AI & production
AI agents have gone from experiment to backlog item for most product teams. The problem is "agent" covers everything from a simple FAQ bot to an autonomous workflow that touches your database and sends email on behalf of users.
The difference between demo and production isn't the model — it's everything around it: what it's allowed to do, what you log, and what happens when it gets it wrong.
TL;DR
Key takeaways
- Start with a narrow use case — not "an agent that can do everything".
- Tool calling needs explicit permissions: which APIs, which data, which side effects.
- Log prompts, tool calls, and outputs — you can't debug what you can't see.
- Plan human escalation and fallback from day one — autonomy is a spectrum, not a switch.
What separates demo from production
In a demo you accept hallucinations as "it's AI". In production that's an incident. In a demo the model runs with full context. In production you must limit tokens, filter input, and ensure sensitive data isn't sent to the model unnecessarily.
We build agents as pipelines: classify intent, choose allowed tools, run with timeout, validate output, and escalate if confidence is low or the action is irreversible. It's less sexy than "autonomous agent" — but it's what holds up in ops.
Chatbot vs. production agent
Chatbot / Q&A
Answers questions from your docs — few side effects, easier to ship.
- RAG over documentation or help centre
- No write actions against your systems
- Fallback: link to support or search
- Good first step for many teams
- Limited value if the user still has to do the work afterwards
Agent with tools
Can perform actions — higher value, higher risk, more engineering.
- Calls your APIs: create ticket, update status, fetch data
- Requires auth, rate limits, and audit log per action
- Human-in-the-loop for destructive or high-risk calls
- Observability on every tool call — not just the final answer
- Only worth building when the use case is clear and repeatable
Questions we get over and over
Which model should we use?
The one that matches latency, language, and tool-calling requirements — not necessarily the newest. We evaluate on your real prompts and tools, and keep abstraction so you can switch models without rewriting the whole agent.
Should data be sent to OpenAI / Anthropic?
Only what must be — and under the data processing agreements you have. Many flows can run with anonymised or aggregated context. Sensitive fields are filtered before they reach the model. If data must not leave the EU, we choose host and model accordingly.
How do we measure if the agent works?
Define success per use case: resolved cases without escalation, correct tool calls, time saved for support. Log everything and manually review a sample the first weeks. Automated evals come later — when you know what "good" means.
Can we start without building a full agent?
Yes. We often start with classification + suggested answer or a single read-only tool. That gives learning and logging without opening write APIs on day one.
Considering AI agents for your product?
Let's clarify use case and boundaries — before you ship to users.
We help cut scope, choose architecture, and set up guardrails — so what you ship is something you can stand behind in ops.