In 2026, software proposes fragrance formulas; it does not smell them. The systems now used inside the large fragrance houses and a handful of independent labs are search tools. They are trained on archives of past formulas and on sales data, they propose combinations of materials in given proportions, and a perfumer then mixes and evaluates a shortlist by nose. Several commercial releases have been developed that way. Nothing in the finished bottle tells you which.
That is the whole answer for a buyer. What follows is what the software is genuinely good at, where it stops, and how to read an "AI-designed" claim when you see one on a shelf in Dubai Mall.
What the software is actually doing
The models are trained on two kinds of data. The first is a formula archive: tens of thousands of past compositions, each a list of materials with their percentages. The second is market data: what sold, where, to whom, and how consumer panels scored it.
From that, the system learns which materials tend to sit together, in what ratios, and which combinations correlated with commercial success. Ask it for "a fresh woody masculine for the Gulf market with strong longevity" and it returns candidate formulas, ranked.
It is good at two things. It can search a very large combinatorial space quickly, which a human cannot. And it can predict, with fair accuracy, what a market has historically liked. Those are real capabilities and they save real time.
What it cannot do
It cannot smell. Every proposal has to be mixed and evaluated by a human nose, because there is still no reliable way to predict the smell of a mixture from a list of its components. Materials interact in ways that are not additive: 2 per cent of one aldehyde can make a rose accord sparkle, and 3 per cent can make it smell of hot metal.
That limitation is fundamental rather than temporary. Odour perception happens at the receptor level in a nose, and the receptors respond to mixtures, not to ingredients one at a time. A model can learn statistical regularities about which pairings worked before; it cannot experience the result.
So the loop remains: propose, mix, smell, adjust. The software shortens the first step. It has not touched the other three.
| Task | Software | Perfumer |
|---|---|---|
| Generating 200 variations on a brief | Minutes | Weeks |
| Predicting what a market liked in the past | Good | Fair |
| Judging whether a trial smells right | Impossible | Essential |
| Reformulating when a material is restricted | Useful shortlist | Final call |
| Deciding what the fragrance is for | None | Everything |
How a perfumer really uses it
The honest description is that it is a search tool with a good memory. A perfumer with an idea can ask for two hundred variations around it and receive a shortlist of ten worth mixing, which compresses weeks of iteration into days.
The second real use is reformulation. When a material is restricted or a supplier changes, the perfumer needs a substitute set that keeps the profile intact. The software is genuinely helpful here, because the question is narrow and the archive contains hundreds of similar swaps.
The third is cost. A model can hold the price of every material and propose a version of a formula that keeps 95 per cent of the character at 60 per cent of the raw-material cost. Whether it actually keeps the character is, again, decided by nose.
The convergence risk
A system trained on what sold will propose what sold. Optimising for historical preference is a recipe for convergence, and the market is already crowded with fragrances that smell like other fragrances: the same ambroxan glow, the same vanilla-tonka base, the same pink pepper opening.
The interesting recent work has come from constraints and intentions, which are human inputs. A perfumer decides to build a whole fragrance from three materials, or to put saffron where nobody expects it. Tygar (AED 120), inspired by Bvlgari Tygar, is grapefruit, ginger and ambroxan and nothing else; that kind of subtraction is a choice, not a statistical average.
This matters in the Gulf more than most places, because the region's taste runs against the global mean. Oud, saffron, heavy rose and dense amber score badly in a European consumer panel and sell out in Riyadh. A model trained on global sales data will steer away from exactly what a Gulf buyer wants.
What it means for a fragrance bought here
For a buyer it changes very little. A fragrance is good or it is not, and the process that produced it is not detectable in the bottle. Two identical formulas smell identical whether a person or a program suggested the ratio.
The test is the same as it has always been. Spray on the inside of the wrist, wait through the first 15 minutes of top notes, then check again at the one-hour and four-hour marks. In a UAE summer, do that once outdoors at 40°C and once in an air-conditioned office at 22°C, because a base that feels quiet indoors can turn heavy in the heat.
An interpretation such as Bleu (AED 120), inspired by Bleu de Chanel Parfum, is judged that way like everything else: does the lemon-zest and lavender opening settle into a cedar, vetiver and sandalwood drydown you want to live with for eight hours? No label about the design method answers that question.
How to read an "AI-designed" label
Read "AI-designed" as a marketing claim rather than a quality signal, in either direction. It does not mean the fragrance is soulless, and it does not mean it is advanced. It means a search tool was used somewhere in development, which by 2026 is true of a large share of mass-market launches whether they say so or not.
What is worth reading instead: the concentration (parfum, eau de parfum, eau de toilette), the note pyramid, and above all how it behaves on your own skin across a working day. Longevity claims of "24 hours" on a box are as unreliable as they ever were, and no algorithm changed that.
Questions people ask
Are there perfumes on sale that were designed by AI? Yes. Several fragrances from major houses since 2019 were developed with algorithmic formulation tools, and the number has grown each year. In every case a perfumer selected, mixed and adjusted the final formula by smell. "Designed by AI" in practice means "shortlisted with software".
Can AI predict what a perfume will smell like? Not reliably. It can predict which combinations resemble ones that worked before, which is a different thing. The smell of a mixture is not the sum of its parts, and no current model replaces a nose evaluating a trial on a blotter and on skin.
Will AI replace perfumers? Not on current evidence. The job has shifted toward briefing, selecting and judging, and away from generating the first hundred trials by hand. The judgement step, which is the one that decides whether a fragrance is any good, remains entirely human.
Does an AI-designed perfume last longer in Gulf heat? No more than any other. Longevity comes from the base materials and the concentration, not from how the ratios were chosen. A program can propose a longer-lasting base, but the same materials were available to any perfumer already, and heat still accelerates the top notes regardless.
Wear what makes you feel remembered.
Gloria Naturals