AI phone assistant costs: building blocks, calculation model and honest limits

At the end of a working day, three slips of paper lie on the desk: a callback request, a name without a number, a note nobody can decipher any more. How many calls never even ended up as a slip of paper? In most businesses the clear answer is: nobody knows. Missed calls leave no trace, apart from a line in the call list that rarely anyone analyses.

This is exactly where the cost question starts, and it is usually asked the wrong way round. People ask what an AI phone assistant costs per month. The better question is what not answering costs, and how much of that an assistant can win back. Without that figure, every price remains a number without a yardstick, whether it seems high or low.

This scenario follows a business whose phone often rings during the day while everyone's hands are full. It is not project documentation but a technical analysis: we explain what the costs are made up of, how you can estimate the benefit with your own figures, and how you can tell that the purchase is not worthwhile in your case.

An everyday example makes the blind spot tangible. Whoever is on a building site, in a treatment room or in a client meeting sees the phone ringing and decides within seconds to let it ring. That decision is right, because the work on site comes first. But it has a price that is booked nowhere, and that price is the basis of any profitability calculation.

If you collect three offers for call handling, you get three billing logics. An answering service often bills by answered calls or minutes, a secretarial service by the hour or a flat rate, an AI phone assistant by setup plus ongoing usage. Comparing monthly amounts says little as long as it is unclear which service is included and which is not.

On top of that, with AI solutions several providers are chained together. The phone number comes from a telephony provider, speech recognition and speech output from voice services, the language model from another service, and the integration and support from the agency. Each layer has its own pricing model, and some bill per minute, others per request or per character.

Another stumbling block is minimum terms and allowances. A low monthly price with included minutes seems predictable until the allowance limit is exceeded and a different tariff applies. So always ask for the price at twice your expected volume, and what happens to the phone number and call flow on termination. A phone number you cannot take with you is a hidden lock-in.

The second difficulty lies on the benefit side. Missed calls are not lost orders: some callers call again, some only had a follow-up question, some were advertising. If you equate the value of calls with the average order without checking, you make any solution look good. If you set it too low, you overlook a real lever.

For the calculation, how high an item is matters less than how it behaves when the number of calls rises. Setup is incurred once and is independent of volume. It covers collecting the typical requests, conversation flow, integration with a calendar or enquiry management, test calls and introduction to the team.

Telephony and voice services are usage-based: more call minutes, more consumption. Here it pays to look at the average call length. An assistant that clarifies what is needed in short, focused dialogues consumes less than one that rambles. Conversation design is therefore also a cost question and not just a question of style.

With telephony, a second look at forwarding is worthwhile. If a call is handed from the assistant to a person, two connections may run simultaneously, and both may be billed. Whether and how this happens depends on the provider and on the integration with your phone system. Ask about it before you estimate the volume, because with frequent handovers this item becomes significant.

Support is the item that is missing most often. A phone assistant is not a device you set up and forget: opening hours change, services are added, callers ask questions nobody anticipated. Without regularly reading through the conversations, quality deteriorates unnoticed. Anyone who does not plan for support only appears to save money.

A fifth type is often forgotten: internal effort. Someone in the business has to handle callbacks, read reports and give feedback. This effort is small but not zero, and it is missing from many offers. We therefore include it explicitly, even though it generates no invoice.

The usual way is to start from the offer and ask whether the price seems acceptable. We would proceed the other way round: first estimate the amount lost through missed calls, then derive the upper limit for running costs, and only then collect offers. That way the yardstick is fixed before a salesperson sets it.

The model needs four figures, all of which can be measured in your business. The first is the number of calls per month. The second is the share that is not answered. The third is the share of those calls that were a genuine enquiry. The fourth is the value of an enquiry, that is, the contribution a handled enquiry makes on average, not the revenue of an order.

Voicemail, answering services and secretarial support are not bad solutions here, but solutions with a different profile. Voicemail costs almost nothing but assumes callers leave a message, and many do not. A secretarial service offers judgement but is tied to attendance. An answering service bridges peaks but knows your business only superficially. The assistant does not replace these routes across the board; it shifts the threshold above which they pay off.

The trade-off against alternatives follows from these figures. If the share of missed calls is low, a better voicemail with a callback promise is often enough. If it is high and the calls arrive at fixed times, a secretarial or answering service should be considered. The assistant plays to its strengths when calls arrive irregularly, simple requests predominate and round-the-clock availability matters.

Any figures we named here would be invented, because they depend on your business. The model therefore works with placeholders that you fill in yourself. A stands for calls per month. V is the share of calls that are not answered. Q is the share of these missed calls that were a genuine enquiry. W is the value of an enquiry, based on the contribution it makes on average.

The value lost per month is A × V × Q × W. An assistant wins back only part of that, because not every caller speaks to an AI and not every enquiry becomes an order. For this we introduce R, the share of missed enquiries that the assistant actually captures and follows up. The recovered value is A × V × Q × R × W.

Set against this are the costs: running costs K per month, consisting of support, phone number and usage-based items, and the setup E, which you spread over a chosen period of M months. Monthly balance = A × V × Q × R × W − K − E/M. If it is permanently below zero, the purchase does not pay off, regardless of how good the technology is.

A worked example only shows the mechanics: enter placeholder values of your own choosing for A, V, Q, R, W, K, E and M and change one at a time. That shows you which figure moves the balance most. As a rule these are V and W, not K. If the value per enquiry is high, you can bear higher costs; if you handle many simple enquiries of low value, you need to calculate more precisely.

The first source is the call list of your phone system or provider. It shows how many calls came in, which were not answered and at what times. A period of four to eight weeks is a sensible start, because individual weeks are distorted by holidays or seasons. Also check whether calls went to voicemail and whether messages were left there.

Q can only be estimated with samples. Call back a selection of the missed numbers, or read the voicemail messages, and sort them into genuine enquiry, follow-up on ongoing matters, advertising and other. This is tedious, but it prevents you from building on an assumption that turns out to be wrong. An hour of effort here often saves a wrong purchase.

You determine W from your business accounts: what is left of a typical enquiry after deducting costs, multiplied by the probability that the enquiry becomes an order? Here honesty helps more than optimism. Better to set a cautious value and check whether the calculation still holds. R is the most difficult figure and can only be read off in operation, which is why you start with a pilot phase.

The most common failure pattern is double counting. A caller who rings three times because nobody answers appears three times in the list but is one enquiry. If all three are counted as lost enquiries, V rises artificially and the calculation comes out too favourable. The same applies to callers who get through later after all or write via a form.

A second failure pattern is usage beyond what was planned. With a pricing model by the minute, a single caller who talks for a long time or gets lost can reach the consumption of a whole day. Sensible measures are upper limits per call, a friendly ending after a defined duration and a warning threshold for monthly volume. That protects the bill as much as the quality.

Third, the calls that are not meant for the assistant at all: suppliers, existing customers with a direct extension, public authorities. If they block or frustrate it, damage arises that appears on no cost line. A clear forwarding logic for known numbers and a quick route to a person help here, as described in the sister study on handover.

A failure pattern on the benefit side is the callback that never happens. The assistant captures an enquiry cleanly, but it sits in a mailbox for two days. Then the advantage over voicemail is small, and R effectively drops to zero. The economic value only arises when someone reliably calls back or the assistant books an appointment itself, as described in the study on appointment booking.

Fourth, the interaction with voice quality. An assistant that stutters or answers with pauses leads to more dropped calls, and R falls. How latency and speech recognition relate to this is covered in the study on voice and latency. For profitability the rule is: a weak voice is not a question of style but directly reduces the part the assistant wins back.

An assistant can make sure that calls no longer go nowhere and that requests reach you in a structured way. How much that is worth depends on the figures of the calculation model, and we name no numbers for it. Promises about revenue growth, saved positions or rates would be invented and would feign a certainty that nobody can have.

The benefit can be measured with figures close to the cause. These include the share of calls that are answered, the number of structured enquiries captured, the time to callback and the share of conversations that had to be handed to a person. From the callbacks you can later read how many enquiries led to appointments or orders, insofar as you record this in the business.

We would recommend a pilot phase that runs long enough to see normal and quiet weeks. During this time, conversations are read through on a sample basis, mishandled cases are collected and the model's placeholders are replaced with measured values. Only then can you judge whether regular operation makes sense and which adjustments to the conversation flow are worthwhile.

It is part of the picture that the effect does not appear immediately. In the first weeks more rework is often needed, because callers say things nobody anticipated. Plan effort for this period and do not judge the assistant by the first week but by the trend.

The most important insight is unspectacular: the price of the assistant is rarely the lever. Anyone who knows the value of an enquiry and the share of genuine enquiries can quickly classify offers. Anyone who does not is comparing apples with voicemails. That is why every consultation with us starts with the call list and not with a product presentation.

The second insight concerns support. A cheap entry without planned maintenance becomes more expensive later, because errors accumulate and callers drop off. Include support as a fixed item and agree what it covers: reading through conversations, adjusting answers, reacting to changes in opening hours or services.

Third: limit the scope at the start. An assistant that takes requests, records callback wishes and answers simple questions is easier to run and evaluate than one that quotes prices, negotiates appointments and resolves complaints. The range of functions can grow once the figures from the pilot hold up.

Fourth, honesty in the calculation includes disclosure to callers. It should be clear at the start that the conversation is being held with an AI and may be recorded or transcribed. Which obligations follow from this in an individual case, clarify with your data protection adviser. We mention it here as orientation and not as legal advice, because it affects the cost of notices, contracts and deletion concepts.

Fifth and last: there are businesses where the assistant is not worth it, and that is a legitimate answer. The conclusion is about that.

You should do without an AI phone assistant if only a few calls are missed and a good voicemail with a reliable callback is enough. The same applies if your calls consist mainly of complex advisory conversations that a person has to conduct anyway, or if the value of a single enquiry is so low that the running costs cannot support it.

Organisational reasons sometimes speak against it too: nobody in the business takes care of the callbacks, the requests are not described, or callers expressly expect a personal voice. Then the assistant shifts the problem instead of solving it. A secretarial or answering service, call forwarding or simply clear times for callbacks can then be the better answer.

One borderline case deserves a mention: the business with strongly fluctuating load. If you get very many calls in a few weeks of the year and hardly any otherwise, variable costs can work well, because little is incurred in quiet months. At the same time, maintaining the conversation flow is a fixed item. So calculate with the year, not the average month, and check whether a time-limited deployment in the peak period is enough.

The next step is a small one: export the call list of the last four to eight weeks and count how many calls were not answered and at what times. That gives you A and V, and with a sample also Q. Send us these key figures, without names or numbers, and we will work through the model with you, with an honest assessment, even if it comes out against a purchase.