Voice AI for service businesses: a buyer's guide
When voice AI pays off
The economic conditions:
- Meaningful inbound volume. Roughly 100+ relevant calls per month minimum.
- Predictable call patterns. Booking, status, FAQ, qualification — calls with shape, not free-form support.
- Cost of missed calls is high. Lost bookings convert poorly to callbacks.
- Recording is legally permissible. Required for evals; if you can't record, voice AI degrades silently.
If all four hold, voice AI almost always pays back. Three or fewer — the math is uncertain.
Use cases that work
| Use case | Why it works |
|---|---|
| After-hours appointment booking | Lost calls convert poorly to next-day callbacks. AI books immediately. |
| Inbound lead qualification | Quick disqualification of bad-fit leads saves AE time. CRM gets clean data. |
| Order / appointment status | High-volume, low-complexity, fits AI perfectly. |
| FAQ + routing | Caller asks a thing, AI answers from KB or routes to right team. |
| Outbound reminders / confirmations | Lower stakes than inbound, but still saves time. |
| Survey / NPS calls | Higher response rates than email; AI cheaper than human surveyors. |
Use cases to avoid (at least initially)
| Use case | Why it's hard |
|---|---|
| Complex medical triage | High empathy, high legal exposure. Use as router only. |
| Billing disputes / complaints | High emotion, high downside on getting it wrong. |
| Sales discovery for €50k+ deals | Discovery needs human judgement and rapport. |
| Crisis hotlines | Not appropriate, regardless of capability. |
| Bookings that require complex eligibility | E.g. medical appointments where eligibility depends on chart review. |
A useful test: would you trust a competent junior with a checklist to do this call? If yes, voice AI can. If no, voice AI shouldn't.
Cost economics
Operational cost per call:
| Component | Cost |
|---|---|
| Twilio inbound voice | ~€0.015/min |
| GPT-4o Realtime | ~€0.10/min (conversation) |
| Tool calls | €0.01-0.05/call total |
| Storage + observability | trivial |
A 90-second booking call: ~€0.12. A 5-minute support call: ~€0.40.
Plus the build cost (one-time):
| Scope | Investment |
|---|---|
| Single-intent booking line | €15-25k |
| Multi-intent qualification + CRM | €25-50k |
| Multi-language / multi-line | €50-100k |
Plus monthly retainer for ongoing evals + tuning (€1.5-3k/month typical).
What to evaluate when picking a vendor
The questions that matter:
1. "Do you record every call?"
If no — pass. You can't evaluate or tune what you can't replay. Voice agents without recording silently degrade over months.
2. "What's the warm transfer story?"
Show me a screenshot of what happens when the agent transfers to a human. The human should pick up with caller context attached — name, what they wanted, where the conversation went. If "warm transfer" is "the caller restarts from scratch," it's not warm.
3. "How do you handle PII in transcripts?"
For healthcare, finance, or any regulated industry. Look for: configurable PII redaction, your-cloud storage of recordings, retention policy controls, jurisdictional residency.
4. "What's the eval cadence?"
Vendors who say "we deploy and watch the dashboards" don't have a real eval discipline. Look for: eval set replayed against historic calls on every prompt change; regressions block deploy; numbers on a dashboard.
5. "Show me the call dashboard."
Not the slick sales screen — the one your ops team will actually use. Per-call volume, outcome distribution, cost per call, latency, drilldown to specific calls with transcript and audio. If they don't have it, they're not running production voice agents.
6. "What's the integration with our calendar / CRM?"
Direct native integration (Google Calendar, Outlook, HubSpot, Salesforce, Pipedrive) is ideal. If your stack is unusual, ask whether the vendor builds custom connectors and what it costs.
7. "What will you refuse to do?"
Any serious voice AI vendor refuses some things: no recording, no warm transfer to humans, deployments without legal sign-off in regulated industries, bypass of jurisdictional consent disclosure. If they'll do anything you ask, they're not opinionated enough.
The two-week pilot pattern
Before signing a multi-month engagement, we recommend a structured pilot:
Week 1: vendor builds a single-intent agent against your real call patterns (or simulated calls if real recordings aren't yet available). You test it personally.
Week 2: agent handles 5-10% of real after-hours traffic. You review every call. Measure outcomes against your baseline.
If the pilot doesn't show clear value in two weeks, the bigger build won't either. €3-6k pilot cost is the cheapest way to de-risk.
Build vs buy
| Build (custom) | Buy (SaaS vendor) | |
|---|---|---|
| Time to live | 3-6 weeks | days |
| Build cost | €15-50k | €0 |
| Per-call cost | €0.10-0.40 | €0.50-2 per call |
| 12-month TCO at 1000 calls/month | €27-60k | €6-24k |
| 12-month TCO at 10,000 calls/month | €40-100k | €60-240k |
| Customization | Full | Limited |
| Lock-in | None (code is yours) | Significant |
The crossover where custom build wins on cost: typically around 3,000-5,000 calls/month. Below that, off-the-shelf SaaS is cheaper at 12-month TCO. Above, custom wins.
What's coming in the next 12 months
Useful to know for buying decisions:
- Voice price drops continue. Realtime model costs have halved twice in two years.
- New voice providers from Anthropic and Google likely to ship realtime in 2026, breaking OpenAI's voice monopoly.
- Better non-English voices continuing to improve.
- MCP standardisation making voice agents more portable across vendors.
Bet on the patterns (recording, evals, escalation, observability) more than on the specific provider. Providers will shuffle; the patterns remain.
Where to go next
For the technical walkthrough of building voice agents see Building a phone agent with Twilio + GPT-4o. For pricing detail see How much does an AI agent cost. For our Voice & Phone Agents service page.
If you have a call surface to automate, drop us a note. One paragraph is enough.
Frequently asked questions
Keep reading
Building a phone agent with Twilio + GPT-4o: a complete walkthrough
Build a phone agent: Twilio provisions the number and streams audio, a Node.js bridge on Cloud Run pipes the audio to GPT-4o Realtime, function-calling tools execute real actions (book appointment, log lead, transfer). Recording, transcript, and observability on every call. Production deployment in 3-6 weeks.
How much does an AI agent cost? Real numbers from real builds
AI agent builds in 2026 typically cost €4-8k for discovery, €15-30k for a working prototype, €25-80k for production, €2-5k/month for retainer. Per-call infrastructure cost runs €0.01-€0.40 depending on shape. Honest numbers from real builds, with the trade-offs explained.
Hiring an AI development agency: 12 questions to ask
Twelve questions that separate serious AI dev shops from demoware vendors. Asks about evals, observability, code ownership, provider lock-in, references, and what they'll refuse to do. If a vendor can't answer cleanly, walk.
Voice & Phone AI Agents
AI receptionists, booking lines, and qualification calls — wired to your calendar, CRM, and ticketing.
Voice & Phone Agent
After-hours bookings, lead qualification, customer service overflow, FAQ lines
Want this delivered in your stack?
If the article describes a workflow you'd like to ship, drop us a note. We reply within one business day.