AI receptionist setup guide

Skip these and you'll either rebuild work twice or launch sounding like every other AI receptionist on the market. A 60-90 minute internal session covers most of it.

Capture brand voice samples

Pull 5–10 examples of your team answering the phone — voicemail recordings, sales calls, training videos. The AI's voice should sound like your highest-performing team member, not like the vendor's stock template. If you don't have recordings, write 3 sample dialogues in the tone you want.

List your top 20 customer questions

Your front-desk staff already knows them. Most are some flavor of pricing, hours, location, services, scheduling, or status updates. The AI needs to answer all 20 without escalating to a human — that's where 70%+ of the call deflection value lives.

Decide your escalation triggers

Before the AI launches, write down the exact phrases or situations that should bypass the AI and reach a human immediately. Common ones: 'I'd like to speak to a manager,' 'this is an emergency,' billing disputes, anything legal. The AI's escalation rules should match this list.

Most vendors give you a fresh tenant within minutes of signup. The standard checklist:

1. 01Verify the billing plan matches what you signed up for (overage rates are where surprises live)
2. 02Set the time zone correctly — wrong time zone breaks after-hours rules
3. 03Configure the business hours schedule, including holidays and lunch breaks if relevant
4. 04Add team members with the right roles (admin / editor / viewer)
5. 05Connect webhooks or native CRM integration (more in the integration guide)
6. 06Set up call recording with the right retention period (HIPAA-friendly setups need encryption + BAA review)

You have three options: a new number from the vendor, porting your existing number, or call forwarding from your current line. Pick by what risk you're willing to take during launch.

Option A — new number from vendor (lowest risk)

Best for first launch. You point a small percentage of inbound to the new number (e.g. add it to your Google Business profile alongside your main line), test for 1–2 weeks, then port if you're happy.

Option B — call forwarding (medium risk)

Your existing phone system forwards inbound calls to the AI's number. Easy to roll back; you can flip forwarding off in 30 seconds. Pay attention to caller-ID display — some carriers strip it during forwarding, which breaks CRM lookups.

Option C — port your existing number (highest risk, cleanest end state)

Once you're confident, port. Porting takes 5–10 business days and there's a brief porting window where calls can drop. Don't port the week of launch — port after 2 weeks of stable testing.

This is where most launches stumble. The default template handles maybe 60% of your real calls; the other 40% need explicit branching. The fastest way to get to a working flow:

1. 01Start with the greeting (one line, on-brand, mentions you're recorded)
2. 02First branch: emergency or routine — emergency routes immediately to on-call
3. 03Routine branch: new customer or existing — new customers go to qualification, existing to lookup-then-route
4. 04Qualification branch: capture the 3–5 fields you actually need (intent, service area, timeline, contact, anything required for routing)
5. 05FAQ branch: the AI fields the top-20 questions from pre-flight without escalating
6. 06Booking branch: if booking is in scope, the AI offers a slot and locks it
7. 07Always-on fallback: 'speak to a person' should always work, every state of the call

Aggressive escalation makes the AI feel useless; lazy escalation makes the AI sound trapped. Find the middle. Recommended escalation triggers:

• Explicit request ('speak to a human', 'manager', 'representative')
• Strong negative sentiment (frustration, swearing, multiple repetitions)
• Off-script questions the FAQ branch doesn't cover
• Specific topics: billing dispute, complaint, legal, refund
• High-value caller (VIP customer detected via CRM lookup)
• Emergency keywords (configurable to your business)

For after-hours escalation, the AI should ALWAYS capture the contact info and reason before escalating — even when it can't transfer immediately. That way the morning team has context, not just 'someone called urgent at 11pm.'

The vendor's test mode is fine; better is testing against a parallel real number before cut-over. 5–10 real-feeling calls covering each major branch:

1. 01Standard new-customer inquiry (most common path)
2. 02Existing-customer call (test CRM lookup if integrated)
3. 03Booking call (test calendar integration)
4. 04FAQ call — try a few of the top-20 questions
5. 05Off-script call — ask something the AI shouldn't be able to handle, verify escalation
6. 06After-hours call — test the after-hours flow
7. 07Emergency call — verify emergency routing works
8. 08Cancellation / reschedule — if booking is in scope
9. 09Spanish-language call — if multilingual is in scope
10. 10Aggressive caller — test the 'speak to a human' fallback under pressure

Once you've tested 10 real-feeling calls and they sound right, point your production number at the AI. Plan to spend ~30 minutes per day for the first two weeks listening to real call recordings and tweaking prompts.

What to listen for in week 1

• Calls that escalated when the AI could have handled them → tighten escalation triggers
• Calls where the AI escalated late (frustrated customer) → loosen escalation triggers
• Awkward phrasing or mispronunciations → edit prompts
• Calls that ended without a clear next step → add a closing line to the flow

What to measure in week 2

• Containment rate (% of calls the AI handled without escalation) — target 60-80%
• Booking rate (if booking is in scope) — should approach your previous staffed rate
• Average call duration — drift up over a few days usually means the AI is over-explaining
• Customer-initiated escalations — these are the AI's failures; investigate each