LLM Integration · Agentic AI
Function calling and tool use with Claude and GPT. Your AI agent books appointments, sends mail, writes invoices, queries your database, calls external APIs. With guardrails, approval workflows for sensitive actions and complete audit trail. This is what sets us apart from a chatbot shop.
What you get
An agent in a demo is easy. An agent that works correctly over months and doesn't suddenly send wrong invoices is engineering. We build the full stack including safety net.
What it's used for
Agent books flight plus hotel plus rental car plus transfer, sends confirmation to the customer, creates the booking in the CRM, schedules the follow-up call. All in one conversation with the customer, human steps in only for special cases.
Agent registers new client in the accounting system, creates the initial data processing agreement from template, sends the document for signature, schedules the kick-off meeting. Sensitive actions (invoice) need approval from the responsible lawyer.
Customer says cancel my order. Agent cancels in the shop, initiates refund via Stripe or Klarna, notifies the warehouse, writes confirmation mail. All in 30 seconds, no human needed.
Agent registers new patient, checks insurance status via insurance API, books matching slot in the practice calendar, sends SMS reminder 24h before. Escalates to reception for private patients.
Site manager says order material X for site Y. Agent checks stock, looks up the part at three suppliers, compares price and delivery time, places the order, schedules delivery in the site plan. Owner approval for orders above 1,000 EUR.
How it works
One week. We walk through all actions the agent could take with you and one to two key employees. Per action: value if right, harm if wrong, reversibility. From there we derive the approval logic.
We build the function schemas for all approved actions. Each schema gets tests against typical mis-calls (wrong parameters, missing fields, hallucinations). Guardrails configuration defines what's alone vs with approval vs never.
We run the agent in a sandbox with test data. You give us 20-50 scenarios (including edge cases), we test each. Success criterion: 90%+ correct actions, 0% harmful actions without approval.
Agent goes live. Audit trail runs from day one. First 14 days we stay close (daily log review), then we move to weekly monitoring. On drift or anomalies we tune guardrails.
Pricing
Simple setups (two to four tools, one clear workflow): from 2,500 EUR one-off plus 199 EUR/mo hosting. More complex multi-tool agents (five to ten tools, multiple workflows, approval hierarchy): from 4,500 EUR setup plus 299-499 EUR/mo. LLM costs run separately on your OpenAI or Anthropic account. We give you a projection beforehand.
See pricing and packagesFAQ
A chatbot answers with text. An agent executes actions. The chatbot tells you the office hours, the agent actually books you an appointment. Function calling is the technical foundation — the LLM gets tool schemas and can call them instead of just generating text. Claude and GPT-4 both do this well, GPT-3.5 only in limited ways.
Top tier: Claude Sonnet 4.6 and GPT-4 Turbo plus GPT-4o. Both very reliable on simple tools and good on multi-tool orchestration. Mid tier: Claude Haiku 4.5 and GPT-4o-mini (good for simple workflows, less reliable on complexity). Local: Llama 3.3 70B or Mistral Large can do it, but with significantly higher hallucination rate. For production we recommend Claude or GPT-4.
Three safety layers: 1) Guardrails determine which actions are allowed at all 2) Approval workflows require human sign-off for sensitive actions 3) Audit trail documents every action for tracking and possibly reversal. Plus: we build reversible operations wherever possible (e.g. soft-delete instead of hard-delete).
Example: agent wants to send an invoice to a customer. Instead of sending directly, it sends a Slack or Telegram or mail message to the responsible employee with the invoice details. They answer *yes* or *no* or *change X*. Only then does the agent execute the action. Approval latency per action: seconds to minutes depending on setup.
From a GDPR and liability perspective the agent is a tool, you are the responsible party. That's why guardrails plus approval workflows are so important — they show you took reasonable measures. We document the setup auditable. For specific industries (tax, medical, law) we bring in a specialist lawyer before production, that's part of the package.
When the agent calls OpenAI or Anthropic, their data processing agreements plus EU hosting options apply. For API calls to your own systems, data stays in your infrastructure. For third-party APIs (e.g. external booking system) data goes there — we do GDPR mapping per tool before production and lay open the data processing agreements.
Next step
We look at which workflows you want to automate, which actions are sensitive and whether agentic AI is the right lever for your case. Honest assessment instead of sales pitch.
