What does an AI chatbot or an AI agent cost in 2026?
A trained, managed AI chatbot costs between EUR 69 and EUR 333 a month with a EUR 69 to EUR 358 setup fee. As a custom build, a pilot chatbot runs EUR 1,100 to EUR 4,100, a RAG knowledge-base agent EUR 4,100 to EUR 22,000, and a multi-agent system wired into an ERP starts around EUR 13,800. The raw model spend behind all of it is a rounding error.
| Tier | Build | Monthly | When this is the right answer |
|---|---|---|---|
| DIY widget | EUR 0 | EUR 4 – 28 | Static FAQ with a few scripted flows and no integration (Kompaas, LeoAI) |
| Managed SaaS chatbot | EUR 69 – 358 | EUR 69 – 333 | Bot trained on your own content, limited integration (Apex Web, SmartAIChat, UseAIEasily) |
| Custom chatbot or pilot | EUR 1,100 – 4,100 | EUR 41 – 220 | One system integrated, 2 to 4 weeks (SocialPro, DField, AI-Smart) |
| RAG knowledge-base agent | EUR 4,100 – 22,000 | EUR 220 – 964 | Your own documents, source citations, permission handling, 6 to 12 weeks |
| Multi-agent or enterprise | EUR 13,800 – 220,000 | EUR 550 – 1,376 | Deep ERP, CRM and WMS integration, audit trail, SLA (AZAR, SocialPro) |
The bottom two rows are not chatbots. There the price depends on how many systems the agent has to reach into, and how much damage it does when it decides wrong. If what you actually need is help picking the layer, we wrote that comparison separately: AI chatbot vs n8n vs custom agent.
Why does almost no vendor publish a price?
Because the price is driven by integration work, and the vendor cannot see your integrations before scoping either. The effect lands on the buyer. With no reference point you can only accept the first number or reject it. You cannot negotiate a figure whose components nobody has shown you.
On 14 August 2026 we asked Perplexity what it costs to roll out an AI chatbot in this market. The answer was, word for word, that there is no public price list giving a concrete implementation price. That is exactly why this article is written the way it is: with numbers, sources and the arithmetic behind them.
Reviewing 17 vendors we found three patterns. The first is monthly SaaS, where a price exists but the message quota decides what you actually pay. The Apex Web market comparison of 12 July 2026 puts it plainly: the advertised price usually refers to the smallest quota. LeoAI opens at 9,990 Ft a month, about EUR 28, for 50 conversations.
The second pattern is banded project pricing. SocialPro, AZAR and DField all publish a range, and none of them explain what makes up the range. The third is a quote form with no numbers on the page at all. Where that happens, third-party estimates circulate instead of the vendor speaking for itself, which serves nobody.
What are you paying for inside a build quote?
An EUR 11,000 RAG project is roughly 200 engineer hours at the EUR 44 to EUR 66 rate typical of Hungarian suppliers. Backend and integration take 41 percent of that, model and prompt work 14 percent, testing another 14 percent. Discovery and deployment together stay below 12 percent. Most of the money goes into plumbing, not into the model.
The percentages are the midpoints of the phase-by-phase USD breakdown Musketeers Tech published on 2 April 2026, applied to an EUR 11,000 project:
| Phase | Share | Of EUR 11,000 | What happens here |
|---|---|---|---|
| Discovery, use case | 4.6% | ~EUR 510 | Process mapping, success criteria, and what not to build |
| Architecture, data prep | 8.3% | ~EUR 910 | Data sources, chunking, permission model, cleanup |
| Model choice, prompting | 13.8% | ~EUR 1,520 | Model bake-off on your data, prompts, tool definitions |
| Backend and integration | 41.3% | ~EUR 4,540 | ERP, CRM and document management wiring, error handling, idempotency |
| Interface | 11.5% | ~EUR 1,270 | Chat or admin UI, source display, feedback capture |
| Eval harness and QA | 13.8% | ~EUR 1,520 | Measurement set, regression runs, error taxonomy |
| Deployment, hardening | 6.9% | ~EUR 760 | Access control, rate limits, spend caps, prompt injection defence |
Two of those lines deserve their own paragraph, because in most quotes they are hidden behind one of the others.
Data preparation
RAG projects do not fail on the model. They fail on the input. If your documents are scanned PDFs, if the same partner name appears in four spellings, if last year’s price list was never archived, the agent will say confident nonsense. This is the line nobody enjoys raising during scoping, because it asks uncomfortable questions about the client’s data. In our projects it is the larger half of the architecture phase.
The eval harness
This is a set of 50 to 200 question and answer pairs that the system re-runs after every prompt change. Without it you cannot tell whether a fix has quietly broken three other things. In quotes it usually appears as “testing”, which implies someone manually tries ten questions. Not the same thing. When you change models, the eval set is the only thing that tells you whether the new one is genuinely better on your data.
If you want the broader picture on AI project pricing rather than agents specifically, we covered that in the AI development cost guide, and the entry level in the chatbot development cost guide.
What does it cost to run per month, token by token?
A RAG chatbot serving 500 conversations a month burns between EUR 1.30 and EUR 68 in raw model spend, depending on the model. An agent processing 200 documents a day through an eight-step loop costs EUR 884 a month on the same Claude Sonnet 5. The difference is not the price of the model. It is the number of steps.
Chatbot: 500 conversations a month
The assumptions are written out so you can redo this with your own numbers. Six turns per conversation. Per turn the input is 1,200 tokens of system prompt and tool definitions, 2,000 tokens of retrieved context across three to five passages, roughly 800 tokens of history, plus a 60-token user message. That is 4,060 input tokens and 250 output tokens per turn.
500 conversations x 6 turns = 3,000 model calls
3,000 x 4,060 input tokens = 12.0M input tokens (rounded)
3,000 x 250 output tokens = 0.75M output tokens
Claude Sonnet 5 ($2 / $10 per MTok):
12.0 x 2 = $24.00
0.75 x 10 = $7.50
monthly total = $31.50 = EUR 27.30
per conversation = EUR 0.055| Model | Monthly USD | Monthly EUR | EUR per conversation |
|---|---|---|---|
| Gemini 2.5 Flash-Lite | $1.50 | EUR 1.30 | EUR 0.003 |
| gpt-5.6-luna | $3.30 | EUR 2.90 | EUR 0.006 |
| Gemini 3.7 Flash (promo price) | $11.81 | EUR 10.20 | EUR 0.020 |
| Claude Haiku 4.5 | $15.75 | EUR 13.60 | EUR 0.027 |
| gpt-5.1 | $22.50 | EUR 19.50 | EUR 0.039 |
| Claude Sonnet 5 | $31.50 | EUR 27.30 | EUR 0.055 |
| Claude Opus 5 | $78.75 | EUR 68.20 | EUR 0.136 |
Behind a EUR 69 to EUR 138 monthly chatbot subscription, at this volume, sits an 80 to 95 percent gross margin. That is not an accusation. The fee pays for the rollout, the integration and the maintenance, not for tokens. It is simply worth knowing what you are buying. Anyone justifying the subscription with “the API is expensive” is not telling you the truth at 500 conversations.
Indexing matters even less. A 500-page knowledge base is around 300,000 tokens, which at the text-embedding-3-small price of $0.02 per million tokens is a one-off $0.006. Embedding the 3,000 monthly queries costs $0.0012. Retrieval infrastructure is never the bottleneck in a RAG budget.
Agent: 200 documents a day
Here the picture flips. A real agent does not call the model once. It works in a loop: read the document, extract the data, validate it, look up the ERP, correct, post, verify, report. That is eight calls per document, and the context accumulates as it goes.
200 documents/day x 22 working days = 4,400 documents/month
8 calls x 12,000 input tokens = 96,000 input tokens per document
8 calls x 500 output tokens = 4,000 output tokens per document
4,400 x 96,000 = 422.4M input tokens
4,400 x 4,000 = 17.6M output tokens
Claude Sonnet 5 ($2 / $10 per MTok):
422.4 x 2 = $844.80
17.6 x 10 = $176.00
monthly total = $1,020.80 = EUR 884
Same workload at a 70% prompt cache hit rate = $489 = EUR 423| Model | Without caching | At 70% cache hits | Saving |
|---|---|---|---|
| Claude Haiku 4.5 | $510 (EUR 442) | $214 (EUR 186) | 58% |
| Claude Sonnet 5 | $1,021 (EUR 884) | $489 (EUR 423) | 52% |
| Claude Opus 5 | $2,552 (EUR 2,209) | $1,206 (EUR 1,043) | 53% |
Now the number that sells this article. Process the same 4,400 documents with a single extraction call, 2,500 input and 400 output tokens each, and the monthly bill on Sonnet 5 is EUR 34. Through an eight-step agent loop it is EUR 884. Same task, same model, a 26x difference.
EUR 34
4,400 documents with one extraction call, Claude Sonnet 5
EUR 884
the same 4,400 documents through an 8-step agent loop
-98.7%
token reduction from code execution over MCP on a Drive-to-Salesforce task
Anthropic, 2025-11-04
So the price of an agent is not set by the model. It is set by how many turns it runs. Ask for that one number in every quote you receive. If the vendor cannot tell you, they have not designed the system yet.
Does a non-English agent cost more?
Probably yes. Tokenizers were trained overwhelmingly on English text, so heavily inflected and compound-heavy languages split into more pieces for the same content, and since billing is per token, the same conversation costs more. The direction is not in question. The magnitude is.
Here we have to be honest: we found no public, reproducible measurement that gives this ratio per model for German or for Hungarian. We know of one Hungarian measurement series, Peter Harang’s Medium post on tokenizer efficiency, and no vendor we checked publishes one of its own. Neither do we, so we are not putting a ratio here. What is usable instead: Anthropic states that the tokenizer introduced with Claude 4.7 produces roughly 30 percent more tokens for the same text at unchanged per-million prices. Same mechanism, arriving as a model upgrade.
Which four line items do most AI quotes leave out?
The vector database, observability, prompt maintenance and model churn. Together they run from about EUR 40 a month to several thousand, and in most quotes not one of them appears as its own line. Three of the four are human work rather than infrastructure, which is why they do not shrink as you scale.
1. Vector database
Pinecone Builder is $20 a month, Standard starts at $50 plus usage, Weaviate Cloud Flex starts at $45, and Qdrant publishes only a calculator. That is EUR 17 to EUR 61 a month, or EUR 200 to EUR 730 a year. For a mid-sized knowledge base of a few thousand passages the whole line is avoidable: pgvector on a Postgres instance you already run costs nothing extra in licence terms. If a quote contains a dedicated vector database, ask why your existing database is not enough.
2. Observability and evals
Langfuse Core is $29 a month, Pro is $199, LangSmith Plus is $39 per seat. The self-hosted Langfuse build is open source and free, so on a mid-market project the real cost of observability is zero licence plus running one container. Anyone billing you $199 a month for Langfuse Pro is selling convenience, not necessity. We compared the two tools here: LangFuse vs LangSmith.
3. Prompt maintenance
This is the biggest hidden line and still the one most often missing. The Azilen breakdown of 18 February 2026 puts prompt updates at $1,000 to $2,500 a month, more than the entire infrastructure combined. That is not surprising, because it is human work. When the model provider ships an update, when a new product type enters the catalogue, when it turns out the agent misreads one category of question, somebody sits down and fixes it. Monthly maintenance retainers in the EUR 80 to EUR 690 range cover exactly that, not the server.
4. Model churn
Model prices move in both directions and no quote we have seen prices that risk in. Three concrete examples from August 2026. Gemini 3.6 and 3.7 Flash are on promotional pricing: after 31 December 2026 input goes from $0.75 to $1.50 and output from $3.75 to $7.50, a straight doubling. Claude Sonnet 5, by contrast, made its introductory $2 and $10 rates permanent and dropped the planned increase. And the Claude 4.7 tokenizer counts roughly 30 percent more tokens for the same text.
When does per-run pricing on n8n, Make and Zapier start to hurt?
When the step count of the workflow and the monthly run count grow together. n8n bills per run regardless of steps, Zapier bills per step, Make bills per module operation. A six-step workflow therefore costs EUR 0.005 per run on n8n Pro, EUR 0.025 on Zapier Pro with annual billing, and EUR 0.037 on monthly billing.
| Runs per month | n8n Cloud | Zapier Pro | Make Core |
|---|---|---|---|
| 500 runs | EUR 20 | 2,500 tasks, $89 (EUR 77) | 3,000 credits, $9 (EUR 8) |
| 2,000 runs | EUR 20 | 10,000 tasks, $129 (EUR 112) | slider pricing, no public list |
| 10,000 runs | EUR 50 | 50,000 tasks, $289 (EUR 250) | slider pricing, no public list |
| 40,000 runs | EUR 667 | 200,000 tasks, $769 (EUR 666) | slider pricing, no public list |
Four breakpoints fall out of those numbers. The first is Zapier’s smallest plan: the 750-task Pro tier at $19.99 covers just 150 runs of a six-step Zap, which works out at EUR 0.115 per run. That is the “but it is only twenty dollars” trap.
The second: above roughly 5,000 tasks a month, per-task billing costs 3 to 7 times more for the same work. The more steps in the workflow, the worse the ratio. The third is the agent loop, where the step count varies per run and typically sits between 8 and 20. At that point the Zapier bill stops being predictable, which is worse than it simply being expensive.
The fourth breakpoint is self-hosting, and here we disagree with the usual advice. The n8n community edition licence is free, but somebody has to operate it. Managed self-hosting from a specialist prices from about EUR 110 a month on a VPS and EUR 220 on-premise, which is more than the EUR 50 n8n Cloud Pro plan. On cost alone, self-hosting only pays above roughly 40,000 runs a month. For data residency, NIS2 or EU AI Act reasons it is often still the right call even when it costs more.
One small trap to close on: the per-run unit price of the n8n Business plan (EUR 0.0167) is higher than Pro (EUR 0.005). Business is a feature and user tier, not a volume tier. Above 10,000 runs the logical step is self-hosting or Enterprise, not Business.
How does Central European pricing compare with DACH agency rates?
A Hungarian supplier bills EUR 44 to EUR 66 an hour. The DACH IT freelancer median was EUR 95, and German agency senior rates run EUR 120 to EUR 160. On identical scope that is a 40 to 60 percent difference, and the same RAG agent priced at EUR 4,100 to EUR 22,000 in Central Europe is benchmarked internationally at $10,000 to $100,000.
| Market | Hourly rate | Source |
|---|---|---|
| Hungary, agency | EUR 44 – 66 (16,000 – 24,000 Ft) | Apex Web / SocialPro, 2026-03-17 |
| Central and Eastern Europe | $50 – 100 | Musketeers Tech, 2026-04-02 |
| DACH IT freelancers, median | EUR 95 (software and web dev EUR 90, data and analytics EUR 100, SAP and ERP EUR 120) | Freelancer-Kompass 2026, 2026-07-02 |
| Germany, agency | EUR 100 – 180 (senior EUR 120 – 160) | software-entwickeln-lassen.com, 2026 |
| Western Europe | $100 – 180 | Musketeers Tech, 2026-04-02 |
| United States | $180 – 250 | Musketeers Tech, 2026-04-02 |
One caveat on that table. We found no reliable public hourly rate published specifically for AI development in Germany or Austria in 2026. The Freelancer-Kompass panel is the best primary data available and it merges Germany, Austria and Switzerland into one DACH figure. Treat the agency rows as vendor-published ranges, not as survey data.
The rate gap is also not the whole argument, and it is the weakest part of it. Two numbers from the German market matter more. Bitkom’s survey of 604 companies with 20 or more employees, published on 11 March 2026, found data protection requirements to be the single largest external barrier to AI adoption at 77 percent. And the ifo business survey of 5 June 2026 found that while 54.5 percent of German companies now use AI, only 18.7 percent of them build their own system. Four in five buy a boxed product, which means the custom question only comes up where the boxed product stops: your own data, your own process, your own ERP. Buying that build inside the EU means no third-country transfer to document, which is why we set out the delivery model on the nearshore development page rather than arguing about price.
What does AppForge charge for this?
We price in three bands, per project, not as a monthly subscription. The final figure comes after a free 30-minute scoping call, with a quote inside 24 hours. We do not publish an hourly rate because every engagement is scoped as a fixed-range project. Figures below are net; Hungarian VAT of 27 percent may apply depending on your country and B2B reverse-charge rules.
| Band | Price | Timeline | What it includes |
|---|---|---|---|
| Chatbot / process automation | 1,000,000 – 3,000,000 Ft (EUR 2,750 – 8,260) | 4 – 6 weeks | Chatbot on one knowledge base, n8n automation, basic integrations (Slack, email), monitoring dashboard, 1 month support, OpenAI or Anthropic API |
| RAG knowledge base system | 3,000,000 – 8,000,000 Ft (EUR 8,260 – 22,020) | 2 – 4 months | Company document ingestion, vector search (ChromaDB or Pinecone), user-level access controls, admin dashboard and analytics, cloud or local LLM, source citations |
| Custom AI solution | 5,000,000 – 15,000,000 Ft (EUR 13,760 – 41,290) | 3 – 9 months | Model fine-tuning, AI agents (LangGraph), local LLM deployment (Llama, Mistral), predictive analytics, computer vision, GDPR and data sovereignty |
| Process automation, per workflow | 100,000 – 10,000,000 Ft (EUR 275 – 27,530) | 1 – 2 week pilot | Simple 1 to 3 step workflow 100,000 – 500,000 Ft; multi-system integration with error handling 500,000 – 2,500,000 Ft; AI-driven complex automation 2,500,000 – 10,000,000 Ft; maintenance 30,000 – 150,000 Ft per month |
Our public AI reference is the AI-based search engine we built for the Hungarian Popular Culture Value Repository at kultura.hu. The service itself is described on the process automation page and the full price list sits on the pricing page. If you want to see the payback first, run your own figures through the AI ROI calculator.
Where do these numbers come from?
Every price comes from the provider’s own public page, retrieved on 14 August 2026. HUF and USD conversions use the Magyar Nemzeti Bank mid-rates of that day (1 EUR = 363.28 HUF, 1 USD = 314.36 HUF). The token calculations are our own modelling on official price lists, with the assumptions spelled out in the text so you can re-run them with your own volumes.
Central European vendor price lists:
- Apex Web market comparison (2026-07-12) and Apex Web pricing
- SmartAIChat, Kompaas, UseAIEasily, AI-Smart
- SocialPro cost breakdown (2026-03-17), AZAR, DField Solutions (2026-05-09), n8n.hu hosting fees
Official model and tooling prices:
- Anthropic Claude, OpenAI, Google Gemini
- Pinecone, Weaviate, Qdrant, Langfuse, LangSmith
- n8n, Make, Zapier
Benchmarks, market surveys and the exchange rate:
- Salt Technologies (2026-02-15), Azilen (2026-02-18), Destilabs (2026-02-23), Musketeers Tech (2026-04-02)
- Freelancer-Kompass 2026 hourly rates (2026-07-02), German agency rate guide (2026)
- Bitkom digitalisation survey (2026-03-11), ifo Institute AI use survey (2026-06-05)
- Anthropic on code execution with MCP (2025-11-04) and MNB exchange rates
Summary and common questions
How much does an AI chatbot cost in 2026?
A managed SaaS chatbot runs EUR 69 to EUR 333 per month with a EUR 69 to EUR 358 setup fee, based on published Central European vendor price lists from 14 August 2026. As a custom build, a pilot chatbot lands between EUR 1,100 and EUR 4,100, and a RAG knowledge-base agent between EUR 4,100 and EUR 22,000. AppForge prices chatbot and process automation work at EUR 2,750 to EUR 8,260 with delivery in 4 to 6 weeks.
What is the monthly running cost of an AI agent?
A RAG chatbot handling 500 conversations a month costs between EUR 1.30 and EUR 68 in raw model spend, depending on which model you pick. An agent processing 200 documents a day in an eight-step loop costs EUR 884 a month on Claude Sonnet 5, or EUR 423 with prompt caching enabled. On top of that come hosting, the vector database, observability and vendor maintenance.
Why does one vendor quote EUR 80 a month and another EUR 40,000 once for the same thing?
Because they are not selling the same thing. The subscription rents you a finished product with limited integration and a message quota. The custom build wires the agent into your own systems and the source code is yours. The dividing line is integration: if a widget on a website is enough, SaaS is cheaper. If the agent has to read from your ERP and write back into it, SaaS cannot help you.
Which cost line do AI quotes most often leave out?
Four of them: the vector database (zero with pgvector, EUR 17 to EUR 61 per month as a managed service), observability and an eval set (zero self-hosted Langfuse, USD 29 per month for Langfuse Core, USD 39 per seat for LangSmith Plus), prompt maintenance (human work, and the largest of the four), and the cost of switching models. Three of the four are labour, not infrastructure.
Is n8n worth it instead of Zapier for an AI agent?
For a six-step workflow, yes. n8n bills per workflow run, Zapier bills per step, so at 10,000 runs a month n8n Pro costs EUR 50 and Zapier Pro on annual billing costs EUR 250 for the same work. In an agent loop the gap widens further, because the number of steps per run varies and typically sits between 8 and 20. Zapier wins on connector coverage, not on price.
What does a RAG knowledge base cost to build?
In Central Europe, EUR 4,100 to EUR 22,000 with a 6 to 12 week delivery, based on published vendor ranges from 14 August 2026. The international benchmark for the same system is USD 10,000 to USD 100,000 according to the Salt Technologies AI development cost dataset of 15 February 2026. AppForge builds RAG knowledge-base systems for EUR 8,260 to EUR 22,020 in 2 to 4 months.
Does a non-English chatbot cost more to run?
Probably yes, but we cannot point you to a reproducible published measurement, and neither can anyone else we found. Tokenizers were trained overwhelmingly on English text, so inflected and compound-heavy languages split into more tokens for the same content. Before you sign, ask the vendor for the token count measured on your own documents with the model they propose.
How does Central European pricing compare with a German or Austrian agency?
A Hungarian supplier bills EUR 44 to EUR 66 an hour. The DACH IT freelancer median was EUR 95 an hour in the Freelancer-Kompass 2026 panel survey published on 2 July 2026, and German agency senior rates run EUR 120 to EUR 160. On the same scope that is a 40 to 60 percent difference, inside the EU, in the same time zone, with no third-country data transfer to document.
If you have a specific process you want an agent for, the fastest route is to tell us how many documents or conversations pass through it each month, and which system it has to write into. We turn that into a price range within 24 hours. If you are still choosing the layer, the chatbot, n8n and custom agent comparison helps, and for the technical background there is our RAG systems article.
