Build AI agents

Design becomes software. In two to three weeks a working prototype emerges, with real data and a small pilot user group. No PowerPoint mock-up, no “we could build it like this”.

What goes into the prototype:

• Model and API connection (Anthropic API, OpenAI API or a local Ollama instance)
• System prompt with the defined rules and tone
• First tool connections (1–3 tools that are critical for the main flow)
• If needed, a RAG index over the most important documents
• A simple frontend — your own web app, a Slack/Teams integration or Open WebUI
• Logging of every interaction for later analysis

The prototype runs on your infrastructure (or one we provide in the meantime), with real anonymised data — not the sample data from a tutorial. Only then do the actual problems show up.

An AI agent that works for the first ten requests isn't yet production-ready. It's about request number 387, where something unusual happens — and whether it then behaves cleanly.

What we systematically test:

• Happy path — typical requests that come up daily. Must be 100% on the mark
• Edge cases — unusual but legitimate requests. We learn these from pilot logs
• Hallucinations — where does the agent invent something instead of saying “I don't know”? Reduce with RAG source verification and lower temperature
• Prompt injection — can someone make the agent ignore instructions or reveal confidential data? Especially important for external user groups
• Acceptance in the team — is the agent used? Do users understand when to escalate? Do they trust the answers?
• Performance & cost — how many tokens per request? Does that stay within budget as usage scales?

The tests lead to iterative adjustments — to the prompt, to tool permissions, to model choice, to the escalation mechanism. Only when all three stakeholder groups (users, business unit, IT) give the green light do we move on.

Prototype becomes production software. This is where the setup that separates a hobby bot from an enterprise-grade agent kicks in — especially with multiple users or in regulated industries.

Multi-user frontend (when several employees use the agent):

• Open WebUI as the frontend — multi-user login, no per-seat pricing like commercial chat tools
• RBAC — who may use which agents/models follows from the role (sales, accounting, management)
• Shared prompt templates — the sales team can use prepared system prompts without maintaining them themselves
• Audit logs — who sent which prompt when, which tools were called, which data was retrieved? Mandatory for compliance

Deployment options:

• Cloud — when there are no compliance hurdles, the fastest path to production
• Self-hosted on your own server — for regulated industries (law firms, practices, agencies, banks). Fully inside your own infrastructure
• Hybrid — frontier model externally for non-critical tasks, local model internally for sensitive routines. Often the best practical solution

What operations include:

• Monitoring of answer quality — automatic samples with human review
• Cost dashboard — token consumption per role, per use case, per month
• Update path for new model versions — controlled, not automatic
• Escalation workflow — when the agent is unsure, it hands over to a human with the full conversation history
• User training — what the agent does well, what it doesn't, when to push back?