You have seen the pattern by now. A product you have used for years sprouts a sparkle icon in the corner. Behind it: a chatbot, loosely acquainted with your data, waiting for you to figure out what to ask it. The feature list says “AI-powered.” The workday feels exactly the same.

That is bolted-on AI, and we think it fails for a reason that no model upgrade will fix: it adds a destination instead of removing work. You still have to stop what you are doing, go to the assistant, describe your problem, and carry the answer back. The intelligence is in the building, but the busywork never left.

What the research is actually about

Rigason runs a serious, ongoing research practice in generative AI and agentic workflows. Not to keep up with a trend, but because the questions are hard and the answers decide whether the technology helps anyone:

• When should an agent act, and when must it ask? Filing a document is reversible; sending a letter is not. Getting that boundary right, case by case, is the difference between a colleague and a liability.
• What does an agent need to understand before it touches a document trail? Terms, dates, obligations, and history: the context a new staff member takes months to absorb.
• What makes assistance feel natural instead of grafted on? The interface question, which we think is the hardest one and the most neglected.

The last one is why our researchers sit with designers, not just with models. A capability that interrupts the work is a demo. A capability that dissolves into the work is a product.

The shipping bar

Everything the lab produces is held to one bar: the intelligence must show up as work already done. The mail read and filed before you sat down. The renewal surfaced six weeks before it lapsed. The draft waiting for your judgment, not your keystrokes. No separate “AI corner” of the product to visit, no prompt to compose, no sparkle icon standing between you and the outcome.

And one line we do not cross: the agent assists, the person decides. Every action an agent takes is reviewable, reversible, and attributed, because for teams stewarding records they may have to defend in five years, “mostly right, unsupervised” is not a feature. It is a risk they cannot afford.

If it feels like AI bolted on, it doesn’t ship. That is what we believe superpowers actually feel like: not talking to the machine, but noticing, at the end of the week, how much of the machine’s work you never had to see.