We'll show how it works in Enad, but most of it applies whichever tool you use.
Why AI copy tends to sound the same
Anyone who has tried AI for product copy knows the result. The text is correct and fluent, but it could have been written for any of your competitors. The jacket is "perfect for every adventure" and the chair "combines form and function".
The model is rarely the problem. An AI that knows nothing about your brand writes like an average of everything it has read, and that average sounds like an advert. It doesn't know that you always say "jacket" and never "coat", that collection names stay in English, or that you never promise anything about materials that isn't in the product data.
The usual fix is a better prompt. Someone writes a long instruction, pastes it in when it's time to translate and gets a better result. Next time a colleague does the translating, with a different version of the prompt, and the tone drifts apart again.
Prompt or profile
The difference between a prompt and a profile is where the instruction lives.
A prompt belongs to one occasion. It is written when someone needs it and disappears when the job is done. A profile is a saved instruction that lives in the system and is used every time the AI writes something for you, whoever presses the button.
In Enad these are called AI profiles. A profile describes how your brand sounds, which words you use and which rules always apply. You choose it when you generate a product description or an SEO field, when you translate a product into one language or all of them, and when you run an AI job across many products at once.
Your tone of voice then becomes something you maintain alongside the product data. It no longer depends on whoever happens to write the prompt.
What an AI profile contains
Profiles are managed under AI → Profiles in Enad Admin. A profile has four settings:
- Name, shown when someone picks a profile.
- Instructions, the description of how the text should be written.
- Include categories and collections, which gives the AI context from the catalogue.
- Default, which marks the profile used when no other is selected.
The instructions are what matter. The other settings decide where the profile shows up and how much the AI gets to know about the product.
Four parts of a good instruction
An instruction works best when it can be followed without interpretation and only contains rules that hold in every context. We split it into four parts.
1. Brand
What do you stand for, who are you writing for, and what should the reader recognise? This is the background that helps the AI choose between two equally correct phrasings.
2. Tone
Describe the voice with concrete traits: direct, warm, technical, understated, premium. "Professional yet personal" says almost nothing, because every brand wants to be that. What helps are traits that rule something out, such as "no exclamation marks" or "short sentences, one point per sentence".
3. Terminology
This is where your glossary belongs. Which product terms do you prefer, how do you write units and capitals, and which words must never be translated? Collection names, material names like GORE-TEX and your own product names are common examples. This is also where most machine translation errors happen, so it pays to be thorough.
4. Rules
Rules are what always applies, whatever the product or language: claims you may not make, information that must always be included and how the text should be formatted.
Here is what a profile might look like for a clothing brand:
Brand:
We write for customers who sail and spend time outdoors, and who choose
clothes that last for many seasons.
Tone:
Factual and calm. Short sentences. No exclamation marks and no exaggerated
promises. Describe what the garment does rather than how it makes the
customer feel.
Terminology:
Use "jacket", never "coat". Collection names stay in English. GORE-TEX is
written in capitals and never translated. Use metric measurements.
Rules:
Never invent materials, certifications, features or care instructions that
are not in the product data. Always mention material and fit when available.
Notice what's missing. Nothing about a particular product, a particular campaign or a particular language. That belongs to the task.
Profile and task are two different things
When you generate text in Enad, you can add a guideline for that one occasion on top of the profile. In an AI job it's called job guidance.
The split is deliberate. The profile holds what applies every time. The guideline describes what you want right now, such as "highlight that this is new in the autumn collection" or "keep the description under 300 characters for this category".
If temporary requests end up in the profile, it soon fills with rules that contradict each other. If the brand's core rules only live in the guideline, someone has to remember to write them every time.
Start from your brand guide
Most brands already have a brand guide, often as a PDF. In Enad you can upload it and let the AI extract a summary, key facts, style guidelines and constraints. This is called Distill PDF.
The result is a draft. You read it through, remove what doesn't apply to product copy and add what's missing before it becomes the profile's instructions. Brand guides often contain rules for logos, colours and campaigns that have nothing to do with product data, and sometimes rules that are out of date. An extract that goes straight into the profile without review takes all of that with it.
One profile per voice
It's tempting to put everything into one big profile. It works worse than it sounds. A profile that has to cover technical specifications, inspiring campaign copy and translations into ten languages becomes long, vague and full of exceptions.
Give each profile one clear job instead, and a name that says which:
- Main brand voice for product descriptions on the webshop
- Technical product copy for specifications and the B2B catalogue
- Swedish translations with the words and phrases that apply in the Swedish market
Mark the profile you use most as the default. It is used automatically for product descriptions and SEO fields when nobody has chosen another.
A few things to keep in mind once the profiles exist:
- Test a changed profile on a few representative products before using it in a large job.
- Update the profile when terminology or guidelines change, the same way you update a glossary.
- Remove profiles you no longer use, so nobody picks an old voice by mistake.
The same voice in every language
Translation is where a profile makes the biggest difference. A translation without a profile translates the words. With a profile, it also keeps the words that shouldn't be translated, the units and the tone.
In Enad you choose a profile when you translate a product, either one language at a time or with suggestions for every language you have set up. The same profile can be used in an AI job that translates whole categories.
One thing is worth being honest about: a profile doesn't guarantee the same quality in every language. Models are better at some languages than others, and a tone that works in English can turn stiff in Finnish. That's why each language is reviewed on its own before it is approved. When you see the same kind of error coming back in one language, it's often a sign that the terminology needs more detail, or that the language needs a profile of its own.
Three-layer GORE-TEX Pro jacket for high-speed sailing. Wind- and waterproof, with high breathability.
Jacka i 3-lagers GORE-TEX Pro för höghastighetssegling. Vind- och vattentät med hög andningsförmåga.
Jacke aus 3-Lagen GORE-TEX Pro für Hochgeschwindigkeitssegeln. Wind- und wasserdicht, sehr atmungsaktiv.
Kolmikerroksinen GORE-TEX Pro -takki nopeaan purjehdukseen. Tuulen- ja vedenpitävä.
Context from the catalogue
Copy gets better when the AI knows more about the product than what's in a single field. That's why a profile in Enad can include the product's categories and collections as context. A jacket in the "Sailing" category and the "Offshore" collection is described differently from the same jacket in "Everyday".
The AI in Enad sits in the same platform as the product data and the asset library. It can therefore also analyse the product's images, such as colour, material, shape and details, and use that when it writes. The copy then matches how the product actually looks, and the alt texts describe what's in the image instead of repeating the product name.
From one field to the whole catalogue
The profile is used in three places in Enad, depending on how much needs writing:
- AI Assist for a single field, such as a product description or an SEO field. Useful when you work on one product at a time.
- Translations for one product into one or all languages.
- AI jobs for many products at once. You choose the type of job (translation, content generation or alt text), which products, fields and languages it covers, and which profile and AI integration to use.
Because the profile is the same in all three, copy written by hand with AI Assist sounds like copy from an AI job across 2,000 products.
The model matters too
A profile shapes how a model writes, but it can't make a weak model strong. A model that's too limited for the task produces weak, incorrect or inconsistent copy, however good the profile is. Long product descriptions, structured SEO fields, translations and image work can need different capabilities.
AI is included in Enad and works straight away, with no contracts or keys of your own. If you'd rather use a provider you already have, you connect your own key for OpenAI, Anthropic or Gemini. You can have several AI integrations active at once and choose which one to use for a given task.
What we recommend before going live:
- Test on representative products, languages and tricky cases.
- Compare against the quality and tone you expect.
- Review extra carefully after changing model or provider.
AI suggests, you publish
Everything the AI produces in Enad is a suggestion. Copy, translations and alt texts are added to the product and reviewed before they are approved and published. When the source text changes, you can see which languages need updating and rerun only those.
That might sound like a brake, but review is what lets you use AI across the whole catalogue with confidence. Nobody wants to publish 2,000 product descriptions that nobody has read. Review is also where you learn what the profile needs: any correction you make more than once is a rule that belongs in the profile.
Common mistakes
Vague instructions. "Write well" and "sound like our brand" give the AI nothing to work with. Write traits and rules that can be followed.
Everything in one profile. A profile meant for every task fits none of them particularly well. Split by voice or use.
The campaign in the profile. Temporary requests belong in the guideline for the task. Otherwise the autumn campaign is still in every description come spring.
No terminology. Most embarrassing translation errors involve words that should never have been translated. List them.
No ban on making things up. Without an explicit rule, an AI can fill in materials, certifications and care instructions that sound plausible. Say that it may only use what's in the product data.
Bulk jobs without a test. Run the profile on ten products first. Changing a profile is cheaper than reviewing a thousand descriptions again.
Checklist for your first profile
- Gather your brand guide, glossary and three product descriptions you're happy with.
- Let Distill PDF draft from the brand guide, and remove what doesn't apply to product copy.
- Write the instruction in four parts: brand, tone, terminology and rules.
- Add the rule about not inventing facts missing from the product data.
- Include categories and collections if they say something about how the product should be described.
- Test on ten products and a couple of languages. Compare with the descriptions you were happy with.
- Adjust, mark the profile as default and run your first AI job.