Brand voice has to be written as rules
Most tone-of-voice guides are adjectives: warm, confident, clear. A model cannot follow an adjective. It can follow rules: sentence length, reading age, whether the brand says we and you, contractions, punctuation habits such as the serial comma or exclamation marks, banned words, and formality per market, du or Sie in German, tu or vous in French. Pair the rules with examples of on-brand and off-brand copy for the same message. Then test before rollout: run identical briefs through each shortlisted tool and have the brand team rank the results blind.
Approved claims, not plausible ones
Language models produce plausible claims: clinically proven, carbon neutral, the fastest in its class. Plausible is not approved. In the EU, food and supplement brands may use only health claims authorised under Regulation (EC) No 1924/2006, and from 27 September 2026 Directive (EU) 2024/825 bans generic environmental claims such as eco-friendly unless recognised excellent environmental performance can be shown. In the UK, the Digital Markets, Competition and Consumers Act 2024 has let the CMA fine companies directly for unfair consumer practices since April 2025. A usable tool draws claims only from a library holding the approved wording, its substantiation and the markets where it may run, and blocks the rest at generation.
Terminology and multilingual output
Product names need exact casing and trademark symbols; some feature names stay in English in certain markets; legal entity names never change. Keep these in a terminology list per language, with do-not-translate entries, so the tool cannot improvise. For multilingual output, decide per market whether copy is generated natively from the brief or adapted from a master, and check the details machines get wrong: number, date and currency formats, quotation marks, and right-to-left layout for Arabic and Hebrew. Synthetic White localises copy into 30 languages with terminology and approved claims enforced, and the banned words in its brand hub apply in every one of them.
What happens to your inputs
Copy briefs are among the most sensitive files in marketing: unreleased products, pricing, positioning against named competitors, sometimes customer data for personalisation. Ask whether inputs are used for training, how long prompts and outputs are retained and why, whether vendor staff can read them, and which model providers process them and where. Tools that draw on internal knowledge, such as shared drives or intranets, raise a further question: whether they respect the file permissions people already have, or let a junior marketer's prompt surface a board paper. Get the answers in the contract, not the help centre.
Updated 25 September 2026