Case study
How a set of rules made ten domains reviewable, and why our first classification was too strict.
A listed manufacturer of fuel cells for off-grid power supply, around 400 employees, subsidiaries on three continents. Its environmental claims appear on ten domains in up to four languages, and under EU Directive 2024/825 each of them needs evidence. Marketing had neither the time nor the people to read several thousand pages and data sheets by hand.
Entry 1
Starting point
The company runs a set of websites that has grown over the years: websites for product lines, markets and subsidiaries, a shop, a service portal. Error codes, help texts and app content exist in eleven languages. Behind them run Salesforce and Pardot for prospects and returns, SAP SuccessFactors for job advertisements and Freshdesk for service. We have looked after these websites and systems for years.
EU Directive 2024/825, EmpCo for short, allows environmental claims such as "zero emissions", "climate neutral" or "environmentally friendly" only if there is evidence behind them. Member states had to transpose it by March 2026, and it applies from September 2026. For a clean energy manufacturer, this is no footnote: such claims sit in product texts, data sheets, press releases and shop descriptions, often for years and worded slightly differently in every language. Whether an individual wording is permissible is ultimately for the client's legal advisers to assess; this case study does not replace them.
The task fell to marketing, five people who also handle trade fairs, campaigns and product launches. By hand, someone would have had to open every page, find every claim, classify it, document it and reword it, then do the same in three more languages. Nobody had three months to spare, and nobody wanted to leave the content unchecked either.
Entry 2
Figures before
| Domains | 10 |
|---|---|
| Languages per website | up to 4 |
| Languages of service content | 11 |
| Text inventory | several thousand pages and data sheets |
| Manual review effort | estimated at around three to four person-months |
| Capacity | marketing team of five, with no free time for the review |
| Overview | none: nobody knew how many passages were affected |
| Risk | cease-and-desist letters over unsubstantiated environmental claims from 2026 |
Entry 3
The path
About ten weeks to regular operation. The week figures are rounded.
- Week 1
Reading in the entire text inventory
A workflow reads all ten domains in all languages, plus the data sheets as PDFs and the press releases. Every page gets an entry with URL, language and content type. For the first time, there is a number: this many pages, this many documents.
- Weeks 2 and 3
First classification
The workflow marks every passage that might fall under the directive, with location, language and classification: uncritical, evidence needed, reword. Marketing checks a sample and says where the workflow is off.
- Weeks 3 to 4
Editorial guidelines as the set of rules
Together with marketing, we draw up guidelines: which claim is permitted with which evidence, how others are reworded, which terms are off limits. The same guidelines steer the workflow and apply to every future text.
- Weeks 4 to 6
A tool for the action needed
Instead of a spreadsheet sent by email, marketing gets its own interface: how many affected passages each page has, which are done, what is still open. The status updates as the work progresses.
- Weeks 6 to 9
Trial run on one domain
For every marked passage, the workflow proposes a rewording according to the guidelines. A person in marketing checks and approves it, and every change is logged. Only once the proposals on the first domain hardly need correcting do the other nine follow.
- From week 10
Handover to regular operation
Marketing works through the action needed at its own pace. New texts are checked against the same guidelines before going live. We look after ongoing operation and update the rules when products or case law change.
Entry 4
Result
All ten domains have been checked in all languages. The action needed is visible for each page, and rewording follows guidelines the company co-wrote itself. The workflow does the reading, the decision stays with a person: marketing approves every passage before it goes online. Anyone who later asks why a claim is worded the way it is will find the proposal, the approval and the date in the log.
Alongside this, the service content in eleven languages is maintained centrally and translated using the client's technical glossary. Only empty fields are filled; editorially reviewed texts stay untouched.
| Metric | before | after |
|---|---|---|
| Domains checked | none systematically | 10 of 10, all languages |
| Overview | none | action needed per page, always up to date |
| Reading and classifying | around three to four person-months, estimated | done by the workflow |
| Evidence | none | log per passage with approval |
| New texts | no rule | checked against the guidelines |
Operating mode: the texts contain no personal data and run via cloud models under contract. Personal data stays in Salesforce and SuccessFactors. Only the client's marketing team may approve.
Entry 5
What went wrong
- The first classification was too strict. The workflow flagged too many passages that were not environmental promises at all: technical specifications with a measured value in a data sheet, the word "clean" in an assembly manual, quotations from funding programmes. Marketing would have spent days dismissing false alarms. We refined the rules, which now distinguish between an advertising claim and a technical specification. The lesson: before the full run, we calibrate on a sample, together with the people who will approve later.
- The translations needed two rounds. In the technical glossary, one product term had two meanings in one language: the same word stood for an assembly and for an accessory. The first translation round followed the glossary faithfully and therefore said the wrong thing in several places. The lesson: before the first run, a glossary is checked for ambiguities, one meaning per term, with a sentence of context.
Entry 6
What we would do differently today
We would put the guidelines before the classification, not after it. The first classification ran on general rules, and the guidelines came in week three. The overly strict hits were the result. Today, a regulated-text project starts with a short draft of the rules, tested on real passages in the first week.
We would set up the tool for the action needed on day one, even if it is empty at first. In the first weeks, marketing was waiting for a status it could not see. An interface that fills up from the first run answers the question "how far are we?" without a status email.
And we would treat the glossary as a work package of its own. It looks like a side issue, but across eleven languages it decides whether a component has the right name.
Entry 7
Key facts
| Industry | Energy technology, fuel cells for off-grid and mobile power supply, listed |
|---|---|
| Employees | around 400 |
| Systems | WordPress Multisite, Salesforce, Pardot, SAP SuccessFactors, Freshdesk, Postmark, Cloudflare |
| Services | Texts and content at scale, Tools that fit you, system integration, ongoing operation |
| Operating mode | Texts without personal data via cloud models under contract; personal data stays in the client's systems; approval for every passage by marketing |
| Duration | about ten weeks to regular operation, in ongoing operation ever since |
Handover
The first step is a 30-minute call.
You tell us about the workflow that costs you the most time. We tell you honestly whether AI pays off there and what the next step would be. Whether a workflow analysis follows is up to you.