The Korea Times reported on 28 July 2026 that Amorepacific has deployed a generative AI research assistant, built with telecom group KT and named Lemon, across its cosmetics R&D and regulatory functions. Source: https://www.koreatimes.co.kr/business/companies/20260728/kt-amorepacific-launch-ai-platform-to-speed-up-cosmetics-rd, with corroborating coverage in the Seoul Economic Daily and Personal Care Insights the same week.
The figures reported are specific. Research and regulatory review tasks that previously took about 12 days now take roughly five minutes. Reviewing the research history of an ingredient has gone from about six days to around five minutes. The system is built on more than 70 years of Amorepacific R&D records.
Most coverage of this will focus on the AI. That is the least interesting part.
The archive is the asset
A generative model with nothing behind it returns plausible sentences. A generative model sitting on seven decades of structured formulation data, stability testing, ingredient decisions and regulatory submissions returns answers that a formulator can act on.
The reason a 12-day task collapses to five minutes is not that the model thinks quickly. It is that the answer already existed somewhere in Amorepacific's records, and the 12 days were spent by a human finding it, reading it, and reconciling it with the other places it appeared. The model removed the search cost from an archive that was already complete.
This is why the story is worth reading for a brand doing £2 million rather than several billion. You are not going to build a 70-year archive. You are, right now, building a five-year one, and the decisions you make about how you keep it will determine whether tools like this are available to you when they become affordable.
What a small brand's version of that archive actually is
Strip away the scale and the categories are the same for everyone.
Formulation records: the current formula for each SKU, every prior version, and the reason each change was made. Most small brands have the current formula and none of the history, which means every time a supplier changes or a cost pressure appears, the reasoning behind past decisions has to be reconstructed from memory.
Ingredient and supplier documentation: specifications, certificates of analysis, allergen declarations, origin documentation. Usually scattered across email and whatever the contract manufacturer happens to hold.
Safety and regulatory: the safety assessment and Cosmetic Product Safety Report per product, notification records, and any correspondence with regulators or retailers about compliance.
Claims substantiation: for every claim currently on pack or on site, the specific study, test or evidence that supports it, and the date it was produced. This is the file that gets requested at the worst possible time, usually by a retailer during onboarding or a regulator during a review.
Testing: stability, compatibility, preservative efficacy, consumer testing. Including the ones that failed, which are often more useful later than the ones that passed.
Why the failed tests matter more than founders expect
The instinct is to keep what worked and discard what did not. That instinct is expensive.
A record that a particular preservative system failed stability in a specific pack format is the thing that stops you repeating the same test in two years when a cost pressure makes that system attractive again. A record that a fragrance caused a compatibility issue with a particular closure is what stops you re-tooling into the same problem.
Amorepacific's system is useful precisely because it can answer "what happened last time we looked at this ingredient" across seventy years, and a meaningful share of what happened last time was a dead end. Dead ends that are written down stop being repeated. Dead ends that live in a former employee's memory get repeated roughly every three years, at full cost.
The realistic timeline for smaller brands
Nobody at £500k-£5m is deploying a bespoke AI research platform. That is not the decision in front of you.
The decision in front of you is what happens when your contract manufacturer, your regulatory consultant or your lab starts offering AI-assisted formulation and regulatory search as a service, which on the current direction of travel is a question of the next couple of years rather than the next decade.
At that point, the brands who benefit will be the ones who can hand over a coherent set of records. The brands who cannot will either pay someone to reconstruct their own history first, or get a service that is only as good as the fragments they can supply.
Reconstructing five years of formulation history after the fact is genuinely difficult and sometimes impossible. Suppliers change. Consultants move on. Contract manufacturers get acquired and archives get culled. The window to collect a document is while the relationship that produced it still exists.
The claims file is the urgent one
If you only do one part of this, do the claims substantiation file.
Regulatory attention on cosmetic claims has been tightening across the UK and EU for several years, and retailer onboarding processes increasingly require substantiation up front rather than on request. The situation that damages brands is not having made an unsupportable claim. It is having made a supportable claim and being unable to produce the support within the window a retailer or regulator has given you.
Every live claim, mapped to the specific evidence behind it, with a date. Claims where you cannot find the evidence are not a filing problem, they are a live commercial risk, and finding them now is considerably better than finding them during a listing review.
The honest read on this story
Amorepacific's five-minute number is a good headline and it is not transferable. No small brand is going to replicate it, and any consultant who tells you otherwise is selling something.
What is transferable is the recognition that the speed came from the archive. The AI was the retrieval layer on top of decades of disciplined record-keeping, and without the records underneath it would have produced nothing worth reporting.
The version of this available to a £2 million brand costs nothing but attention. Decide where the records live, put them there consistently, and keep the ones that failed alongside the ones that worked. In three years that is either a genuine asset or a reconstruction project, and the difference is decided by what you do this quarter.