How much an AI agent for customer service costs in 2026

The question “how much does an AI agent for customer service cost” has three legitimate answers, and that is exactly the problem. Answer 1: €200 per month if you take a preconfigured SaaS chatbot. Answer 2: €15,000-80,000 upfront if you want something custom integrated into your systems. Answer 3: €150,000+ if you are a large company with enterprise requirements. Which answer is the right one for you depends on factors that probably nobody has ever explained to you clearly. Let’s map the territory, with real numbers and no made-up flat rates.

Answer in 3 lines

• Preconfigured SaaS tier: €100-500 per month, setup in 1-2 weeks, limited brand and logic.
• Custom integrated tier (the one we serve): €15,000-80,000 upfront + €500-2,000 per month, setup in 4-12 weeks, real integration into your systems.
• Enterprise tier: €150,000+ upfront for volumes above 50,000 requests per month and complex compliance.

The three levels of AI agent for customer service

These are not three different products: they are three different philosophies of how to think about the problem.

1. Preconfigured SaaS chatbot (€100-500 per month)

Products like Intercom Fin, Zendesk AI, Tidio, Crisp, Freshdesk AI. You buy a subscription, you upload your knowledge base (FAQs, help center articles, possibly company documents), you configure some basic rules, it goes into production in 1-2 weeks.

Real pros: cost predictability (you pay monthly per user or per conversation), setup speed, zero maintenance because the vendor updates the AI model behind the scenes, ready standard integrations (Shopify, HubSpot, Salesforce, some business management systems).

Real cons: the brand of the conversation remains partly the vendor’s (typical phrases, standard fallback patterns, predetermined response formats). Integrations with specific Italian systems (TeamSystem, Zucchetti, ESA, vertical business management systems) are often absent or require costly custom APIs. Complex business logic is not easily expressible.

When it makes sense: e-commerce with standardizable FAQs, B2B SaaS with repeated onboarding, businesses that have a well-structured help center and just want to answer the same 50 questions better.

2. Custom integrated AI agent (€15,000-80,000 + €500-2,000 per month)

This is the tier we serve. You build an agent specific to your context, integrated directly into your systems (CRM, business management system, e-commerce, ticketing, warehouse). The agent doesn’t just answer FAQs: it reads real orders, checks real shipment statuses, accesses actual client data, can write to systems (create tickets, open requests, update records).

Real pros: 100% brand-aligned, integrated with the systems you actually use, evolvable over time by adding capability without replacing the foundation, total control over data (no vendor lock-in on the LLM model, which can be swapped between Claude, GPT, self-hosted open-source models).

Real cons: higher upfront investment, 4-12 week setup time, requires a serious partner to build it (a custom agent is software engineering, not configuration).

When it makes sense: SMBs with proprietary non-standardizable processes, businesses with volumes exceeding 500 requests per month (above which the per-message pricing of SaaS becomes expensive), companies that want to maintain control over the conversation flow and the data that passes through.

3. Enterprise solution (€150,000+ upfront)

For volumes above 50,000 requests per month, complex multilingual requirements, heavy compliance (PCI-DSS, sector-specific healthcare or banking), integration with enterprise stacks. Typically involves vendors like Accenture, Capgemini, or internal teams at large companies that build in-house.

When it makes sense: large companies, banks, healthcare, sectors where regulatory compliance weighs as much as technology. For Italian SMBs it is rarely the right answer: you pay a compliance and governance overhead designed for organizations of a different scale.

Which factors determine price in the custom tier?

Six factors in decreasing order of impact.

1. Number of systems to integrate. An agent that reads only from the CRM costs less than an agent that reads from CRM + business management system + e-commerce + WMS. Each integration typically adds €2,000-8,000 to the upfront cost, depending on the API maturity of the target system. Italian systems without modern APIs (e.g., old versions of TeamSystem, on-premise vertical business management systems) require more expensive custom adapters.

2. Expected monthly request volume. Above 5,000 requests per month you need a serious caching and rate-limiting architecture, and LLM costs become significant. The €500-2,000 monthly range is optimized for volumes under 3,000. Beyond that, you move to a different design with proportional costs.

3. Channels it must be available on. Web chat only: minimum. WhatsApp Business: adds €1,500-3,000 setup + cost per Meta conversation. Email parsing: €2,000-5,000 for classifier setup. Voice (phone call with voice AI): €8,000-20,000 setup, significant cost per minute.

4. Supported languages. Italian only is included. A second language (typically English) adds €2,000-5,000 of prompt engineering and testing work. Three or more languages: it depends on complexity, but we are above €10,000 additional.

5. Level of tone of voice customization. A “neutral professional” agent is the default. An agent that replicates a specific brand voice (formal, technical, ironic, regional Italian) requires 2-4 weeks of iterative prompt engineering work: typically €3,000-6,000 additional.

6. Required compliance. Basic GDPR is included. Regulated sectors (healthcare with patient data, finance with transaction data, public administration) add work on audit logs, data retention, anonymization: €5,000-15,000 depending on scope.

Three concrete scenarios with real numbers

Scenario A: SMB e-commerce, 1,000 requests per month

Online clothing store, ~30k visits per month, ~1,000 customer service conversations per month. Typical requests: where is my order, can I change the size, how does the return work, is size X of this product available.

Typical setup: agent integrated with Shopify (catalog + orders), Brevo (transactional emails for follow-up), existing help center as FAQ source. Italian only. Web + WhatsApp Business channel.

Costs: €15,000-25,000 upfront (3-5 weeks of work), €500-700 per month of operations (LLM tokens ~€150-300, monitoring + maintenance ~€300-400). Typical payback: 4-7 months thanks to the reduction of human work on customer service.

Scenario B: Professional firm, 200 active clients

Accounting firm or consultancy with ~200 business clients. Typical questions: tax deadlines, status of case X, missing documents, appointment confirmations. Requires integration with firm business management system (TeamSystem or equivalent) and calendar.

Typical setup: agent integrated with the firm’s business management system, shared calendar, internal regulatory knowledge base. Strong client disambiguation logic (it is essential that the agent doesn’t mistake which client it is responding to). Email + website web chat channel.

Costs: €20,000-35,000 upfront (4-7 weeks of work, much of which on the connection to the business management system that is often not well documented), €700-1,000 per month. Typical payback: 6-10 months via reduction of junior collaborators’ time on repetitive requests.

Scenario C: B2B distribution, 5,000 requests per month

Industrial or materials distributor, ~800 active business clients, ~5,000 conversations per month between orders, invoicing, logistics. Multiple channels (WhatsApp Business is the first, email the second).

Typical setup: agent integrated with the ERP business management system (order statuses, availability, invoicing), the WMS (shipments), CRM for historical interactions. Complex logic: the agent does things (opens tickets, updates records, registers reminders), it doesn’t just read. B2B GDPR compliance.

Costs: €40,000-70,000 upfront (10-14 weeks of work), €1,500-2,500 per month. Typical payback: 8-14 months. Above this scale the value is not only in savings on personnel but in faster response (orders are processed in minutes, not hours), which has a direct impact on conversion rate and churn.

What is included and what is not

A serious estimate for the custom tier includes:

• Discovery and mapping of real use cases (1-2 weeks)
• Integration with the planned systems (2-6 weeks depending on number and API quality)
• AI agent development with iterative prompt engineering (2-4 weeks)
• Testing on real cases, training of the internal team (1-2 weeks)
• Production deployment with basic monitoring
• First 1-2 months of hyperattention post-go-live (bug fixing, prompt fine-tuning)
• Technical and operational documentation

What is NOT included (and should be estimated separately):

• Cost of LLM tokens to the provider (Anthropic, OpenAI, etc.): typically €100-500 per month
• Costs of external channels (WhatsApp Business API to Meta, voice gateway, etc.)
• Deep re-trainings if the underlying product changes drastically
• Expansion to new channels or new integrations after go-live (these are additional projects)
• Any third-party monitoring software licenses (LangSmith, Helicone, etc.)

Red flags of bad estimates

Four phrases that should set off an alarm.

“All included starting from €5,000”. A serious custom AI agent does not cost €5,000. Whoever sells at that price is selling either a rule-based chatbot disguised as an AI agent, or a setup that breaks at the first case outside the loaded FAQs.

“We do it in 2 weeks”. Two weeks are enough for a preconfigured SaaS setup. A custom integrated agent requires a minimum of 4-6 weeks even in the simplest cases, because the time is taken by integrations with your systems, not the AI part.

“It works out-of-the-box without training”. No custom AI agent works out-of-the-box. The value lies precisely in training on your data, on your processes, on your exceptions. Whoever promises the opposite is selling a generic product passed off as custom.

“Guaranteed ROI in 30 days”. Guaranteed ROI does not exist, on any IT project. Realistic ROI for a custom agent is 4-10 months depending on the scenario. Whoever guarantees 30 days is either lying or selling something that doesn’t match what you think you are buying.

Bonus red flag: the detailed one-page estimate with line items that don’t correspond to real technical objects (“general AI setup”, “complete integration”, “AI training”). A serious estimate specifies which systems are integrated, with which APIs, in what timeframes, with what delivery milestones.

How a serious estimate is made (in 4 steps)

1. Discovery call (30-45 minutes, free). What your customer service does today, on which channels, what are the 5-10 most frequent cases, which systems are involved, what is the real monthly volume. Without this information any estimate is hypothetical.

2. Technical assessment (1-2 weeks, free if we proceed, typically €1,500-3,000 standalone). The systems are actually examined: which APIs they expose, what data is there, what regulatory constraints exist. Output: assessment document with clear scope.

3. Technical-economic proposal (3-5 pages). Precise line items: integration with system X, agent development with use case Y, delivery milestones. Costs broken down by phase, not in an opaque flat rate.

4. Optional POC (€3,000-8,000 in 2-3 weeks). If the total investment is significant (over €30k), a POC on a narrow use case allows you to validate technical fit before committing to the full project. Often the POC cost is discounted from the project if we proceed.

FAQ

How much does a basic AI agent cost?

SaaS tier: €100-500 per month (Intercom Fin, Zendesk AI, Tidio). Simple custom tier: €15,000-25,000 upfront + €500-1,000 per month of operations. The choice depends on how much you want to customize and integrate into your systems. For an e-commerce with standard FAQs, SaaS is often enough. For a business with proprietary processes and real integrations, custom becomes the only sensible choice.

How long does it take to put it into production?

SaaS: 1-2 weeks of configuration, excluding the time to write the knowledge base well (which is often the real hidden cost). Custom: 4-12 weeks (4-6 for simple implementations like scenario A, 8-12 for complex integrations like scenario C). Enterprise: 3-6 months typical. The timelines include setup, integrations, prompt engineering, and testing, but not the training of the internal team (1-2 additional weeks).

Is there a monthly cost after implementation?

Yes, always. For custom agents: €500-2,000 per month. This includes LLM tokens (€100-500), monitoring and observability (€100-300), evolutionary maintenance (€300-1,200). Costs vary with volumes. Whoever proposes you a custom agent “without monthly costs” is selling a system that will not have maintenance, and in 6 months will break at the first change of the upstream API.

Can I try a POC first?

Yes, and it is almost always the right thing to do if the total investment exceeds €30k. A typical POC costs €3,000-8,000 and lasts 2-3 weeks. It serves to validate the technical fit (can it really read our business management system?) and the business fit (do users really accept the AI interaction?) before committing to the complete project. Often the POC is discounted from the project cost if we proceed.

Can I change the LLM model afterwards (e.g., from OpenAI to Claude)?

In a well-built custom agent, yes. The prompt engineering part may require adjustments, but the architecture does not change. In SaaS agents no: the vendor decides which model to use and rarely exposes it as a choice. For those who want to avoid lock-in on a single LLM provider, the custom tier is the only sensible option.

Want a serious estimate in your context?

If you have made it this far you are probably evaluating an AI agent for your customer service and you want real numbers on your case. No automatic flat rates, no sales bots. A real conversation with whoever would build the agent. Let’s talk.

To better contextualize the investment, also useful: the pillar page AI agents, the page dedicated to the customer service AI agent, and the one on the WhatsApp AI agent for e-commerce.