This smell startup says its AI nose is as good as a human’s

WorkAI.TV Editorial Desk
4 Min Read

Share with your CMO

Osmo, a 2023 Google Brain spinout, is making a credible claim that its AI model has reached human-level smell perception, defined as matching the average person at odor detection and a junior perfumer at scent creation. The company built a proprietary dataset of 10 million human smell labels and 800 million molecular “scent tokens” to get there. Its AI-driven fragrance design platform already serves major consumer goods clients, and its model reportedly outperformed in-house perfumers at one of the world’s largest CPG companies last month.

What this means for your business

The fragrance and personal care categories have always been bottlenecked by expert scarcity: a world-class perfumer takes years to train, costs a fortune to retain, and can evaluate only so many formulations per day. Osmo’s model breaks that constraint open in a way that’s directly relevant to any CMO owning a product line where scent is part of the brand experience, which is most of personal care, home goods, food, and hospitality. If the benchmark claim holds, the question isn’t whether AI can do this. It’s whether your fragrance development timeline is about to get cut from months to days by a competitor who moved first.

The “outperformed in-house perfumers” claim deserves scrutiny, partly because Wiltschko is making it about his own company’s technology, which creates the obvious incentive to frame benchmark conditions favorably. What matters strategically isn’t whether Osmo beat a perfumer on one task at one company. It’s the structural argument underneath: 10 million labeled smell descriptions and 800 million molecular tokens represent a dataset that no individual expert or fragrance house has ever operated at, and scale advantages at the data layer tend to compound. The comparison to language models isn’t just a metaphor Wiltschko deploys for press. It’s the actual architecture logic, applying transformer-style (meaning the pattern-recognition engine that powers large language models) pattern matching to molecular sequences instead of words.

The business that should feel this most acutely is any brand paying retainer costs for fragrance development while signing NDAs to protect formulas that an AI trained on enough molecular data could now approximate from first principles. Osmo’s bus-sized smell-printing robot producing prompt-driven custom scents in a week isn’t a curiosity demo, it’s a compression of a proprietary supply chain into a commodity API. The CMO who treats fragrance IP as a durable moat should start asking what that moat is actually made of, because if the answer is “tacit expert knowledge,” that’s the specific thing this technology is designed to dissolve.

Concept deep-dive: Scent tokenization

Scent tokenization is the process of breaking a smell down into its constituent molecules and concentrations, then representing each combination as a discrete unit of data, the way words are broken into tokens for a language model. It exists because raw scent descriptions (“smells floral”) are too vague for a machine to act on, but molecular fingerprints are precise and reproducible. The business implication is direct: once smell becomes structured data, it can be searched, combined, and generated at software speed.

Based on reporting from This smell startup says its AI nose is as good as a human’s, originally published 2026-11-18 05:20:00.

TAGGED:
Share This Article