Skip to Content

The part of your job AI will never take… or will it?

The Innovation Kitchen No. 2 – what a chatbot, a mussel-inspired glue and one missed patent teach us about the human part of innovation work
17 November 2026 by

A chatbot once told an inventor that his invention was "likely first of its kind". It even backed this up with a reference: Larsen and colleagues, 2021, Journal of Adhesive Science. That paper doesn't exist.

The inventor was me, or rather the inventor I played on stage at the SNITTS conference in Gothenburg this autumn. Timo Lagerbjelke from HH Innovation and I did a role-play. A researcher has spent two years on a glue inspired by blue mussels, one that holds tissue and bone together inside the human body, where everything is wet. He has a meeting with his patent attorney in an hour. So he pastes his whole formula into a chatbot and asks: is this new?

The answer was confident, well written and wrong. Mussel-inspired adhesives have been researched for about twenty years, and in our story a patent filed eighteen months earlier was waiting to be found. On top of that, an unprotected invention had just gone into a public chatbot.

The inventor was invented. What happened to him wasn't. Many in the room recognised themselves.

Not a you problem, a structural one

Timo didn't scold the inventor. A general chatbot is brilliant with language, he said, but for this job it has four built-in weaknesses:

  • It produces what sounds right, not what is right.
  • Some chatbots can now search patents, through Google Patents for example, but not nearly thoroughly enough for a real state-of-the-art search.
  • What you type may not stay with you.
  • It knows nothing about your institution.

The research agrees. When Walters and Wilder checked the references ChatGPT gave in literature reviews, more than half were invented with the older model and about one in five with GPT-4. OpenAI's own researchers explain why: models are trained like students sitting a multiple-choice exam, where a confident guess scores better than "I don't know".

The bigger risk is us. In a Microsoft and Carnegie Mellon survey of 319 knowledge workers, the more people trusted AI, the less they checked its work. And in a Harvard–BCG experiment, consultants given a task just outside what GPT-4 does well were 19 percentage points less likely to get it right with AI than without it. The researchers call this the jagged frontier: AI is excellent at some tasks and weak at others that look almost the same.

Why one missed patent is a plate of fries

A proper search needs those four things the other way round: complete patent and literature sources, references you can check, data that stays where it should, and context about your own institution. Timo mentioned aaxon OS as one example of a platform built that way. Full disclosure: it's the one my team builds.

When I first tried our state-of-the-art search properly, I was impressed. That rarely happens to me. I told Rüdiger, one of my co-founders. His answer: "No. It's still crap."

Then he explained. In a state-of-the-art search, missing one single patent makes the whole result useless. You take it to your patent attorney, pay maybe €7,000, and a few days later you learn there's no freedom to operate, all because of the one patent you missed. That risk is just as real with a traditional manual search. Most other AI agents don't have to be that perfect, though. If a funding agent misses one call that might have fitted, it's no disaster, as long as it finds others you go on to win.

Think of ordering a cordon bleu. A missed funding call means the salt is missing. A missed patent means you ordered a cordon bleu and only got the fries.

A machine can't tell you what a mistake costs in your particular job. People like Rüdiger can.

Or will it?

The line keeps moving. In 2025 an AI-written paper passed peer review at a workshop of a major AI conference. This year Google's PaperOrchestra turns messy lab notes into a full manuscript and checks that its references exist. Searching, sorting and first drafts are moving to the machine side.

Other things aren't. Stanford researchers found that AI research ideas were judged more novel on paper, but lost more of their shine than human ideas once researchers actually carried them out. A celebrated MIT study claiming AI made materials scientists far more productive was disowned by MIT in 2025, after someone asked awkward questions about the data. And Germany's Federal Court of Justice confirmed in 2024 that only a natural person can be named as an inventor, even when AI helped find the solution.

When our colleagues at Atrineo mapped the tasks in technology transfer, the ones beyond any current model had three things in common: knowledge that isn't written down, judgement with real stakes, and accountability. Economists Agrawal, Gans and Goldfarb put it neatly: AI makes prediction cheap, and when coffee gets cheaper, the cream becomes more valuable. Judgement is the cream.

So my key learning is this: AI is taking over the searching. People keep the judging, connecting and deciding. The line will keep moving, so the most human skill may be knowing where it is today.

Three things to try

  1. Leadership: add one sentence to your AI rules, "For every decision that matters, a named person signs." And decide in advance what the saved time is for.
  2. TTO: sort your tasks into three piles: the machine does it, the machine does it with a person checking, or people do it. Redo it every six months, and give researchers a safe place to ask AI, so the next invention doesn't end up in a public chatbot.
  3. Innovators: don't paste anything you wouldn't put on a billboard. Click every reference. And when the AI says your idea is brilliant, ask it to argue the opposite.

A short reflection

AI helped me find the studies in this piece in an afternoon instead of a week. I chose which ones were worth your time and read them, and if one turns out to be wrong, that's on me.

Which part of your job would you never hand to AI, and has your answer changed in the last year?

Listen to episode 2 of The Innovation Kitchen with Claudine, my AI sous-chef, above or on Spotify.

Sources

Join the conversation

Comments and discussion happen on LinkedIn, where this edition is part of The Innovation Kitchen newsletter.

Comment on LinkedIn

Martin's podcast voice is an AI version of his own; Claudine is an AI character. Every word is written and approved by Martin. The Innovation Kitchen is supported by aaxon.

17 November 2026
Share this post
The many recipes of innovation
The Innovation Kitchen No. 1 – why most innovation is evolution, not revolution