With a growing appetite for faster content turnaround times, it is no wonder TechDoc teams are looking to AI to speed up localisation.
And since technical documentation is subject to such strict quality and compliance standards, some ask: Why not build our own AI model for localisation, to keep everything in-house, and have full control of the output?
To explore the question, we spoke with Signe Winther Poulsen, COO at LanguageWire. With more than 20 years of experience in localisation, Signe has helped global organisations manage complex technical documentation ecosystems.
Signe's observation is that it's easy to get started with your own AI model, but much harder to maintain it. In this article, she explains why that is and what you can do instead.
Why scaling an AI model is harder than building one
Organisations that work with technical documentation are often mature in their content management approach. They tend to work with structured content environments built around XML, DITA, or component-based authoring.
In other words, they are used to complexity and scale. Many are very successful in experimenting and improving their content creation processes with AI.
However, in Signe Winther Poulsen's experience, using AI for content creation and using AI as part of a larger localisation workflow are two different beasts:
"I would be mindful of the cost over time. It can seem easy and tempting, but what we're seeing and observing, and the feedback that we're getting from customers, is that the return on investment is hard to keep up once it starts to become daily business."
When you build your own AI model, you have to maintain all the infrastructure, governance, quality controls, and continuous optimisation needed to make it safe and reliable for TechDoc localisation.
Over time, changes in technology or environment may cause the AI to make mistakes, even if it performs well at first: "What we see in practice is that the AI translations need continuous training, continuous modification, and new AI agentic flows to be set up in order to avoid hallucinations."
As content volumes increase, the model must continue to perform consistently across products, markets, and languages. It must preserve approved terminology, avoid hallucinations, and comply with technical and regulatory requirements, every single time, no matter the content type or the language.
This is particularly important in a niche like TechDoc where quality is often defined by accuracy.
A minor inconsistency in terminology or a poorly translated instruction can create confusion, operational risks, or compliance breaches.
But accuracy and precision can be particularly finicky once AI is involved.
"What we're observing is that it's quite easy to start a project like this, but scaling it and maintaining the quality is harder, especially when precision is a key component of quality," Signe explains.
To achieve proper quality control, every step needs to be carefully managed. You can improve precision by training the model, but also by combining it with technologies that are much better at being precise, such as termbases and translation memories.
AI is just a fraction of the localisation process. Here’s what else matters
One of the biggest misconceptions Signe sees is that translation is the main part of the localisation process, and that training AI to translate well is the biggest challenge teams will face.
In reality, translation is only one step in a complex workflow.
A typical TechDoc localisation workflow often includes several important steps:
Content creation
File transformation and processing
Translation memory matching
Terminology application
AI-assisted translation
Human editing
Proofreading and/or validation
Final delivery and reintegration into the documentation environment
Each stage plays a role in maintaining quality, consistency, and technical accuracy. When organisations ask whether they should build their own AI model, they often focus on only one piece of a much larger localisation ecosystem.
Signe points out that it's the holistic workflow that makes the biggest difference for TechDoc. AI only creates value when integrated into that setup, working alongside the other tools rather than sitting "on top" of them:
"I think we should look at technology as a package that's been evolving over time. It's not that once there was one technology and now there's a new technology that's AI. We would always recommend AI translations as a core, just like we would always recommend the translation memory as the first step because that's maybe the most precise."
Konecranes: What good AI-use looks like in practice
Viewing AI through that lens can also help demystify some of the fear TechDoc teams have around AI.
Signe says: "The risk mitigation for fear of AI translation going rogue is a level of human in the loop plus technology, such as language quality assurance. You can utilise the benefits of AI without having to close your eyes and say, 'I hope it goes okay'."
An example is Konecranes, one of LanguageWire's customers and a global industrial manufacturer that creates custom cranes. For many years, the company had a decentralised and manual setup, where technical content was created from scratch across different business units.
With the help of LanguageWire, the company decided to centralise its TechDoc communications to reduce costs. Rather than having each department maintain its own terminology, workflows, and translation setup, they created a single, shared localisation ecosystem.
"We've been working in tandem with them on optimising our mutual processes consistently. We've been part of that process, both in terms of the technical setup, and workflow construction," Signe explains.
Since they had already established a localisation workflow that was designed to scale, AI could simply become another component within this system.
According to Manager of Documentation at Konecranes Pasi Savola, human-in-the-loop may be the most important component of their AI-localisation workflow. In a recent Slator article, he highlights how the complexity of his business means that localisation must follow that complexity, both in terms of linguistic assets, processes, and AI, saying that AI “cannot create content on the go for a guy who’s going to service a crane.” Customer-facing content must be verified by humans.
“We cannot leave everything to AI, of course, especially in our business… We need the human touch to make sure that the content is right.” Pasi Savola, Manager of Documentation, Konecranes.
That is why it is so important to not treat AI as an either-or but rather focusing on building a controlled workflow. AI can then fit into that workflow in a way that it can add value without increasing the risk.
In Signe’s words:
“It’s important that we look at the AI space as not an either or, but instead as: How do we supplement the value-add the technology can create with human-in-the-loop, but also other non-AI technologies that are a bit more rigid and stringent like terminology management.”
Should you build your own AI model?
Before investing in your own model, ask yourself: is your biggest localisation challenge translation quality? Or is it terminology governance, workflow integration, review capacity, and scaling across languages?
How will you prevent hallucinations and maintain the model over time? And do you have the resources to continuously train and optimise it?
Signe's advice is to make sure you can answer those questions before starting to build your own model. However, in her experience, the best option is to bring your existing content maturity and experience to a trusted partner like LanguageWire for the best results:
"Customers have a lot of opportunities to improve their content creation processes with AI. Let's partner on making those two ends meet: the optimised content creation with the optimised translation process we have. Then we can manoeuvre together in this great, exciting, big, and sometimes a little scary ocean of AI."
Finding the right way to implement AI
For most organisations, the challenge is not launching an AI model. It is building a localisation operation that can scale, maintain quality, and support business-critical content across markets and languages.
At LanguageWire, we have spent more than 25 years helping enterprises manage multilingual content, terminology, localisation workflows, and quality assurance at scale. Our approach combines AI, automation, linguistic expertise, and proven localisation processes to help organisations get the benefits of AI without compromising control or quality.
If you're evaluating AI localisation for technical documentation, speak with one of our experts. We can help you determine the right localisation strategy for your TechDoc team. Consult an expert.
This article is the August 2026 edition of Lost & found in translation, a monthly newsletter sharing insights, opinions, and reflections by real people who work in localisation. Subscribe here to get notified when the next edition comes out.
