Setting up an AI receptionist is faster than provisioning a phone line — most platforms get you live in 15 to 60 minutes. The work that matters happens before you ever sign up: deciding what to delegate to AI, deciding what to keep with humans, and deciding how forwarding will route the call.
This guide is product-agnostic. We use OnCall as the reference implementation because we know it best, but the steps apply to most modern AI receptionists in 2026 — Smith.ai, MyAIFrontDesk, Goodcall, NextPhone, Bookipi, Jobber AI, RingCentral AI, and the rest. If you're still choosing between platforms, read the cost comparison first.
Step 1 — Decide what your AI should handle (15 min)
Before touching a settings page, write down what kinds of calls you actually get. Most service businesses are surprised how few categories there really are. For an HVAC shop: emergency service, scheduled service, replacement quotes, maintenance plan members, warranty calls. For a dental office: new-patient bookings, emergency toothache, existing-patient rescheduling, insurance questions. For a roofer: storm damage, insurance claims, planned replacement, warranty leak callbacks. For an auto repair shop: diagnostics, scheduled maintenance, tow inbound, fleet customers.
Now sort each category into one of three tiers:
- Pickup-and-message: AI answers, captures basic info, sends you a text. Use for low-volume edge cases or anything that needs human judgment.
- Qualify-and-book: AI answers, asks 3-5 qualifying questions, books on your live calendar, sends caller a confirmation SMS. This is where most of the business value lives.
- Full-service with CRM sync: AI answers, qualifies, books, then pushes the lead into your CRM with tags and a tagged call recording. Requires CRM integration setup; usually adds $39-99/mo to most platforms (free with OnCall Sync if you're on the $39/mo Pro tier).
For most service businesses, 70-80% of calls fit Tier 2. Don't over-engineer. Start there and expand.
Step 2 — Choose a platform (15 min)
Pricing models vary more than feature sets. The big decision is per-call vs per-minute vs flat-rate, and that depends on your call volume:
- Under 50 calls/mo, high per-call value: per-call pricing (Smith.ai is the obvious choice).
- 50-300 calls/mo, variable volume: pay-as-you-go (OnCall, MyAIFrontDesk Business tier).
- 300+ calls/mo, predictable volume: flat-rate unlimited (NextPhone, Bookipi).
- Already on a unified communications platform: their AI add-on (RingCentral AI Receptionist, GoTo).
Once you've picked a model, run your real call volume through 2-3 vendors' calculators. The platform that wins on sticker price doesn't always win on real-world cost. Side-by-side breakdown here.
Step 3 — Configure your industry profile (10 min)
Modern AI receptionists ship with industry templates. Pick the closest match to your trade — most platforms have HVAC, plumbing, electrical, dental, real estate, and a few dozen more out of the box. The template pre-loads:
- Voice / persona (a calm dispatcher voice for HVAC; a warm front-desk voice for dental).
- Qualifying questions specific to your trade (vehicle make/model for auto repair, lawn size for landscaping, insurance carrier for dental).
- Emergency keywords your AI should escalate (gas smell, sewage, no-heat-with-elderly-resident, smoke).
- After-hours behavior (capture lead vs. promise callback vs. live transfer).
- Default qualifying script and 14-day SMS+email follow-up cadence.
Don't override defaults yet — start with the template, run a week of real calls, and iterate. The defaults are battle-tested. Your custom rules will be better, but only after you've heard your AI fail at something specific. OnCall ships 23 industry templates; most platforms have between 5 and 20.
Step 4 — Build your knowledge base (15 min)
Your AI is only as good as what it knows. Three knowledge sources to load on day one:
- Your website URL. Most platforms auto-extract services, hours, pricing, and FAQs from your site. Run this first; check the extracted summary; correct anything wrong.
- Your service list and pricing. Diagnostic fees, flat-rate jobs, package pricing — anything that has a published number. Don't make the AI guess. Don't load anything that requires a human conversation (custom quotes, complex coverage).
- Your FAQ. The 5-10 questions every customer asks. Insurance carriers accepted, payment methods, service area, warranty terms, weekend hours. These are usually on your website already; copy them in.
A small, accurate knowledge base beats a large hallucination-prone one. Resist the temptation to dump 50 PDFs in. If you don't know what's in a document, your AI won't either.
Step 5 — Pick a phone number and forward your line (5 min)
The mechanical part. Most AI platforms include the first phone number free or for a small monthly fee ($10/mo on OnCall; included on Bookipi and NextPhone). Pick a local area code. Activation is usually under 5 minutes.
Then forward your existing business line to the new AI number. This is a carrier-side setting — Verizon, AT&T, T-Mobile, Bell, Rogers all have one-page guides for it. Two patterns:
- Forward all calls. Simplest setup. AI answers everything. You can still make outbound calls from your real line.
- Forward when busy / unanswered. Your line rings first, AI catches anything you don't pick up. Best for small teams who answer when they can.
Step 6 — Test it before the public hears it (10 min)
This is where most setups fall down. Don't flip the switch and walk away. Make at least three test calls before going live:
- Call yourself with a routine inquiry (a typical request your business gets). Listen end-to-end.
- Call from a coworker's phone with an emergency-keyword scenario specific to your industry. Verify the escalation actually triggers.
- Call from your own phone with a question that should NOT be answered by AI (something complex, requiring human judgment). Verify the AI captures the lead and routes correctly instead of guessing.
Listen to the call recordings on your dashboard. Most platforms (including OnCall) record by default with disclosure. Hearing your AI answer your calls in your voice is the only test that matters.
Common setup mistakes
Five mistakes we see in nearly every first-week pilot. None of these is fatal. Avoiding them saves you a week of iteration.
Mistake 1 — Vague qualifying questions
"Tell me about the issue" is a bad qualifying question. "Is the unit running but blowing warm, or not turning on at all?" is a great one. Specific binary or short-list questions get specific answers; open-ended questions get rambling answers your AI can't parse. Replace every open-ended question in your default script with something more specific.
Mistake 2 — No after-hours behavior set
Most platforms default to "capture lead" after hours. For HVAC, plumbing, towing, water-damage, and other emergency-prone trades, that's wrong. You want "promise callback" or "live transfer to on-call tech." Configure this on day one. The 11pm furnace-died call is the entire reason you bought AI in the first place.
Mistake 3 — Ignoring KB curation
The auto-extract from your website is a starting point, not the final answer. Spend 15 minutes after the extract reviewing what your AI thinks it knows. We've seen pilots where the AI confidently quoted "starting at $19/sqft" because the website had an old roofing flyer; the real number was $4-7/sqft. Verify what got pulled.
Mistake 4 — Skipping voice testing
Voices that sound great on a marketing demo can sound robotic on your specific traffic — accents, background noise, network conditions all change perception. Test the voice you've picked on at least 5 real-world calls. Most platforms let you swap the voice in seconds. OnCall has a public demo where you can hear three voices answer your phone before signing up.
Mistake 5 — Setting it and forgetting it
Week-one performance is a starting point, not the final result. Plan a 30-minute review every Friday for the first month: scan call recordings, identify the 1-2 calls that didn't go well, fix the knowledge base or qualifying script, redeploy. Most platforms surface low-confidence calls automatically; OnCall does this in the Errors UI.
Realistic expectations for week one
AI receptionists are good in 2026. They're not perfect. Expect 80-90% of calls to go correctly out of the box, dropping to 95%+ after two weeks of iteration. The remaining 5% are usually:
- Callers asking unusual questions (about a specific niche service you didn't document).
- Callers with strong accents or background noise that confuse the speech recognition.
- Callers in unusual emotional states (very angry, very upset) where AI handling reads as cold.
For each of those, the fix is the same: capture the lead, route to a human, and update your AI's knowledge base or escalation rules so the next similar call goes better. Most platforms (OnCall included) make this a 60-second fix in the dashboard, not a developer task.
When to expand
After two weeks of stable inbound coverage, consider adding the next layer:
- Outbound campaigns. Use AI for cold outreach to old leads or unconverted form fills. Different compliance rules — your platform should respect TCPA windows and DNC scrubbing.
- Web chat widget. The same AI that answers your phone can answer chat on your website. See examples here.
- SMS follow-up sequences. 14-day automated cadences for unconverted leads.
- CRM sync. Push every lead and conversion into HubSpot, GoHighLevel, Pipedrive, Salesforce, or any custom webhook.
The compounding effect of AI answering 100% of inbound, qualifying every lead, and following up automatically over 14 days is what changes the math. Most operators don't realize the bottleneck wasn't leads — it was follow-up. AI fixes both.