AI for Cosmetic Formulation: What It Can Draft and a Chemist Must Check

Yes, AI can help a small cosmetics brand with the thinking and paperwork around a formula: briefs, ingredient options to discuss, tidy ingredient lists and batch sheets. It cannot replace the chemist who judges stability and preservation, or the qualified safety assessor who signs the safety report that many markets require before sale.

The useful way to think about it is a split. AI drafts and organises; a person with the right qualifications decides what is safe, stable and truthful. Most of the risk sits in the gap between a formula that looks right on screen and one that has been made, tested and assessed.

Follow me on Instagram@sagnikteaches

What does a cosmetic formula project actually involve?

A typical small-brand project runs through six stages: a concept brief; a draft formula with a chemist or contract lab; lab samples and revisions; stability and preservation testing; a safety assessment and product information file; and a scaled-up batch with its records and labels. Each stage has a different owner.

Connect on LinkedInSagnik Bhattacharya

Can AI write the formula itself?

It can produce something that looks like a formula, and that is the danger. A chat assistant has read a great deal of text about cosmetics and will happily output a table of phases, ingredients and percentages that adds up to 100. What it cannot do is know whether that mixture will emulsify, stay stable on a warm shelf, resist microbial growth for the life of the pack, suit the packaging, or be tolerated by skin. Several of those only reveal themselves in a lab and over weeks.

Subscribe on YouTube@sagnikteaches

A realistic example. A founder asks for "a vitamin C day cream with natural preservation". The assistant produces a neat formula with a plant-based preservative at a plausible level. Nothing in the answer tells the founder that vitamin C ingredients can be unstable in water-based products, that many "natural preservation" ideas fail challenge testing, or that the pH needed for some actives conflicts with the preservative's working range. A chemist knows to raise those; the assistant's confident table hides them. Treat any AI formula as a conversation starter to bring to a chemist, never as something to make in a kitchen and sell.

There are specialist formulation tools that suggest ingredients and ratios from databases. Treat these the same way: the output is a hypothesis, and a person responsible for the product still tests it. If you are weighing a tool, ask what data it is built on, whether it checks ingredient restrictions for your markets, and what happens to your inputs.

Where AI is useful at the concept stage

This is the best place for AI, because mistakes cost nothing yet. A good brief saves expensive back-and-forth with a lab or chemist. Ask the assistant to interview you.

I run a small skincare brand. I want to brief a cosmetic chemist
on a new product. Interview me one question at a time to build the
brief. Cover: product type and format, target user and skin type,
texture and finish, fragrance or fragrance-free, packaging type and
size, retail price and target cost of goods, markets I will sell in,
ingredients I want or must avoid and why, claims I hope to make
(I know they must be provable), and the launch date.
When I say "done", write a one-page brief with a table of
requirements and a list of open questions for the chemist.
Do not suggest ingredients or percentages.

The last line matters. You want the assistant to organise your thinking, not to pre-empt the chemist. A completed brief (illustrative) might read: "Face moisturiser, lotion format, normal to dry skin, light non-greasy finish, fragrance-free, 50 ml airless pump, retail $28, target ingredient and packaging cost under $4.20 a unit, markets: two named regions, avoid mineral oil, hoping to claim 'moisturising', launch in nine months." Then open questions: which preservative system suits the packaging, whether the airless pump is compatible, and what minimum batch size a lab can make.

That brief takes about 20 minutes and is worth more than a page of AI-invented formula.

What an ingredient list needs, and where AI helps

Once a chemist has a formula, the pack needs an ingredient list in the standard naming system, in the right order. AI is helpful for the clerical work: checking that names are spelled consistently, that the list matches the formula sheet, and that a supplier's change has been reflected. It is not the authority on what must be declared, how to treat fragrance components or what the legal order should be; check against the rules for your markets and your chemist's sheet. The tutorial on INCI ingredient list checks with AI covers this comparison in detail.

A worked check (illustrative). The formula sheet has twelve lines; the label draft has eleven names. Ask the assistant to compare them and report any ingredient in one list but not the other, any spelling differences and any ordering differences. It finds that a pH adjuster used in small quantities is on the sheet but missing from the label. A person confirms with the chemist whether it must be declared, then fixes the label. The assistant did the comparison in seconds, and a human made the judgement.

Batch sheets: the arithmetic AI handles well

Scaling a lab formula to a production batch is arithmetic, and arithmetic with many lines is where people make slips. Give the assistant the approved formula in percentages and the batch size, and ask for weights. Then check by adding the weights and confirming they equal the batch size, which takes a calculator and thirty seconds.

Say the approved formula (illustrative, shortened) is: water 70.4%, glycerin 4.0%, emulsifier 3.5%, oil phase blend 12.0%, thickener 0.4%, preservative system as per supplier, 1.0%, other 8.7% in lines you list. For a 20 kg batch, glycerin is 20,000 g x 0.04 = 800 g and the thickener 80 g. A good batch sheet adds columns the formula lacks: the phase each ingredient belongs to, the supplier lot number field, a weighed-by and checked-by signature box, the order and temperature of additions copied from the chemist's method, and the time of each step.

Ask the assistant to build the sheet as a table and not to alter the method. Then compare it against the chemist's document line by line before it is used. A mistake to watch for: percentages in a formula rarely sum to exactly 100 if rounded, and a model may "fix" this by adjusting water. Never let it adjust a quantity. If the numbers do not add up, the chemist corrects the formula.

File the signed sheet with the batch's supplier lot numbers and a retained sample, so a later complaint can be traced to the exact materials and the people who weighed them.

Safety, stability and claims: what stays with qualified people

Three areas never leave qualified hands, whatever the AI produces.

Safety assessment

Many markets require a safety assessment before a cosmetic product goes on sale. In the EU, for example, Regulation (EC) No 1223/2009 says the responsible person must ensure, before placing a cosmetic on the market, that it has undergone a safety assessment and that a cosmetic product safety report is set up. The assessment considers the formula, the intended use, the exposure and more, and is done by a person with the qualifications the rules require. Rules differ between markets, so read the regulator's pages for each place you sell, and ask your assessor what they need. AI can help assemble the documents the assessor asks for; what an assessor signs, and what AI can and cannot write goes further.

Stability and preservation

Stability testing looks at whether the product stays acceptable in its pack over time and across temperature conditions: appearance, smell, viscosity, pH, separation. Preservation testing (often called challenge testing) looks at whether the product resists contamination with microbes. Both take weeks, use a lab or a trained person, and have pass criteria set before the test, not after. AI can draft a test plan outline or a table to record readings, but it cannot run the test or interpret a borderline result, and a wrong answer here can mean a contaminated product on a customer's face.

Claims and labels

Every claim on the pack and the page needs proof you hold. The routine in what AI copy must not say about skincare applies, and so do the label rules for each market, which a cosmetic label checklist sets out. The brief you wrote at the start should list the claims you hope for; the chemist and assessor tell you which the evidence will support.

Comparing lab samples without losing track of them

A chemist usually sends three to six sample versions before one is approved. Each comes with a short note, and you add your own impressions. By version four, nobody remembers which one felt tacky. A table fixes that, and AI can build it from your messy notes.

Below are my notes on four lab samples of a face lotion
(versions A to D). Make a table with one row per version and
columns: what changed from the previous version (from the chemist's
note), texture on application, finish after 5 minutes, smell,
appearance in the jar, my verdict. Use only what the notes say.
If a cell has no information write "not noted". Do not add opinions.

An illustrative result: version B changed the thickener level, "texture slightly stringy", finish "tacky", smell "not noted", verdict "no". Version C changed the emulsifier, "smooth", "light", verdict "keep". The gaps marked "not noted" become your questions for the next session. The chemist then confirms the changes between versions, because your notes of what changed may be incomplete, and the formula is the chemist's record, not the table.

Questions to put to a contract lab or chemist

A small brand often meets a lab for the first time with a vague idea and a budget. AI can turn your brief into a question list so the first meeting is productive. Ask for questions in four groups and take the list with you.

• Scope: do you formulate from a brief, or only manufacture a formula I supply? Who owns the formula? What is the minimum batch size?
• Testing: which stability conditions and how long? Is preservation testing included or extra? Which pass criteria do you use, and can I see them before testing starts?
• Documents: what do you provide for the safety assessment: ingredient specifications, certificates, test reports? Can you recommend a qualified assessor, or does the lab have one?
• Changes: what happens when I want to change a raw material or a supplier? Who re-tests?

An honest warning on cost and time: formulation, testing and assessment take months, not days, and prices differ widely between labs, so ask for quotes in writing and compare line by line. An assistant can lay two quotes side by side in a table, which is a good use for it, but it cannot tell you which lab is competent. Ask for references and samples of past work.

Supplier changes: the quiet risk

A formula is only as stable as its raw materials. A supplier changes a grade, a plant extract's composition shifts with the harvest, or a substitute is offered when a material is out of stock. Each change can affect the texture, the preservation or the safety assessment. A small brand should keep a change log: date, material, what changed, who approved, whether re-testing was needed. AI can read a supplier's change notice and summarise the difference in plain words, and it can check whether the new specification sheet has different or additional components than the old one by comparing the two documents.

Illustrative case: a supplier emails that the thickener will now come as a pre-dispersed liquid, not a powder. The assistant summarises the notice and compares the two specification sheets, noting that the new version contains a small amount of a preservative carrier not present before. That flag goes to the chemist and the assessor, who decide if the ingredient list, the preservation system or the assessment changes. Nobody changes the batch sheet until they answer.

Six safe prompts to start with

1. "Interview me to write a one-page product brief; do not suggest ingredients."
2. "Turn these lab-sample notes into a comparison table; mark gaps as 'not noted'."
3. "Compare this formula sheet with this label ingredient list and report every difference."
4. "Scale this approved formula to a 20 kg batch; show each weight and the total, change nothing else."
5. "Summarise this supplier change notice in 100 words and list what to ask my chemist."
6. "Build a checklist of documents an assessor typically asks for; I will confirm it with them."

Notice what is absent: nothing asks the model to choose ingredients, decide safety, pass a test or approve a claim. That is the line, and keeping to it is what makes the time AI saves safe to take.

What goes wrong when AI is used as the chemist

1. A formula that separates in the jar. A founder scales an AI formula and finds the oil layer rising after a fortnight. Cost: a batch of packaging and ingredients, plus a delay.
2. A product that grows mould. The preservative level looked reasonable on paper but was never challenge-tested. This is the risk that matters most, because it harms customers and the brand together.
3. An ingredient that is restricted in one market. The assistant's formula contains something that is limited or banned where the founder sells. Each market publishes its own lists, and a model's memory of them may be out of date.
4. A "supported" claim that nobody supported. The assistant writes that an ingredient "reduces the look of fine lines". The formula has too little of it for any such effect to be shown.
5. A leaked formula. A full recipe pasted into a free consumer tool. Use a business plan, or share only structure.

A sensible workflow for a small brand

This sequence uses AI heavily in the dull parts and keeps people in control of the risky parts. Times are illustrative.

1. Brief (one afternoon). Use the interview prompt. Review the brief with your chemist or contract lab.
2. Formula (chemist's time). The chemist designs and revises samples. You keep a log; AI can tidy your notes of each sample's feel and look into a consistent table so you can compare versions.
3. Test plan (a day). Ask your lab what they will run. AI turns their email into a one-page schedule with dates and owners.
4. Assessment pack (a few days). Use AI to build a checklist of documents the assessor needs: formula, supplier specifications, packaging details, test results, intended use. You send; they assess.
5. Batch sheet and label (a day). AI drafts the batch sheet from the approved formula; the chemist approves the method; the ingredient list is compared against the formula sheet.
6. Launch checks. A person reads every claim against the proof table, and a retained sample from the first batch is stored.

Whether this saves money depends on the size of the project, but it tends to save chemist hours, because you arrive with a clear brief, a clean data table and no basic questions left to ask. That is a sensible use of expensive time. Six jobs a small skincare brand can delegate lists the wider set of tasks around a product, and AI for beauty brands: copy, photos and customer questions follows the finished product into the shop.

If you want help mapping which of these steps to put AI on, and how to keep your chemist and assessor at the centre, that is what my AI implementation consultation covers, and you can reach me through the contact page.

Questions founders ask about AI and formulas

Can I sell a formula an AI wrote?

Not on the strength of the AI alone. A formula needs a qualified chemist's review, a safety assessment by a qualified assessor where the rules require one, stability and preservation testing, and a check against the ingredient restrictions of each market you sell into. The AI draft is a starting idea, nothing more.

Which parts of formulation is AI good at?

Writing a clear brief, listing options to discuss with a chemist, tidying an ingredient list into a consistent format, turning a formula into a batch sheet with arithmetic, and explaining jargon. It is weakest where physical behaviour matters: emulsion stability, preservation and skin tolerance.

Is it safe to paste my formula into a chat tool?

A full formula with percentages is your key trade secret. Use a business plan where content is not used for training by default, or share only structure and ingredient names. Ask your supplier or lab about confidentiality before sharing anything with third parties.

Does a safety assessment have to be done by a person?

In regimes such as the EU Cosmetics Regulation, the safety assessment and report are prepared by a qualified safety assessor before the product is placed on the market. AI can help organise documents for that person, but it cannot be the assessor.

Further reads

• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Regulation (EC) No 1223/2009 on cosmetic products, safety assessment provisions (read on EUR-Lex, 5 November 2026); the author's guidance. Rules differ between markets, so check the regulator pages for each place you sell.