer who started as 'routine new customer' might surface an emergency halfway through).
Qualification: capture the 3–5 fields you actually need
More qualification questions = lower completion rate. Pick the fewest that your sales/dispatch team can't live without:
- Service area / location (gates whether you can take the call at all)
- Service or matter type (decides routing and rep assignment)
- Urgency or timeline (decides priority)
- Contact info (so you can follow up if the call drops)
- Decision-maker status (for high-stakes verticals like remodeling, legal)
Phrase questions naturally
'What's the address of the property?' beats 'Please provide your service address.' Vendors that let you customize prompt phrasing have a big advantage — use it.
Skip what you can derive
If you have the caller's phone number and they're an existing customer in the CRM, you already have the address — don't ask again. Use CRM lookup to skip redundant questions.
FAQ deflection: where most of the deflection value lives
70–80% of inbound calls at a typical SMB are one of: 'what are your hours,' 'how much does X cost,' 'do you service Y area,' 'when's my appointment,' 'can I reschedule.' If the AI handles these without escalating, your front-desk staff gets their time back.
Build the FAQ from real call data
Pull 30 days of voicemail or call recordings. Sort by frequency. The top 20 questions are your FAQ. Don't write a generic FAQ — yours is specific to your business, your hours, your services.
Write answers in natural sentences, not docs
'Our hours are Monday through Friday, 8am to 6pm, and Saturday 9am to 2pm. We're closed Sunday.' beats 'Hours of operation: M-F 8am-6pm; Sat 9am-2pm; Sun closed.' The AI reads what you write.
Escalation: the trigger list
Escalation that's too aggressive makes the AI feel pointless; escalation that's too lazy makes the AI feel like a wall. Recommended triggers, in priority order:
- 01Explicit request ('speak to a human', 'manager', 'representative', 'agent')
- 02Strong negative sentiment (frustration, repetition, raised voice)
- 03Specific high-risk topics (billing dispute, complaint, legal threat, refund)
- 04VIP customers (detected via CRM lookup; configurable per business)
- 05Off-script questions the FAQ branch can't handle
- 06Emergency keywords (define per business: 'leak', 'fire', 'urgent', 'pain' for medical, etc.)
- 07Caller-detected ambiguity after 2 failed clarification attempts
Edge cases worth designing for
Language switch mid-call
Caller starts in English but is more comfortable in Spanish. Better vendors detect this and offer to switch; lesser vendors miss it entirely. If your customer base is meaningfully multilingual, this is a must-design.
Background noise and bad lines
Construction sites, busy kitchens, driving — your real callers won't always be in quiet rooms. The AI should ask the caller to repeat (politely) when it can't parse, not pretend it heard correctly.
After-hours emergency vs after-hours routine
After-hours flow shouldn't be a flat 'we're closed.' Real emergencies need the on-call escalation; routine after-hours should book the next-available morning slot.
Call recording disclosure
Two-party-consent states require explicit consent for recording. Most AI vendors handle the disclosure phrase; review the default language with your counsel before launch.
Iterate post-launch
No call flow is right on day one. Plan to spend ~30 min/day for the first 2 weeks listening to recordings and adjusting:
- Calls that escalated unnecessarily → tighten escalation criteria for that case
- Calls where the AI sounded awkward → edit the prompt for that branch
- Calls where the customer asked something not in the FAQ → add it
- Calls where qualification took too long → cut a question
- Calls where qualification missed context → add a question (carefully)
Checklist
The version you can copy.
Top-level branch structure documented (4–7 branches)
Qualification questions narrowed to 3–5 critical ones
FAQ built from real 30-day call data (top 20)
Escalation triggers written out + tested
Edge cases designed: language switch, after-hours, recording disclosure, legal/medical advice
All prompts written in your brand voice (not vendor default)
Always-on 'speak to a human' fallback tested in each branch
Post-launch tuning calendar blocked for 2 weeks