AI
Integrating an LLM into your existing application
The API call to the model is the easy part. The integration around it decides whether it amounts to anything: model choice, cost, quality monitoring, and the boundary of what you trust the model with.
Model choice is a disposable decision. Build it that way.
Whichever model is best for your task today will be outdated in a quarter. Models improve and get cheaper faster than any roadmap keeps up with. The most important architectural choice is therefore not which model you pick, but that your application must not notice when you switch. We build LLM integrations model-independent: changing model or provider is a configuration choice, not a rebuild, and each task can run its own model. The expensive model for the hard work, a cheap one for the routine work.
Control cost per task, not per month
LLM costs are variable costs: every call costs money, and an integration that makes ten calls per document costs ten times the one that does it in one. We calculate the cost per processed unit up front, an e-mail, a document, a search, and watch it in production. Then "does this pay" is arithmetic instead of a surprise on the invoice.
Guard quality with a yardstick, not a feeling
An LLM fails differently from ordinary software: not with an error message but with a convincingly wrong answer. Every integration therefore gets an evaluation set: fixed examples with known good outcomes, run against every change, different model, different prompt, different data. You see quality decay before your users do, and you dare to switch when a better model appears.
What you do not delegate to an LLM
Our boundary is simple: the model prepares, a person decides wherever the outcome is irreversible. Commitments to customers, payments, legal conclusions, anything where a mistake costs money or trust, gets an explicit approval moment. And everything the model does is recorded, so afterwards you can answer what happened and why.
From production
Through our workflow platform Supertaak, an AI-driven e-mail automation runs at a client: incoming mail is read, categorised and prepared, and a person signs at the moments that matter. Every run leaves a record. It is the kind of integration described above, for real and in production.
Related: RAG implementation for the search side, agentic workflows for the process side, and what we build for the cases.
An LLM in your application?
Tell us what your application does and where the language work sits. Within two working days we give an honest first assessment, cost side included.