Voice agents are sold on a sound: a phone ringing out in an empty reception at half past six. Effective pitch, poor basis for a purchase. Whether one earns its keep in a UK service business is arithmetic — and that arithmetic has a line no demo shows you, because here the way a call is placed is regulated separately from what the call says.
TL;DR
Work out what is leaking through calls nobody answers, then subtract the whole cost of answering them: build, per-minute running cost, and the compliance work UK law attaches to automated calling. For inbound that last line is a policy. For outbound it is most of the project. Do the sum before you watch a demo, not after.
First, reject the receptionist comparison
The standard pitch puts a monthly licence next to a part-time salary and lets the gap make the argument. That comparison is wrong twice.
It is wrong about the person. Someone on your front desk is not a call-answering device that costs more: they chase the no-show, read the room when a patient is upset, and know which regular gets squeezed in. Swapping that for a voice agent is not a saving but a narrower service at a lower price — possibly the right decision, certainly a different one.
It is also wrong about the alternative, pricing the agent against the most expensive option available. What matters is the comparison with your phones today — voicemail, an answering service, a call-back at four o'clock — and with the cheap fixes that have no model in them: a booking link in your Google listing, an automatic text after a missed call, an out-of-hours greeting more useful than "we are closed".
So the question is not whether an agent is cheaper than a person. It is: what goes unanswered, what is that worth, and what is the cheapest thing that answers it.
The leakage side: three numbers you have to go and get
Revenue leaking each month = calls nobody answers × the share that would have booked × what one booking is worth.
None of the three comes out of an article. Any figure that is not yours describes somebody else's phones.
- Calls nobody answers. Your VoIP dashboard or carrier can produce the count. Take a full month rather than a bad Monday and break it down by hour: where they cluster decides whether you are buying out-of-hours cover or overflow cover at 11am, and those are different builds.
- The share that would have booked. This is the number the answer is most sensitive to, and the easiest one to flatter. Somebody ringing at nine in the evening may have rung two competitors in the same ten minutes, and answering second is not answering first. Pick a pessimistic figure and see whether the case survives it.
- What one booking is worth. Use the first appointment, not lifetime value. Lifetime value makes any business case work on paper, which is exactly why it is the wrong input for a sum you intend to check against a bank statement.
A worked example, to show the shape of the sum. The three inputs are round numbers picked so the arithmetic is legible — not benchmarks, and not drawn from anyone's data:
- 100 unanswered calls × 1 in 10 that would have booked × £100 a booking = £1,000 a month
If your own three numbers round to noise, stop: nothing below rescues a small leak.
The cost side, including the line the demo skips
The build is the line with a real number attached: our AI integration work starts at £4,500 excl. VAT, and a scope tells you what that buys before you commit.
The third line is the UK-specific one. AI calling here sits under PECR and UK GDPR at the ICO, and under Ofcom's rules about how a calling operation behaves. PECR was drafted long before any of this existed. It has one rule for automated calling systems, which send "sounds which are not live speech", and another for other marketing calls, and which one covers an agent holding a genuine two-way conversation has not been formally resolved. The ICO's informal advice, which it says is not legally binding, put even a short AI-made greeting in a marketing call under the automated-call rule. Our guide to whether AI receptionists are legal in the UK sets out the rules in full.
For an inbound agent, the resulting cost is real but bounded. If calls are recorded, callers must be told so, and why. You need a retention decision on those recordings rather than a default, and a handover to a person that genuinely works. That is a policy decision, a scoped piece of build and a line in your privacy notice: worth costing, not worth fearing.
Outbound changes the sum, not just a setting
Point the same product at a list and the economics change shape.
Outbound marketing calls made with automated calling systems need the subscriber's prior consent under PECR — consent that, in the ICO's words, "must specifically cover automated calls", evidenced rather than assumed, and not satisfied by the fact that the number belongs to a business. Ofcom's persistent-misuse policy then applies to the operation itself: calling parties should not make silent or abandoned calls, an abandoned-call message should say on whose behalf the call was made and give a basic-rate number to decline further calls, and Ofcom is likely to regard misuse at unsociable hours as more harmful. On top of that sits caller ID. Ofcom's General Condition C6 requires phone providers, where technically feasible, to take all reasonable steps to identify and block calls whose number is invalid, not dialable or does not uniquely identify the caller, and its caller ID guidance, applying from 29 January 2025, expects them to block calls from abroad showing a UK number outside limited legitimate cases — so non-compliant traffic stops connecting, which hurts commercially long before anyone writes to you. For breaches from 5 February 2026, the maximum PECR penalty for an undertaking is £17.5m or 4% of its total annual worldwide turnover in the preceding financial year, whichever is higher.
Cost those properly — consent capture, dialler configuration, the price of being wrong — and outbound is not the same purchase with a different toggle. That is precisely why inbound comes first: it is the case where the compliance line is small enough not to swing the answer.
Inbound — the compliance line is a policy
- The caller rang you, so nothing is being marketed at them and PECR's marketing-call consent does not arise
- Recording disclosure if calls are recorded, a retention period and a working human handover
- The value sits in calls that already exist and already go unanswered
- Nothing in the build is waiting on an unresolved legal question
Outbound — the compliance line is the project
- Prior consent per subscriber, specifically to automated calls, recorded before the first call is placed
- Ofcom's policy on silent and abandoned calls and the abandoned-call message
- A valid, dialable number shown, or providers block the call
- Two regulators can reach the same campaign for different reasons
Why there is no per-minute figure in this post
The running-cost line has no number against it, deliberately. A per-minute rate turns on your telephony provider, your call mix and how long a booking conversation takes against your particular calendar — and a figure invented for a web page would quietly become the figure you budgeted against. A blank sends you to your own call data; a plausible number stops you looking. Price it during scoping, next to the cost of the calls currently going unanswered.
Where the sum says no
Heads up
Low volume kills the case fastest: a handful of unanswered calls a month against a one-off build is a payback period measured in years. So does a misdiagnosed leak — if calls are answered but bookings still do not happen, the problem is the booking process and an agent will simply reach the same dead end faster.
A third case matters in regulated sectors. Where the honest answer cannot come from software — clinical questions, complaints, anything that would require naming a prescription-only medicine — the agent must decline and hand over. That is correct behaviour, but it narrows the calls it can finish, so run the sum on that narrower set rather than total call volume.
What to do this week
Pull a month of call data, count what rang out and when, and multiply it out. Then price the three cost lines against it — build, per-minute, compliance — and be honest about which side of the line the total lands on.
If it clears comfortably, start inbound: reception, booking and out-of-hours capture, with a real route to a person, and a booked call landing somewhere it cannot be lost — CRM automation is the receiving end. If the number is marginal, the cheaper fixes deserve their chance first — and if the website may be losing people before they ever reach the phone, that is the side a free audit looks at: search, speed and conversion-readiness on your own pages, with its scope set out there.
Either way, our guide to whether AI receptionists are legal in the UK has the full UK legal position, and getting in touch with what your phones actually look like is quicker than another demo.