The Cost of Automating Customer Service: Real Ranges for 2026

The cost of automating customer service depends on the level of autonomy you buy: an FAQ bot runs €1,500-3,000; an agent that answers from your company's real knowledge (catalog, policies, documentation) runs €5,000-12,000; and an agent that also executes actions — checking an order, processing a return, changing an appointment — runs €15,000-40,000 as a typical SMB project. Add a recurring cost of €50-400 per month for inference and infrastructure, plus annual maintenance of 10-20% of the project. The cost per resolved conversation lands in cents, versus the several euros each human-handled interaction costs.

Those are the headline numbers. The rest of this article takes them apart piece by piece: what each level actually buys, what pushes you to the top or bottom of the range, how to calculate the return against a human team, and which hidden costs to examine before signing anything. These are the ranges we use in our own proposals — not brochure figures.

The three automation levels and their prices

There is no such thing as "a chatbot": there are three levels of system with very different costs and results. Choosing the level is the first budget decision.

The column that matters is the last one. An FAQ bot is cheap but resolves little, and whatever it doesn't resolve still lands on your team — now with a customer irritated by the detour. The real value jump is level 3: when the agent can check the order status in your actual system and act on it, the conversation ends right there. "Where is my order?" — which in most ecommerce operations is on the order of a third of total volume — is only truly resolved at that level.

What pushes you to the top or bottom of the range?

Within each level, four factors move the price. None of them is the AI itself; all of them are integration and scope:

• Number of channels. Web chat is the base; adding WhatsApp, email and phone adds integration and testing per channel. Voice is the most expensive channel to do well.
• Number of systems to connect. Each system (ecommerce, ERP, CRM, logistics) is a connector with its own authentication, permissions and edge cases. Two clean integrations beat five half-done ones.
• The state of your documentation. If your return policy lives in the team's heads, the project includes writing it down. That's a cost — but one your company keeps forever.
• Languages and volume. Extra languages barely raise the project price (modern AI practically gives them away); high volume raises the recurring cost, not the project.

Cost per conversation: the metric that settles the decision

The honest comparison is per resolved conversation. A human-handled interaction costs, fully loaded, somewhere between €2 and €6 on an in-house team (more for specialized technical support). The same conversation resolved by an agent costs cents: inference for a typical conversation runs €0.05-0.30 depending on model and complexity, plus a proportional share of infrastructure.

With those numbers, the napkin math is direct. An SMB receiving 2,000 queries a month at an average cost of €4 spends €96,000 a year handling them. If a level-3 agent resolves 70%, those 1,400 conversations drop from €5,600 a month to under €500 including inference and prorated maintenance. The €15,000-40,000 project pays for itself within the first year — and that's before the second effect: service becomes instant and 24/7, which in ecommerce shows up in conversion and reviews.

Below 300-500 queries a month, the direct-savings case is weaker and the real motive is usually different: freeing the two people who currently drop whatever they're doing every time the chat pings. That's a valid case too — you calculate it the same way, using the value of the hours freed — but you should know which of the two you're buying. The general breakdown of what makes up any agent's price is in how much a custom AI agent costs.

The hidden costs proposals don't mention

Four line items show up after signing if nobody put them on the table first:

1. The content. The agent answers only as well as your documentation. If policies, product sheets and guides need creating or rewriting, that's days of work from someone who knows the business; put it in the budget rather than discovering it in week 3.
2. Exceptions and escalation. Designing what happens when the agent can't or shouldn't resolve — who it escalates to, with what context, on what schedule — is 20-30% of the project's work. It's also what separates a good experience from a customer trapped in a loop.
3. Real maintenance. Your products, policies and system APIs all change; the 10-20% per year is not a sales extra, it's what it costs for the agent to still be telling the truth six months from now.
4. Subscription platforms. The alternative to a custom agent is SaaS support platforms with AI that charge per resolution or per agent/month; at volume, the annual bill routinely exceeds the cost of a custom build. Always compare over 24 months, not 3.

What NOT to automate in customer service

The budget is also protected by knowing what to leave out. Serious complaints and angry high-value customers need a person with authority to decide: the agent detects the anger and escalates with priority, but it doesn't negotiate compensation. Same for legal cases, insurance claims, or any conversation where the economic or reputational outcome matters. The rule we apply in every deployment: automate the volume, never the judgment. And on every channel, the path to a human must exist and be visible — hiding it saves cents and costs customers.

How to budget it properly: the sequence

1. Measure what you have: queries per month by channel, the 10 most frequent contact reasons, current cost per interaction and average response time.
2. Classify the reasons into informational (level 2 resolves them) and operational (they need level 3). That ratio decides which level to buy.
3. Start with one channel and the top 10 reasons. Covering 70% of the volume well beats covering 100% badly.
4. Demand supervised mode at the start: for the first weeks, uncertain answers pass through a person. That's the phase that calibrates the system.
5. Compare against your baseline at 6-8 weeks: automatic resolution rate, response time, satisfaction and cost per conversation.

To see how customer service fits alongside the other automatable processes in your business, the map is in AI agent use cases by department. And if you'd rather have numbers for your specific case — your volume, channels and systems — before deciding level and budget, that's exactly what we do in our AI consulting engagements.

Frequently asked questions

What does automating customer service cost a typical SMB?

For an SMB with 500-3,000 monthly queries, the typical complete project — an agent with your own knowledge that executes the main actions, on web and WhatsApp — lands between €15,000 and €40,000, plus €150-400 per month. A more modest level-2 start begins around €5,000.

Is a SaaS platform or a custom agent better?

With low volume and standard processes, a subscription platform launches faster. With volume, or with actions against your systems (orders, returns, bookings), the custom agent wins at 24 months almost every time: no per-resolution fees, integrated with your real tools, and yours. The expensive mistake is choosing on the demo instead of the two-year total cost.

What share of queries can AI genuinely resolve?

A well-built level-3 system autonomously resolves 60-85% of volume in most B2C businesses, and the remainder reaches your team classified and with context. Distrust anyone promising higher figures without having seen your contact reasons.

How long until it's live?

A level 2 goes to production in 2-4 weeks; a level 3 with integrations, in 4-8 weeks. The first weeks run in supervised mode while the accuracy rate is calibrated, and the return is evaluated against the baseline at 6-8 weeks.

Does the agent replace my support team?

It replaces the repetitive volume, not the judgment: the team goes from answering the same thing a hundred times to handling complex cases, important customers and the improvement of the system itself. In practice, most SMBs don't cut headcount — they absorb growth without hiring.