It is now easy to attach a language model to almost anything. It is much harder to know where it makes a system genuinely better for the people relying on it. We would rather be slow and right here than fast and decorative.
Good fits
- Turning unstructured input into structured records that a person then confirms.
- Summarizing across many records so a decision-maker sees the shape of things.
- Helping people find and understand what a system already knows.
- Drafting, where the human remains the author.
Poor fits
- Silent automation of decisions with real consequences for people.
- Any flow where the model's confidence is presented as certainty.
- Replacing clear, deterministic workflows that already work.
Our research interest is in the first list. Specifically, in how AI-assisted features can sit inside operational systems with clear boundaries, visible provenance and a human in control. That is where we are running experiments, and we will write about what we learn, including what does not work.
Note: This is a statement of research direction. It does not describe shipped functionality.