The AI that writes a formula from your own shelf

Can an AI write a real perfume formula, rather than a list of suggestions? In PerfuMate it does: AI Studio returns a complete formula, every line a material from your own shelf with a percentage and a role, renormalised to exactly 100% in code and validated on the server before you ever see it. There are three doors into that engine and one set of grounding rules behind all of them, and this page walks both, including what the AI will refuse to do, because the refusals are what make the rest believable. The fuller account lives at /ai; the workflow the AI drops its results into is its own page, and where AI sits in the wider category is on the software page.

Door one: can the AI write a formula from a brief?

Write a brief in plain language and AI Studio composes the whole thing: every line a shelf material with a percentage and a role, and one tap lands it on the bench as a working formula scaled to a batch, with each line's rationale kept as a note. Constraints ride along with the brief: an ingredient cap, styles, creativity, perfume or accord, longevity, intensity, shelf-only, prefer naturals. The conversation is kept, so "warmer, drop the aldehydes" re-runs against the previous answer rather than starting over, and the model's own open questions come back as one-tap prompts.

Door two: can it improve a formula I already have?

From any open formula the AI reads the bench directly, no upload. It answers first as advice: a full critique, make it last longer, what is missing, smooth it out, reduce cost. A second tap turns that same advice into a complete revised copy, with every unchanged line kept at its exact share. The revision lands as a new formula, nothing is saved until you have reviewed it, and the original is never touched.

Door three: can it rebuild a community formula from my shelf?

On any community post, one button hands that formula to AI Studio with the shelf-only constraint on, and what comes back is the same idea recomposed from bottles you actually own. That is deliberate: a post names materials you may not have, so the result is a recomposition rather than a clone. The same door takes a PDF or a pasted formula from anywhere else.

What is the AI allowed to use?

All three doors run on the same rules, and they are enforced in code rather than requested politely. The AI is grounded on your real shelf and saved formulas. The model cannot answer a build request in prose: it is forced to return a structured formula, and with shelf-only on, the server discards the whole answer if any line names a material you do not own. A returned formula that fails validation is re-asked once with the exact faults listed, and it is corrected or refused rather than shipped unverified. Nothing is ever edited in place: a revision is always a new copy you look at before saving. And when a result lands on the bench, every dose is corrected for the strength of the bottle you actually own, so if your only linalool is a 10% dilution the weighed mass is multiplied up to match. The Scent Equalizer feeds the same engine: its slider arithmetic travels as evidence in the brief, so "more mossy at hour eight" is composed as one idea rather than patched line by line.

Taken together, this is probably the most advanced AI shipping in any perfumery formulation tool today. Probably, because nobody has measured every tool, and this page does not claim what it cannot check. What can be checked: none of the tools on the comparison page ship any AI at all, and every mechanism described above is verified against the shipping code.

What will the AI not do?

It does not browse, cannot smell, never edits a formula in place, does not read a GC-MS report, and is not free.

• It does not browse the internet.
• It cannot smell anything, has no feedback from your bench, and does not learn from your results.
• It never edits a formula in place; a revision is always a new copy.
• It does not reverse-engineer a formula from a GC-MS report, and has never claimed to. That answer is a plain no.
• It is not free: one run per request, including a question that returns no formula.

One more boundary worth stating: AI Studio sends your brief and the relevant part of your shelf to the server to compose the answer, and that content is used only to answer your request. Using it means signing in, because the AI has to be able to read your shelf; if you would rather nothing left your device, the rest of the bench works without an account.

What renews, and what does not?

The bench itself is a one-time purchase, never a subscription, and the app is complete without ever buying a single AI run. AI Studio is the one thing that renews, priced separately and by the month, because every run is a live bill to the model on the day it happens. Plans are sold by runs per month, unused runs roll over up to a ceiling of two months of your plan, and whatever you hold when a plan ends stays yours: cancelling stops the next charge and nothing else. Pro includes 2 runs to start and Lifetime Unlimited includes 5, so those two rungs come with enough to judge it. The prices are here.

The claims above are the checkable kind, and the cheapest check is your own shelf. Open the bench, stock it with what you own, and hand the AI a brief.