AI agents built for production , not for the demo

What we mean by an AI agent

An AI agent is a program that receives a goal rather than a script. It decides at runtime which tool to call in which order and when to stop. A chatbot produces an answer in one step, a workflow runs a pre-written sequence, an agent plans for itself. That distinction is also the cost and the risk distinction.

Anthropic's own engineering guidance (Building Effective Agents, 2024-12-19) states that a pre-coded workflow beats an autonomous agent for most tasks. That is our default position too. In roughly half of our scoping calls we end up proposing a much cheaper workflow instead of an agent. The three layers are compared in detail in our chatbot, n8n and custom agent article, and the conceptual background to ReAct and tool use is in AI agents fundamentals.

Which agents we deliver

Four types reach production with us: document processing, customer service, internal knowledge base and process orchestration. Each has its own price band, its own lead time and its own risk profile. Three of the four start with a pilot, because that is where you find out whether the underlying data is good enough.

For the knowledge base agent the boundary matters: RAG handles search and answers, while the document agent extracts structured values. The two often live in one project but are measured differently. The engineering side of writing back into ERP, CRM or WMS is covered on our system integration page.

How a production agent is put together

Seven layers: orchestration, tool layer, integration, permissions, cost ceiling, human approval and fallback. The difference between a demo and a production system is precisely the last six. Very few vendors publish an actual architecture, so here is ours in full.

One engineering detail that saves real money: when the agent writes code that filters intermediate data in a sandbox instead of making separate tool calls, context collapses. In Anthropic's measured example (4 November 2025) a Drive to Salesforce copy task went from 150,000 tokens to 2,000. It does not apply to every task, but for steps that move large data sets we use it.

How you know the agent actually works

Through a golden dataset and an eval harness. The golden dataset is 100-300 hand-annotated real cases from your own material, each with the correct output. Every prompt, model and schema change runs against it, and in CI that run is a gate. A regression does not ship. Without it the system cannot be maintained, only hoped about.

• Retrieval quality, with numbers

  In Anthropic's measurement a plain embedding index gave a 5.7% retrieval failure rate, contextual embeddings 3.7%, adding lexical search 2.9%, and reranking 1.9% (Contextual Retrieval, 19 September 2024). That is why hybrid search and a reranker are standard equipment in our builds rather than an upsell.

• Language-specific measurement

  Global embedding leaderboards do not predict domain performance in a morphologically rich language. For Hungarian there is no maintained public retrieval benchmark as of August 2026, and the most recent independent test (Harang, 9 January 2025, 2,132 sentence pairs) found substantial overlap between the similarity distributions of related and unrelated pairs. German compound nouns create a related problem for lexical search. In both cases a project-specific eval set is mandatory, not optional.

• Deterministic validation after the model

  For documents: arithmetic reconciliation (net plus VAT equals gross, line items sum to the total), format checks on VAT numbers, IBANs and dates, and cross-checking on critical fields. The most dangerous vision-model failure is a value that is formally perfect and factually wrong, which no reviewer catches by eye.

• Observability, at published prices

  Langfuse self-hosted USD 0 licence (MIT), Langfuse Core USD 29 per month, LangSmith Plus USD 39 per seat per month, Braintrust Pro USD 249 per month (vendor price pages, 2026-08-14). Our default is self-hosted Langfuse inside the EU. The comparison is in our LangFuse and LangSmith article.

Most proposals skip this section, and not by accident. The eval harness is the least impressive part of the project and the most expensive one to retrofit: building it later costs the same as building it upfront, except the system is already making mistakes on live data while you do.

Where the data goes, and what leaves your estate

EU-region endpoints by default, with the data processing agreement part of the project. We never route client data through free API tiers. Where data cannot leave the building, the whole chain runs on premise: both the document parsing and the model stay on your hardware.

Two specifics that shape the vendor decision. On the Gemini API free tier Google states that content is used for product improvement, which rules it out for client data. Anthropic states that uploaded images are not used for model training, that uploads are ephemeral, and that a zero data retention option is available for PDF processing. For a fully closed setup and self-hosted open-weights models, see our article on local AI deployment and the GDPR-compliant chatbot with EU hosting page.

What it costs, and what you pay monthly

Delivery runs between EUR 2,750 and EUR 41,300 depending on type, with maintenance between EUR 83 and EUR 413 per month. Raw model spend is small next to that: 500 conversations a month costs between 470 and 9,900 HUF (roughly EUR 1.30 to EUR 27) depending on model choice. Most of the monthly fee is human work, not tokens.

Bands come from the AppForge published price list (updated 2026-08-14), converted at the Hungarian National Bank mid-rate of 1 EUR = 363.28 HUF on 2026-08-14. All figures net; Hungarian VAT of 27% may apply depending on customer country and B2B reverse-charge rules. We invoice EU clients in EUR. Fixed price after a 30-minute scoping call, quote within 24 hours.

On the rate side: Musketeers Tech (2 April 2026) puts Western European AI engineering at USD 100-180 per hour and Central and Eastern Europe at USD 50-100. We found no DACH-specific published rate survey for 2026, so treat the Western European figure as a proxy rather than a quote. We do not publish hourly rates ourselves, because every engagement is scoped as a fixed-range project. The full price list is on our pricing page, the payback maths with your own numbers is in the AI ROI calculator, and the nearshore case for a German or Austrian buyer is set out on the nearshore development page.

How a project runs, week by week

A RAG agent in the EUR 8,250-22,000 band typically takes 12 weeks, and the system has no write access until halfway through. That is deliberate. The most common failure mode is not the model: it is granting production permissions before anyone has measured how accurate the thing is.

1. Week 0: scoping call

   30 minutes, free, in CET. Output: which process, measured on which number, who owns it on your side. If the process itself is broken, we say so instead of selling AI on top of it.

2. Weeks 1-2: discovery and golden dataset

   We collect 100-300 real cases with their correct output. That becomes the benchmark for the rest of the project. Success gets a number here, not at the closing demo.

3. Weeks 3-4: tool layer and integration

   Read access to ERP, CRM or the document management system, the permission model, an MCP server or API adapter, idempotency and error handling. No write access yet.

4. Weeks 5-8: agent loop, guardrails, eval harness

   Planning separated from execution, cost ceiling, human approval points, tracing. The golden dataset moves into CI as a gate.

5. Weeks 9-10: shadow run

   The agent runs on live data but does not write. A daily variance report against the human decision shows where another approval point is needed.

6. Weeks 11-12: go-live on a narrow scope

   One supplier group, one customer type, one process. Expansion follows the measurements. Handover includes source code and runbook documentation.

In the chatbot band this compresses to 4-6 weeks; for a process orchestrator it stretches to 3-9 months. Working in CET means the review loop closes the same day for a Munich, Vienna or Zurich team rather than overnight. The full automation service description is on our process automation page.