
The next chapter of AI is specialization
For the past few years, AI has been defined by scale: bigger models, broader capabilities, better reasoning. General-purpose large language models can now write, summarize, analyze and chat about almost anything.
But as AI moves into daily business operations, the question changes. It is no longer what a model can do. It is how reliably it can do one specific job, in one specific industry, under real commercial pressure.
Analysts see the same shift. Gartner lists domain-specific language models (DSLMs) among its top strategic technology trends for 2026 and predicts that by 2028 more than half of the generative AI models enterprises use will be domain-specific. Its July 2026 forecast has spending on domain-specific and specialized generative AI models growing 210% in 2026, from about $1.6 billion to $4.9 billion. It is the fastest-growing segment in the forecast, and the direction is clear.
I also think “domain-specific” is on its way to becoming a label vendors attach to anything. So it is worth being precise about what it means, and about why beauty is where it matters most.
What domain-specific actually means
A general-purpose model is built for breadth. It can explain a mortgage, debug code or help someone pick a lipstick in the same conversation. A domain-specific system is built for depth: relevance, accuracy and efficiency inside a defined area.
That does not always mean training a new foundation model. A strong domain-specific system usually combines an existing LLM with specialized models, structured knowledge, retrieval, proprietary data, business rules and purpose-built agents. The useful question is never “did you train your own model?” It is “which architecture gives the most reliable result for this task?”
Why beauty demands more than generic intelligence
On the surface, a beauty catalog looks like any other ecommerce assortment: products, prices, categories, stock. Underneath sits a dense web of ingredients, formulations, textures, finishes, shades, skin characteristics, preferences and intended results.
Take a shopper looking for a foundation. She wants a natural finish, light coverage, a specific undertone and a formula that works with her existing skincare. Climate, occasion and past disappointments matter too.
A general-purpose assistant will give her a convincing answer. A convincing answer is not the same as a suitable product. To get it right, a system has to connect three kinds of knowledge:
- Product intelligence: formulations, shades, textures, finishes, compatibility.
- Customer intelligence: skin characteristics, stated preferences, goals, shopping context.
- Commercial intelligence: assortment, availability, pricing, promotions, retailer rules.

These cannot be treated separately. The perfect match may be out of stock in her market. A bestselling formula may give the wrong finish. Two products can share a category label and serve completely different needs.
In beauty commerce, relevance is not a similarity score. It is a relationship between the product, the person and the context.
Here is the test I find most telling. A general chatbot almost always has an answer. A good beauty system sometimes has to say: the shade that matches your undertone isn’t available where you shop, or the finish you want comes in a different formula. Knowing when the best match isn’t on the shelf, and what the next-best option really offers, is a domain skill.
From behavior to intent
Traditional ecommerce personalization runs on behavior: what you browsed, what you bought, what similar customers bought. Those signals still matter, but they leave out the why. A past purchase doesn’t tell you whether the product worked.
Conversational and camera-based experiences add the missing context. A shopper can say what she wants, what she has tried and what she disliked. With consent and proper safeguards, a scan can add observable signals about her skin. When those signals meet structured product knowledge and the retailer’s own rules, the system can do more than recommend what similar shoppers bought. It can work out what suits this shopper, explain why, and guide her within the assortment that is actually available.
ARview does not store images. Each image is processed in real time to generate insights (metrics), and only those insights are kept and passed to the next agent.
Why one chatbot is not enough

Different shopping tasks need different capabilities. Virtual try-on needs realistic visual processing. Skin diagnostics needs computer vision and carefully defined interpretation. Routine building is about sequencing and compatibility. Gifting is about occasion and recipient.
A chat window can be the shared front door, but the tasks behind it are not interchangeable. ARview is one example of building this way: Aria, a domain-specific AI model trained on each brand’s own data, powers 8 specialist agents. They are Try-on, Skin Diagnostics, Routine, Personalization, Basket Building, Loyalty, Gifting and Fragrance.
Aria works as part of a hybrid system, alongside the other components. Computer vision handles skin and face analysis. Structured product data and retrieval supply facts about the catalog. Explicit rules, such as which ingredients address which concerns and what a brand is allowed to claim, set the boundaries. Language models rank options and explain them within those boundaries. Each component does what it is best at, so no single model has to carry everything.
Where the advantage really sits
As capable foundation models become available to everyone, general capability stops being the differentiator. Models will still differ in reasoning, cost and reliability, but for many everyday tasks the leading ones are increasingly interchangeable. The gap that matters moves from the model to what surrounds it: the knowledge, data and workflows a business builds on top.
Beauty brands and retailers start with real assets: product knowledge, formulation expertise, assortment logic and years of merchandising judgment. Most of it is scattered across product databases, documents and experienced people. Making it usable by AI is the first step, and it is where many projects stop.
There is a further point that gets lost in the “beauty AI” conversation. Brands don’t need AI that understands beauty in general. They need AI that understands their version of it. Two brands can sell similar foundations to different audiences, with different formulations, shade philosophies, price positioning and ways of talking to customers. There is no universally correct recommendation, only the right one for this brand’s product truth, commercial strategy and customers. That is why dropping a generic beauty chatbot onto a website is not the same as building domain intelligence.
The bigger advantage comes next. Every conversation, scan and try-on shows the brand something a purchase record can’t: why a shopper rejected a product, which attributes matter to which customers, where the assortment has gaps, what stopped a sale. A competitor who licenses the same foundation model cannot license that history. The platform learns continuously from scans, try-ons and purchases, and each retailer’s data stays inside its own tenant ecosystem. It compounds, provided the findings feed back into real decisions about assortment, merchandising and service.
Specialization is not a guarantee of quality, though. A domain-specific system has to prove its recommendations are reliable, not just look more sophisticated. Data quality, evaluation, governance, privacy and human oversight still decide whether it works.
Four questions to ask any vendor
If you are evaluating “beauty AI” this year, start here:
- What exactly is specialized? The model, the data, the rules, the agents, or just the prompt?
- Whose knowledge does it run on? Can it work from your catalog, your shade logic and your commercial rules, or only from generic beauty content?
- How do you measure that it is right? Ask for evaluation methods and results, and look at what happens when the best match isn’t available.
- What intelligence does every interaction generate, and who owns it? Ask what insights are captured, how consent is handled, and how those insights inform better recommendations, assortment and merchandising decisions.
From knowing everything to understanding what matters
General-purpose AI makes intelligence accessible. Domain-specific AI makes expertise operational. Beauty doesn’t need AI that can talk about beauty. It needs AI that understands products, people and commercial reality together.
The real advantage isn’t teaching AI what a brand already knows. It’s letting the brand learn what it couldn’t previously see.