An AI assistant within the CMS creates verifiable drafts directly in the content management system, but assumes no editorial responsibility. Translations, meta texts, alt texts, and structured fields are only published after expert review and human approval. For me, this separation is crucial: The AI prepares the work and points out potential errors, but it doesn't change your website without your knowledge.
Especially for South Tyrolean SMEs, AI-powered editorial assistance can reduce the maintenance effort for multilingual content and frequently changing company data. This requires clearly defined roles, reliable data sources, and a binding approval process.
What AI assistance in a CMS actually does
AI assistance within the CMS operates within predefined boundaries. The system receives selected content and fields, generates a proposal, and saves the result as a draft. The status remains "draft" until an authorized person has completed the review and explicitly approved publication.
The basic process always consists of four steps:
- Editorial input: You provide a source text, reliable company data, and the intended use.
- AI design: The AI formulates a proposal within the defined fields and rules.
- Technical review: A responsible person controls language, facts, context, brand voice, and relevant guidelines.
- Documented release: An authorized person confirms the version. The change log records the draft, corrections, time, and approval.
This logic belongs in an editorial system with clear approval stages . This ensures that AI remains a tool within a transparent workflow.
One example farm for all four use cases
As a simplified example, consider a family-run carpentry workshop in the Eisack Valley with a website in German, Italian, and English. The workshop offers custom-made furniture, provides consultations in a workshop in Klausen, and employs a small team.
This example does not describe a real client. It shows what information the AI receives, what design it prepares, and at which points humans must review and decide.
1. CMS translation with AI: Design and professional approval
AI-powered CMS translation only works reliably if the source text, target language, terminology, and brand voice are clearly defined. Linguistic accuracy alone is not enough. For German, Italian, and English websites, local language customs, technical terms, and the actual meaning of the offering must also be accurate.
Step 1: Editorial input
German source text: "We manufacture custom-made furniture from local wood and advise you personally in our workshop in Klausen."
In addition, the AI receives binding guidelines: "Custom-made furniture" is translated into Italian as "mobili su misura." "Workshop" refers to the artisanal production and consultation location. The tone should remain personal and factual.
Step 2: AI Design
Italian proposal: “Realizziamo mobili su misura in legno regionale e ti offriamo una consulenza personale nel nostro laboratorio a Chiusa.”
The proposal is understandable, but not yet approved. The term "legno regionale" may be too imprecise depending on the proof of origin. Similarly, "laboratorio" must correspond to the actual name of the business.
Step 3: Technical review
A person with linguistic expertise checks the grammar, terminology, and regional usage. The person responsible for the project also verifies whether the wood demonstrably originates from the region and whether the personal consultation actually takes place in the workshop.
Verified version: “Realizziamo mobili su misura con legno proveniente dalla regione e ti accogliamo su appuntamento nella nostra falegnameria a Chiusa.”
Step 4: Documented Release
The translation initially receives the status "linguistically reviewed." After verification of the source text and the consultation context, it receives the status "professionally reviewed." Only final human approval makes the content publishable.
In my work with multilingual businesses, a standardized terminology list has proven invaluable. It prevents services, materials, or place names from being translated differently on various subpages. You can find further information in the article about German-Italian website translations.
2. Metadata with AI: Preparing meta texts in a controlled manner
Creating metadata with AI means preparing meta titles and meta descriptions as editorial suggestions. AI can formulate variations, but it cannot guarantee a ranking or reliably assess search intent on its own. Each draft must fit the page content, be unambiguous, and differ from the meta texts of other pages.
Step 1: Editorial input
The editor selects the "Custom Kitchens" page. The following information is verified: individual planning, production in Klausen, consultation by appointment, and delivery within South Tyrol. Statements such as "cheapest custom kitchens," "100 percent regional," or "delivery in two weeks" are unsubstantiated.
Step 2: AI Design
Suggested meta title: "Custom kitchens from Klausen | Individually planned"
Suggested meta description: "Individually planned custom kitchens from our carpentry workshop in Klausen. Personal consultation and delivery throughout South Tyrol."
Step 3: Technical review
The editorial team checks four points:
- Does the suggestion match the search intent?
- Does the website actually cover all the services mentioned?
- Is the wording used for display in search results precise and understandable?
- Does each language version have its own independent, matching text?
The AI must not add benefits that are not substantiated on the page. Similarly, a proliferation of location and performance terms does not improve an unclear meta text.
Step 4: Documented Release
The reviewed meta text is saved along with the page version, language, and approver. If the service or delivery area changes later, the weekly content check can flag any potential discrepancies. Corrections are made exclusively by the editorial team.
3. Alt text: Image function takes precedence over search terms
Alt text is not a place to dump keywords. Meaningful alt text conveys the information or function that an image fulfills in the specific context of the page. The same photo can be treated differently on a team page, a services page, or as a purely decorative element.
The W3C Web Accessibility Initiative explains that text alternatives depend on the use, context, and content of an image. Purely decorative images are generally given an empty alt attribute (alt="") so that assistive technologies do not display unnecessary content.
Step 1: Editorial input
The CMS contains an image with the rather uninformative filename "IMG_2847.jpg". The editorial team adds the context: The photo shows a carpenter planing a larch wood drawer front. The image appears on the services page for handcrafted custom furniture and illustrates a specific step in the process.
Step 2: AI Design
Suggestion: "A carpenter is working on a piece of furniture in a workshop in South Tyrol."
The design is too general. The image shows a specific work step, while the location reference is not necessary for the image information.
Step 3: Technical review
The editorial team checks what is actually recognizable in the image and what information is still missing in the surrounding text.
Verified version: "Carpenter planes a drawer front made of larch."
If the same photo is used merely as a decorative background and conveys no additional information, the alt text remains empty. The AI can prepare this decision, but requires information about the image's function and the page context.
Step 4: Documented Release
The CMS saves the approved alt text each time an image is used. For multilingual pages, it checks whether the alt text needs to be translated or whether the image serves a different function in a particular language version.
4. Structured fields: Pre-fill with facts, don't invent them.
Structured fields separate individual pieces of information. Services, opening hours, contact persons, locations, FAQs, contact methods, and validity periods are not hidden in a long text, but rather maintained in defined data fields. This facilitates controlled output across multiple pages and reduces inconsistencies.
The conceptual basis is explained in our article about a website content model with structured fields.
Step 1: Editorial input
The carpentry workshop transmits the following information from its approved company schedule: "Workshop open Monday to Friday from 8 a.m. to 12 p.m. and from 14 p.m. to 17 p.m. Consultations by appointment only." The source is the internally approved company schedule with the date.
Step 2: AI Design
The AI distributes the information across the designated fields:
- Days of the week: Monday to Friday
- Late morning: 08: 00 12 to: 00 pm
- Afternoon: 14: 00 17 to: 00 pm
- Consultation: by appointment
- Saturdays and Sundays: not specified
The AI must not automatically interpret "closed" as "no information" being given. Missing information remains marked as unresolved.
Step 3: Technical review
The responsible person checks mandatory fields, data types, time format, holiday regulations, and source. For contact persons, it is also checked whether the name, function, telephone number, and intended publication are current and permissible.
Step 4: Documented Release
The approved fields are saved with source, verification date, and responsible person. If opening hours change, the same confirmed information can be distributed to multiple locations in a controlled manner. The database remains under human control.
The weekly content check is a reporting system.
Websites are often rendered unreliable by small inconsistencies: the footer contains an old phone number, the contact page has different opening hours, or an Italian service page lists a product that is no longer offered.
A weekly content check can collect such anomalies. Each report should include:
- the exact source and language version
- the detected deviation or missing information,
- the comparison source used,
- an understandable explanation,
- a non-binding proposal for action,
- the status "open", "reviewed", "rejected" or "approved".
The content check must not make any silent changes. It reports and prioritizes potential problems. It is explicitly prohibited to automatically publish or correct business-critical information.
Roles and rights for small editorial teams
Technical features alone don't create a reliable process. You need a role-based logic that also works in a small team. These organizational rules are summarized as AI governance.
- Trigger draft: Editors are allowed to submit defined content to the AI and request suggestions.
- Check language: A person with linguistic expertise checks translations, tone of voice, and local phrasing.
- Check the facts: The person responsible for the company confirms services, prices, times, names and other company details.
- Final release: A person authorized to do so deliberately publishes the reviewed version.
- View the logs: Those in charge can track who created, modified, rejected, or approved a draft.
For sensitive or business-critical content, I recommend the four-eyes principle. The person working on an AI design should not also be the sole expert reviewer. This reduces AI errors as well as typos, outdated assumptions, and misunderstandings.
The change log should still show the original version, the AI suggestion, manual changes, the person who reviewed it, the person who approved it, and the time. It is crucial that the published decision remains traceable.
Own AI access and data sovereignty
The company should use its own AI access point. This allows the company to retain control over provider selection, usage scope, billing, and access management. If a provider is changed, the connection can be transferred, provided the technical integration was planned accordingly.
However, having your own AI access does not guarantee compliance with data protection regulations. In 2024, the European Data Protection Board clarified that the processing of personal data in connection with AI models is subject to the requirements of the GDPR and requires a valid legal basis.
Before connecting, at least the following questions must be answered:
- What content and data are transferred to the AI service?
- Does the information you provide contain personal data, confidential information, or special categories of personal data?
- For what specific purpose is the processing carried out?
- On what legal basis is the processing based?
- Where is data processed and how long is it stored?
- Who is allowed to use the access and who manages the permissions?
- Is user input used for the further development of the service, and can this usage be deactivated?
- How will the editorial team continue working if the AI service fails or is replaced?
Data sovereignty doesn't mean you can't use external services. Data sovereignty means your company understands and controls data flows, roles, export options, and dependencies. You can find a more detailed decision logic in our article on GDPR and AI for SMEs.
EU AI Act: Correctly classifying competence and oversight
The EU AI Act does not treat all AI applications equally. The transparency and supervisory requirements for high-risk AI arise in particular from Articles 13 and 14. Article 50 contains specific transparency obligations for certain AI systems. In contrast, the obligation to promote sufficient AI competence under Article 4 generally applies to providers and operators and takes into account, among other things, knowledge, experience, and the context of application.
Editorial AI assistance within a content management system (CMS) should not be classified as a high-risk system solely due to its editorial function. The legal assessment depends on the specific function, the data processed, and the context of use. Regardless of the classification, trained personnel, defined boundaries, and human oversight are necessary for a reliable editorial process.
Questions and answers about AI assistance in the CMS
Does the AI automatically publish changes?
No. A controlled process does not include automatic publication . The AI submits proposals as drafts; publication only occurs after expert review and human approval.
Why should my company use its own AI access?
Having your own AI access point gives your company more control over roles, costs, provider selection, and technical connections. However, this access does not automatically guarantee data protection compliance; data flows, legal bases, and contractual terms must be reviewed separately.
Am I allowed to enter sensitive data into the AI assistant?
Personal, confidential, or specially protected data should not be transferred to an external service without prior review. Define clearly which data is permitted, which must be anonymized, and which must be excluded.
Who is responsible for incorrect published content?
The AI assumes no editorial or entrepreneurial responsibility. Your company must regulate responsibilities and approvals in such a way that published statements are professionally reviewed. Specific liability issues should be assessed legally.
Does every AI translation need to be reviewed by a human?
Published company content should be reviewed by a human. This is especially important for technical, legal, or brand-related texts, because a translation may sound fluent but still inaccurately reflect terminology or the scope of services.
Can AI generate alt text completely on its own?
The AI can suggest alt text if image content and page context are provided. The editorial team must then determine the image's function and whether descriptive, functional, or empty alt text is required.
What should be included in the change log?
The changelog should include the original version, the AI draft, significant corrections, review status, approver, and time. This ensures that it remains traceable how a proposal became published content.
What happens if the AI service fails?
The CMS and manual editing must remain functional. Existing content must not depend on ongoing AI access. Only the generation of new AI suggestions is paused until the service is available again or replaced.
Does a small team need a four-eyes principle?
Not every minor text correction requires two people. However, when it comes to opening hours, service promises, prices, personal data, and technically sensitive translations, the four-eyes principle reduces avoidable errors.
Can AI automatically fill in missing company data?
No. Missing facts must be marked as unresolved. The AI may only pre-populate structured fields from a reliable source and may not generate opening hours, contact persons, prices, or services based on assumptions.
Here's how to introduce AI assistance in a controlled manner in your CMS.
After over 20 years in web development, branding, and digital projects, one principle has become clear to me: A narrowly defined and easily understood process is more reliable than extensive automation without clear responsibilities. Therefore, don't start with all content at once.
- Choose a use case: For example, start with only meta-texts or a clearly defined type of translation.
- Limit fields: Define which CMS fields can be read and which can be filled as drafts.
- Identify reliable sources: Define where approved services, opening hours, terminology, and brand rules are located.
- Document the test criteria: Define how language, facts, image context, uniqueness, and timeliness will be controlled.
- Assigning roles and rights: Separate the draft, linguistic review, technical review, and final approval stages.
- Use test content: First, test the process using non-critical and non-personal sample data.
- Test the change log: Check whether decisions and corrections are comprehensible.
- Plan for outages and provider changes: Normal content maintenance must remain possible even without an AI service.
- Expand gradually: Only add additional languages, page types, or structured fields after a successful test.
If you want to integrate AI assistance into your website's CMS, our AI and digitalization services combine technical implementation with editorial processes, brand voice, and clear responsibilities. The goal is less routine work, more reliable data, and a system that your team controls.
