AI Voice Agent for Inbound Calls: An SMB Guide

Every unanswered call is a lead that just rang your competitor instead. An AI voice agent for inbound calls picks up, has the actual conversation, checks your calendar, books the appointment, and logs what happened, without a person standing by to catch it. This is what one looks like in production, what it costs, and what has to be true for it to be safe to run.

What an AI voice agent for inbound calls actually does

An AI voice agent answers the phone, understands the caller in real time, and takes action inside your systems, not just a scripted phone tree. That’s the line between it and an IVR: an IVR routes you through “press 1 for sales”, an AI voice agent has the conversation a good receptionist would have.

Under the hood it’s a pipeline: speech-to-text turns the caller’s voice into text, an LLM decides what to say and what to do, text-to-speech turns the response back into voice, and the whole thing runs over telephony infrastructure like Twilio. The part that actually makes it useful for a business isn’t the voice quality, it’s the tool calls in the middle: looking up the caller in your CRM, checking real calendar availability, writing a booking back, or transferring to a human with context attached. Without those, it’s an expensive way to leave a very convincing voicemail.

Why SMBs miss calls, and what it costs

The scale of the problem is bigger than most owners assume. Research from Aira on business phone data found that 62% of business calls go unanswered, a substantial share of potential revenue lost before a conversation even starts (Aira, 2026). And when a call does roll to voicemail, most callers don’t wait around: research cited by Capture Client puts voicemail abandonment at 67%, meaning two-thirds of callers who reach voicemail hang up rather than leave a message (Capture Client, citing BIA/Kelsey).

Put those together and the pattern is obvious: a missed call isn’t a minor inconvenience, it’s usually a lost enquiry, because the caller doesn’t leave a message, they call the next business on the list. For a service business where a booked job or a signed contract is worth real money, that’s the entire problem an AI voice agent exists to solve. It doesn’t need sleep, doesn’t put people on hold during a busy Monday, and doesn’t lose the thread when three calls come in at once.

How we build one: the platform decision

Two platforms dominate this space right now, Retell and Vapi. On the fundamentals they’re close: similar latency around 600ms, both run on Twilio, both can use ElevenLabs for voice. The difference is who owns the orchestration layer.

Vapi is BYOK: you wire together speech-to-text, the LLM, text-to-speech, and telephony yourself, and you manage retries and race conditions across those vendors on your own. Retell owns that orchestration layer instead of exposing it. For a client build that had to be reliable and hand off cleanly to a business that would run it themselves, Retell’s managed runtime plus custom functions got to production faster with less left to maintain afterwards.

The tools are where the real engineering sits. On a production inbound voice agent we built for a UK B2B client, there were four webhook tools: an availability check that reads real calendar slots, a caller and CRM lookup against HubSpot, a booking write-back, and a warm transfer to a human that carries the call’s context with it. Each one is a typed, tested function the model calls, not a hope that the model will describe the right thing convincingly. That’s the same discipline we cover in more depth in how we build production AI agents: the model does the conversation, plain code does anything that touches your real systems.

Turning every call into structured data

A voice agent that just talks is a demo. The value shows up when every call produces data your team can act on without listening back to a recording.

On that same build, Retell’s post-call analysis pulled nine fields from every call: call outcome, intent category, a qualified-lead flag, sentiment, callback-requested, contact details, service mentioned, competitor-mentioned, and a one-line summary. A mix of Boolean, enum, number and text fields, all firing straight into HubSpot so the team gets a triaged record on every call automatically.

The thing that wasn’t obvious going in: a meaningful share of callers weren’t ready to book, they were price-checking or comparing options against competitors. The intent enum surfaced that clearly, call by call, and it changed how the team scored lead quality in HubSpot. Instead of treating every call as a hot lead, they started prioritising the ones the data actually flagged as ready to move. That’s the difference between an AI voice agent and a call recording: one produces a transcript, the other produces a decision-ready record.

Can AI voice agents book appointments automatically?

Yes, and this is usually the single feature that pays for the build. The agent reads real slots off your calendar through a webhook tool, offers times to the caller, and writes the booking back the moment the caller confirms. No double-booking, no “someone will call you back to confirm”, the slot is taken while the caller is still on the line.

The important detail is sequencing. The booking has to actually succeed in your system before the agent tells the caller it’s booked. Anything less is a promise the business might not be able to keep.

The hardest part: making sure it never overpromises

Making the agent talk naturally is the easy part. The hard part is making sure it never confirms something the business hasn’t actually authorised. On a phone call, a hallucinated commitment can be legally binding, so this can’t be solved with prompt-tuning alone.

The way to build this properly: the agent can’t confirm anything it hasn’t actually executed. Booking confirmation only fires when the booking tool returns success. Prices and policy come from a grounded source, not the model’s memory. Anything the agent is about to act on gets read back to the caller first, so the caller hears exactly what’s about to happen before it happens. Anything outside the agent’s authorised scope escalates to a human instead of the model improvising an answer. On top of that, a post-call check flags any unauthorised commitment that might have slipped through, so the business catches it before the customer does.

That’s the actual engineering problem behind “AI receptionist for small business”: not making it sound convincing, but making sure it’s constrained to only do what the business has actually signed off on.

How much does an AI voice agent cost?

There’s no honest per-minute number to give you here, because the cost is driven by the tool integrations, not the voice model. An agent that just answers FAQs is a different build to one wired into your calendar, CRM, and a warm-transfer flow with context.

What we can tell you: it’s fixed scope and fixed price, agreed in a week-1 audit where we map how calls actually move through your business and write a one-page build spec. Most builds ship in 2 to 6 weeks, with working software shown every week so you’re not waiting until the end to see if it works. You own the build on handover, with full documentation, and an optional retainer covers ongoing iteration if you want it. No lock-in either way.

AI call answering service UK: does location matter?

Functionally, no. The stack, telephony over Twilio, a real-time voice model, webhook tools into your calendar and CRM, works the same wherever the business is based. What actually determines whether an AI call answering service is worth it isn’t the country on the invoice, it’s whether the build is wired into your real systems or just bolted on top of them as a glorified voicemail greeting.

If you’re a UK small business evaluating this, the questions worth asking a vendor are the ones this guide covers: what tools does the agent actually call, what happens when it doesn’t know an answer, and what data do you get back after the call ends.

Where to start

Missed calls are usually the easiest leak to find and one of the most expensive to ignore. If you want to know whether an AI voice agent is the right fix for your business specifically, that’s what the audit week is for: we look at how your calls actually come in, where they drop, and give you fixed scope and price before any code is written. See our AI agent development service for the full picture, or book a strategy call to talk through your specific setup. 30 minutes, no pitch.