Using AI in technical documentation? 

AI is part­ic­ul­arly worthwhile for recurring tasks, large amounts of information and processes in which content needs to be searched, evaluated, created, translated, checked or provided for different target groups. Possible fields of application include, for example, research in existing docum­ent­ation, creating initial text drafts, quality checks, translat­ions or answering product-related questions.

Whether this actually results in an advantage depends on the specific process, the available information and the quality requir­em­ents. Therefore, the starting point should not be what is technically feasible, but a clearly demarcated problem. A pilot project can be used to test whether the application measurably saves time, improves quality or mean­ingfully relieves employees.

AI can support technical writers with many time-consuming tasks. These include, for example:

• searching and evaluating existing documents and product information
• comparing content and identifying missing or contrad­ictory information
• creating text drafts based on information provided
• stand­ard­ising, rephrasing or preparing content for different target groups
• checking terminology and editorial rules
• preparing or autom­at­ic­ally generating translat­ions
• providing information for different media and usage situations

Which tasks are actually suitable depends on the quality of the source information, the process and the requir­em­ents for the result. Prof­ess­ional reviews, approvals and clear resp­ons­ib­il­ities therefore remain necessary. AI supports editorial work, but does not assume resp­ons­ib­ility for the correctness and complet­eness of the technical docum­ent­ation.

For every AI use case, information does not need to be fully structured or stored in a content management system. Existing documents, files and other unstruct­ured information sources can also be used, depending on the use case, and evaluated with the help of AI.

Structured content can offer clear advantages. It facilitates reuse, automated processing, updating and provision of information. The crucial question is therefore not in general whether structured data is available, but which information quality and structure the specific use case requires.

We examine which existing information can already be used, which functions an existing content management system offers and where additional structure, interfaces or other technical solutions are required. This prevents you from carrying out unnecessary preparatory work before your first use of AI.

A good use case solves a specific problem and can be tested with limited effort. Part­ic­ul­arly suitable are recurring tasks that cause a lot of manual work and where time spent, quality and costs can be compared before and after the use of AI.

An example from our own practice is the conversion of product photos into line drawings. Manually creating such illustrat­ions can be time-consuming. That is why we have developed an AI service that autom­at­ic­ally converts photos into line drawings. Specialists review and edit the result before it is used in technical docum­ent­ation.

Not every application is equally suitable. If only a few processes are involved, the source information fluctuates greatly or the results are difficult to verify, developing your own solution can be more expensive than the potential savings.

At the beginning, we therefore clarify:

• Which task regularly involves a lot of effort?

• How often does it occur?

• What information is available for this?

• How can quality and benefit be measured?

• Which specialist checks remain necessary?

In technical writing, AI can analyse existing information, create draft texts, compare content and check editorial rules. However, the results may be technically incorrect, incomplete or not suited to the product. A ling­uist­ic­ally convincing text is not autom­at­ic­ally a correct operating manual or service information.

How reliable the results are depends in particular on:

• whether up-to-date and unambiguous product information is available

• whether target group, intended use and type of information are clearly defined

• whether terminology, writing rules and docum­ent­ation spec­if­ic­at­ions are taken into account

• whether the AI accesses suitable and approved information sources

• how results are checked from a technical and editorial perspective

Especially for safety-relevant or legally significant content, AI output must not be adopted without review. AI can accelerate research, creation and quality assurance. However, prof­ess­ional review and approval by qualified persons remain indisp­ens­able.

Sensitive information is frequently processed in technical docum­ent­ation. This includes design data, product photos, internal spec­if­ic­at­ions, service information or content relating to products that have not yet been released. We do not enter such data unchecked into publicly accessible AI tools.

For our work at kothes we use a centrally managed AI environment. This allows us to define which applic­at­ions and models may be used, who can access them and which rules apply to handling customer information. The technology is thus embedded in defined organ­is­at­ional and technical processes.

Before a specific use, we clarify, among other things:

• which information may be processed

• which model and which application are suitable

• where and how the data is processed

• whether inputs are used for training a model

• who receives access to data and results

• which storage and deletion rules apply

• which customer requir­em­ents addit­ion­ally need to be taken into account

Depending on the use case, information can be reduced, anonymised or processed in a separate protected environment. For part­ic­ul­arly sensitive projects, we coordinate the permissible processing and the technical solution used with the customer in advance.

With suitable workflows, prod­uct­ivity increases of 20 to 40 percent are realistic. In individual use cases, the potential can be sign­if­ic­antly higher. How large the savings actually are depends on the task, the available information and the required verif­ic­ation effort.

One example is the creation of line drawings from product photos. Depending on complexity, manual tracing can take around 15 to 60 minutes per illustration. With AI, an initial draft can often be generated within a few minutes. Including review and reworking, the effort can then be around one to five minutes.

In this way, we look at the entire docum­ent­ation process step by step. We examine which tasks can be accelerated, what effort is required for checking and reworking, and how the output changes. Some projects show that for certain tasks the output can more than double.

However, this figure cannot autom­at­ic­ally be transferred to the entire docum­ent­ation process. Technical coord­in­ation, reviews and approvals remain necessary. In some cases, new process steps are also created to check the results. Whether AI ultimately reduces costs, shortens processing times or enables more output with the same resources should therefore be measured using a specific use case.

That mainly depends on whether you want to use AI yourself or whether kothes takes over the editorial work including the use of AI.

If kothes provides the service, you pay for the agreed result. Where AI reduces the effort, processing can be cheaper or faster than with purely manual production. Altern­at­ively, you receive more output for the same budget. How great this advantage is depends on the task, the source material and the effort required for checking and post-editing.

If you want to use AI in your own technical writing department, tool costs and labour costs will arise. Tool costs include existing AI functions in the content management system, additional applic­at­ions such as Microsoft Copilot or specialised cloud solutions, as well as usage-based costs, for example for processed data or tokens.

There is also internal or external effort involved in implem­ent­ation. The application must be selected, set up and sensibly integrated into day-to-day editorial work. In addition, clear rules are needed for data protection, quality assurance, checking and approval. Only when the solution is used regularly and demonstrably saves time or increases output does an economic benefit arise.