Deployment
AI voice agents for SMEs: what they handle well, and what they do not
Where AI voice agents pay off for SMEs: the calls they handle well, realistic per-minute costs, and the guardrails to set before going live.
The short answer
AI voice agents suit high-volume, structured phone work: answering and qualifying inbound calls, booking and confirming appointments, chasing no-shows, taking simple orders and handling out-of-hours cover. They are a poor fit for complex complaints, sensitive conversations and anything needing negotiation. Costs are driven by minutes rather than seats, and the deciding factor for success is a clean escalation path to a human.
Key takeaways
- Judge voice agents on containment rate and booked outcomes, not on how human they sound.
- Inbound qualification and appointment logistics are the reliable early wins.
- Every deployment needs a confidence threshold and a live transfer route.
- Pricing is per minute plus telephony, so volume decides the business case.
- Disclose that the caller is speaking to an AI system, and keep records.
Voice is the AI capability that changed most in the last two years, and it is the one SMEs most often deploy badly. The technology is good enough for real work. The failures come from pointing it at the wrong calls.
Where voice agents work
| Use case | Direction | Why it works |
|---|---|---|
| Answering and qualifying new enquiries | Inbound | Structured questions, high volume, speed matters more than nuance |
| Booking, rescheduling and confirming appointments | Both | Clear success condition, calendar integration, repetitive |
| Out-of-hours and overflow cover | Inbound | The alternative is voicemail, so the bar is low and the gain is real |
| No-show and reminder calls | Outbound | Pure volume work with a measurable revenue effect |
| Order status and simple account questions | Inbound | Answer lives in a system the agent can query |
| Lapsed customer reactivation | Outbound | Low expectation, high volume, human takes over on interest |
Where they do not
- Complaints with an emotional charge. Escalate on the first sign of frustration.
- Anything involving negotiation on price or terms.
- Clinical, legal, financial advice or safeguarding conversations.
- Highly technical diagnosis where the caller cannot describe the problem in structured terms.
- Calls where the relationship is the product, such as key account management.
What it costs
Voice pricing is usage-based rather than per seat, which makes the business case unusually easy to model. The components are:
- Speech recognition, model reasoning and speech synthesis, priced per minute of conversation.
- Telephony minutes and number rental, which are cheap but not zero.
- Integration and build cost, driven by how many systems the agent must read from and write to.
- Ongoing tuning, transcript review and monitoring, which is where quality is maintained.
Per-minute rates have fallen steadily and continue to. Rather than quoting a figure that will age, model it this way: take your monthly call volume for the target call type, multiply by average handling time, and price the minutes at your supplier's current rate. Compare that against the loaded cost of the hours those calls consume today, plus the value of the calls currently going unanswered. In most SMEs the second number is larger than the first and nobody has ever measured it.
The metrics that matter
| Metric | Definition | What good looks like |
|---|---|---|
| Containment rate | Calls fully handled without a human | Rises steadily over the first two months as edge cases are fixed |
| Outcome rate | Calls that produced the intended result, such as a booking | The only metric that maps directly to revenue |
| Escalation quality | Whether transfers happened at the right moment | Reviewed by sampling transcripts weekly |
| Time to answer | Speed of pickup versus your human baseline | Usually the fastest visible win |
| Caller sentiment | Sampled from transcripts | Watch for a drop after any script change |
Note what is not on that list: how human the voice sounds. It matters far less than callers being told what the system can do and being moved to a person quickly when it cannot.
Guardrails before launch
- Disclose that the caller is speaking with an automated assistant at the start of the call.
- Set a confidence threshold that triggers transfer, and test it with deliberately awkward calls.
- Define recording, retention and consent handling in line with UK GDPR or the relevant US state rules, including two-party consent states.
- Keep a hard stop list: topics the agent must never attempt.
- Run a shadow period where the agent handles a slice of traffic and every transcript is reviewed.
- Give one named person ownership of weekly transcript review. Voice quality decays without it.
A sensible rollout
Start with out-of-hours or overflow, where the comparison is voicemail rather than your best receptionist. Run two to four weeks, review every transcript, fix the top five failure patterns, then extend to daytime overflow. Only after that should you consider a primary inbound line. Businesses that reverse this order tend to pull the whole thing after a bad week.
Frequently asked questions
- What can an AI voice agent do for a small business?
- Answer and qualify inbound enquiries, book and confirm appointments, cover out of hours and overflow, make reminder and no-show calls, and answer simple account or order status questions by querying your systems. It hands the call to a person when the conversation moves outside those boundaries.
- How much do AI voice agents cost?
- Pricing is per minute of conversation plus telephony, rather than per seat, with a one-off build cost for integrations. Model it by multiplying your monthly volume for the target call type by average handling time at your supplier's current per-minute rate, then compare with the loaded staff hours those calls consume today.
- Do customers mind speaking to an AI voice agent?
- Far less than expected when two conditions hold: they are told at the start, and they can reach a person quickly. Frustration comes from being trapped, not from the fact of automation. A fast, clean transfer path matters more than how natural the voice sounds.
- Do we have to tell callers they are speaking to an AI?
- Disclose it. It is the right default for trust, it is increasingly expected by regulators on both sides of the Atlantic, and it costs you nothing in performance. Handle call recording consent separately, including the US states that require all-party consent.
- What is a good containment rate for an AI voice agent?
- It depends entirely on call mix, so the useful measure is the trend rather than the number. Expect it to climb through the first two months as edge cases are fixed, and treat outcome rate, such as appointments actually booked, as the metric that decides whether it pays.
Work through this with us
brep runs AI operating reviews, workforce training and agent deployment for businesses of 10 to 250 people in the UK and US. One diagnosis, one workflow, someone accountable for keeping it working.