What We Automate, What We Don't, and Why It Ends Up in the Bottle

This company has no back office. The bookkeeping, the document filing, the monthly close and the calendar of regulatory deadlines all run on a schedule, without anyone starting them. The perfume does not.

People ask how a company this small handles the paperwork of selling a regulated product across borders. The honest answer is that we drew a line early — this side is a machine's work, that side is a person's — and we have been strict about it since. This is where the line sits today, and what holding it actually pays for.

The cost nobody smells

Every bottle of perfume you buy has to pay for several different things at once: the materials, the bottle and box, production, transport, tax, whatever the retailer takes, and marketing. Those are the costs people usually talk about.

There is one more, and it is the one nobody smells: the cost of the company existing at all.

Selling a fragrance across borders means somebody has to do the bookkeeping. Somebody has to file the annual accounts, hold the shareholders' meeting on time, submit the financial statements to the register, renew the import VAT authorisation before it lapses, make the packaging declaration in France, check twice a year whether the EU cosmetics rules have changed under your product. Somebody has to name and file every invoice so that it can be found two years later by an auditor. Somebody has to answer the email.

None of that work touches the perfume. All of it has to happen, or you cannot legally trade.

The traditional answer is to hire — an assistant, a bookkeeper, an operations person, an agency on retainer. Each one is a fixed monthly cost that arrives whether or not you sold anything that month. And every one of those costs has to be recovered somewhere. It is recovered in the price of the bottle.

The industry is automating the other end

It is worth being clear that we are not describing something unusual. The largest fragrance companies in the world are automating too — they are simply pointing it at a different part of the work.

In October 2025 The Estée Lauder Companies opened a Global Fragrance Atelier in Paris, inside La Maison des Parfums on Rue Volney. By its own description it runs an AI-enabled, end-to-end creation process, combining olfactive and neuroscience modelling with real-time monitoring of patents, research and regulation. The stated goal, in the company's own press release, is "reducing fragrance development lead times by up to 30–50% in the coming years" — for brands including Jo Malone London, Le Labo, TOM FORD and KILIAN PARIS.

That is automation aimed at creation, in order to go faster.

We have aimed ours at the exact opposite end, for the exact opposite reason. We are not trying to make perfume faster. We release rarely, deliberately. What we wanted removed was everything that is not the perfume.

List one: what runs without us

These things happen on a schedule. Nobody remembers them, because nobody has to.

• Accounting intake. Documents go into a single folder. They get read, renamed to one standard convention — supplier, account category, description, period, invoice number, amount and currency — and filed into the right month by issue date. A filing system is only useful if it is boring and identical every time, which is precisely what a person is worst at.
• The monthly close. On the 11th of each month, the previous month's invoices and bank statements are read and turned into a report. Nobody requests it. It appears.
• The regulatory calendar. Annual accounts and the corporate tax return in March. The shareholders' meeting in June. The financial statements filed to the register in July. The French packaging declaration in February. The import VAT authorisation renewed every six months. A twice-yearly check of whether EU cosmetics regulation has been amended in a way that touches our formula. Small companies do not usually miss these deadlines because the work is hard. They miss them because a human being forgot.
• Watching. A job checks a public government register every three days to see whether our dangerous goods shipper application has been entered yet. It will keep checking until it appears, then tell us and switch itself off. That is a job no person should ever be given.
• A daily digest. Sales and sign-ups, one message a day.
• First drafts and research passes. Not the final word on anything — the first pass, so that the human time goes into judgement rather than into starting.

List two: what we will not hand over

This list matters more than the first one, and it is shorter.

The perfume. The formula is made by a perfumer — a person, in a laboratory, with a nose that has been trained for decades. There is no version of this where a machine composes our fragrance. We have never been tempted, and if you had asked us to explain why in one sentence, it would be that machines do not smell anything.

The decision about what to make. What a fragrance is for is not a technical question and does not have a correct answer. It is a choice, and choices belong to people who have to live with them.

Saying no. Rejecting a trial you have waited months for is not a task. It is a judgement, made against something you are trying to reach, and it costs money every time.

The relationships. Suppliers, laboratories, freight agents, customs brokers. Everything genuinely difficult in this business has been resolved by a specific person choosing to help us, usually because they had spoken to us before.

Signatures. This one is not a philosophy, it is the law, and it is the clearest illustration we have. When we send perfume by air, the shipper's declaration must be signed by a named human being — the qualification belongs to a person and a limited company cannot hold it. Similarly, our sales invoices are still issued by hand, one at a time, into the national tax system, because that requires logging in with personal credentials. We do not give credentials to software. That rule is written down internally, and we would keep it even if the law relaxed.

Anything we say in public. Every external claim in this journal is checked against a primary source by a person before it is published. Which brings us to the part that is actually useful to anyone else considering this.

The failure mode, honestly

If you are going to run a company this way, you need to know how the tools fail, because it is not how software used to fail.

Old software failed loudly. It crashed, or it returned an error, and you knew. These tools fail confidently. Ask a language model about a regulation it half-remembers and it will hand you a clean, well-organised, correctly formatted paragraph with an article number in it, and the article number will be wrong. Nothing in the output looks any different from the output that is right.

We have caught this happening. The rule we run now is a direct result: for anything outward-facing, legal, or regulatory, the model may find the source, but the model may never be the source. We open the actual law, the actual standard, the actual manufacturer's document, and check the number against the text. When we cannot verify something, we either drop the claim or we publish and say plainly that we could not verify it — which is why you will occasionally see us do exactly that in these articles.

That rule exists because we did not always have it.

What the line actually buys

Here is the whole argument in one paragraph.

Every recurring cost this company does not carry is a cost that does not have to be recovered in the price of a bottle. We have no payroll for administration, no agency retainers, no back office. What that changes is not a number on a spreadsheet — it changes the first question we ask about a raw material.

A company carrying a large fixed overhead has to ask "what fits the budget" before it asks "is this the right one", because the overhead arrives every month whether the perfume is good or not. We get to ask the second question first. That is why we can build a formula around expensive naturals, in quantities that would be difficult to justify if we were also paying for an office of people to keep the lights on — and it is the same reason we chose the heavier board for a shipping box that most customers will throw away, and why the website works the way it does.

We are not claiming this is virtuous. It is a trade, and it has a cost of its own: with no back office, there is nobody to absorb a mistake, and the failure modes described above are ours to catch. We think it is the right trade for what we are trying to make. We would not pretend it is free.

What happens next

Two things are in progress. Invoice issuance will be wired into the national e-invoicing system once it goes live for us, which removes the last piece of routine typing from the finance side. And the regulatory watch is getting wider, because the lesson of the past year is that the rules move faster than anyone's memory of them.

Nothing on the second list is moving. It is not going to.

The short version

• Every bottle has to pay for the company that made it, not just the materials. Administration is a real cost and it is invisible in the product.
• What we automate: bookkeeping intake and filing, the monthly close, the regulatory deadline calendar, monitoring public registers, daily reporting, first drafts.
• What we don't: the formula, the decision about what to make, rejections, supplier relationships, signatures and credentials, and any factual claim we make in public.
• The characteristic failure of these tools is confidence, not collapse. Our working rule: the model may find the source, but never be the source.
• The point of all of it is the first question we get to ask about an ingredient: not what fits the budget, but is this the right one.

If you would like to follow the rest of it, our weekly newsletter The Hidden Nose goes out once a week; the link is in the footer.