What does good AI customer service cost? An honest calculation

Everyone who sells AI customer service puts a price on their website. Almost nobody tells you what it actually costs. The license is only part of the bill. On top of it come setup, maintenance, and the hassle of a team learning to trust a new system. This piece is about the cost of AI customer service without the polished numbers: which pricing models exist, where the hidden costs hide, and how to work out for yourself when it pays for itself. Whether you are trying to automate customer service to save time or to handle more volume, the math is the same.

The two pricing models you will run into

Broadly, there are two ways vendors make you pay: per seat or per conversation. They sound interchangeable, but they push your behavior in completely different directions.

Per seat (per agent, per month)

You pay a fixed amount for every employee who works in the system. This comes straight out of classic helpdesk software, where people do the work and the software sits around it. Predictable on your invoice, that is the appeal. The downside sits exactly where AI is supposed to help. If the AI handles 60 percent of your questions, you need fewer people, but you still pay per chair. You are locked into a price built for a model you are moving away from.

Per conversation (or per resolved conversation)

You pay for each interaction the AI handles. That fits more logically with how AI customer service works: the system carries the volume, you pay for that volume. Watch the definition here. A conversation that gets handed to a human after two sentences, does that count? And a customer who comes back three times a week, is that three conversations or one customer? In busy months a per-conversation price can climb fast. Run the numbers on your peak, not your average.

The costs that are not on the quote

The subscription price is the easy part. The real bill sits in the things around it, and those only show up once you have already signed.

• Setup and training. Someone has to feed the AI your information: return policy, delivery times, common questions, the tone that fits your brand. For a simple webshop that is a couple of days of work. For an organization with lots of exceptions and jargon it can take weeks.
• Maintenance. Your catalog changes, your prices change, your policy changes. Someone has to keep the AI current, or after three months it gives outdated answers. That is not a one-off job but a recurring task.
• Integrations. An AI that cannot look inside your systems only answers general questions. If you want it to see a real order status or stock level, it has to connect to Shopify, your CRM, or your calendar. Some vendors charge for that separately.
• The wrong answers. Every AI gets it wrong sometimes. A bad answer costs you an annoyed customer, sometimes a return, sometimes a dent in your reputation. That is on no invoice, but it is a real cost to factor into your decision.
• The transition period. In the first few weeks your team runs double: the old work plus monitoring whether the AI does it well. That time is not free.

None of these items is a dealbreaker. But if you only compare the 49 or 99 euro a month and ignore the rest, you are comparing apples to half an apple.

How to calculate the payback time

The question is not whether AI customer service costs money. It does. The question is whether it brings in more than it costs, and how fast. You only need a handful of numbers for that, and you probably already have them.

1. Count your monthly volume. How many questions come in across chat, email, and phone combined? Look in your inbox or ticket system, not your gut.
2. Estimate the repeatable share. What percentage are variations on the same handful of questions: where is my package, what is the delivery time, how do I return something? For a lot of webshops that sits somewhere between 40 and 70 percent.
3. Work out the time per question. How many minutes does an average question take, including looking things up and typing? Often that is a few minutes, more for phone calls.
4. Multiply. Volume times repeatable share times minutes gives you the hours the AI can realistically take off your plate. Put your hourly wage next to that and you have the saving per month.
5. Subtract the real costs. Subscription plus the hidden items above. What is left, divided by the one-off setup cost, gives you the payback time in months.

An example to make it concrete. Say you get a thousand questions a month, half are repeatable, and each question takes four minutes on average. That is about 33 hours of work a month you can shift. What that is worth depends on your hourly wage. For most teams a subscription of around a hundred euro pays for itself well within the first month, provided the AI actually handles that repeatable share well. That last condition is the whole game.

Where the math goes sideways

The most common mistake is calculating as if the AI solves 90 percent. Almost nobody hits that, and whoever promises it is selling air. Count on clearing most of the repeatable share, and leave the tricky cases with your people. A good AI knows when it does not know, and hands off to a colleague with the context of the conversation attached, so the customer does not have to tell their story again. That is where the quality lives.

The second mistake is confusing saving with firing. In practice people rarely disappear. They do different work: fewer standard questions to tick off, more time for the customer who is really stuck or about to walk away. So do not count the gain only in saved wage costs, but also in what your team does with the freed-up time. That is harder to put a number on, and often more valuable.

What an honest choice looks like

Pick a pricing model that fits how your customer contact moves. Stable volume all year? Then predictability is nice. Big peaks? Then you want to know how the price behaves at that peak before you sign. Ask about the hidden costs before they appear, and test with your own real questions instead of the salesperson's demo. Tools like Conveya offer a short trial period for exactly that: you see on your own data what the AI actually catches, instead of betting on a brochure. And because Conveya pulls webchat, email, WhatsApp, Instagram, Messenger, and phone into one EU-hosted inbox that reads live from your systems, you can test the full picture, not a slice of it.

The cost of AI customer service is easy to get a handle on once you look past the monthly fee. Count the repeatable work, calculate with sober assumptions, and account for the items that are not on the quote. Then you know not only what it costs, but also whether it is worth it.