The Translation Workflow for the AI Era

13 Aug 2026

Read time: 11 minute(s)

We are building Content Fusion as a context layer where people and AI agents create together, organizations govern and approve content, and trusted knowledge becomes available for both humans and agents to consume and act on.

The translation support we added delivers on that vision as well as on other values behind Oxygen Content Fusion, such as no lock-in and adaptability to your content.

When designing a translation workflow for AI Agents, there are two constraints from the classic translation workflows that change: translation is fast (minutes or hours rather than days or weeks) and cheap (around one dollar per topic).

So, starting from first principles, we created a workflow for incremental translation that keeps that state inside the project as files with a documented XML-based format.

This open-format state management lets you integrate your existing workflows while gradually adopting the built-in AI Translator in specific scenarios, such as last-minute changes before a release.

How it works

In your project workspace, open the Translation Jobs panel and choose New translation. Pick your publications and your target languages — both are multi-select, so one job can produce several translations. Add instructions for this particular job if you have any, and start it.

The job appears in the list and you can watch the execution log while it runs. When it finishes, Bring into Workspace puts the translated files into your workspace as ordinary uncommitted changes. These include a parallel folder for each language, the translated maps, the binaries referenced by the publication, and the translation index. Then you have to review the translations and commit.

The next translation job will use the translation index to translate only what was changed.

Gradual adoption

The workflow state is stored in a translation index file with a documented format, and the agent knows how to update it even if content is translated elsewhere, for example by a translation agency.

Even if you do not use the built-in agent for the translation, you can ask it to ingest the translated content so that it keeps track of what is already translated and what needs to be reworked. This means you can continue to use the your language service providers for most of the tasks while gradually adopting AI translation.

Quality comes from context

One common concern with AI translation is the quality of the output.

First of all, Content Fusion uses an AI Agent based on frontier LLMs which is a radically different technology than legacy Machine Translation and its performance continues to evolve at an unprecedented pace.

We also found that the quality can be brought to a good level by providing the right context. There are several kinds, and they are worth separating.

The content itself. The agent reads the whole topic before it writes anything along with any previous translation of that topic and related files in the project.
Your preferences. Terminology, house style, do-not-translate lists, the register you address readers in. You commit these once in .contentfusion/i18n/agent-instructions.md and they apply to every job:
```
Use a formal register and address the reader with the formal second person.
Never translate: Content Fusion, oXygen, WebHelp, DITA-OT.
For terms, look them up in ./terms.tbx and reuse the canonical translation there.
```
While it already knows the standard DITA semantics, you can configure how it translates your DITA specialization.
The domain. Although small AI models provide good translation capabilities, we opted for a large model (currently Sonnet 5) because in its many weights it encodes vast knowledge even about niche domains. Of course, your content is innovative and cannot be in the model weights, but it can pick up any specifics from the same agent-instructions.md file.

So, the quality of the translation can be improved by investing some effort to configure the agent. For example, when we translated the Content Fusion Cloud user manual, we noticed that the quality improved after we included the translated UI strings in the agent instructions. We will publish a dedicated blog post about that experiment.

Reviewing the translation

Since AI can make mistakes, you need to review the output before accepting it in your project.

This may be challenging if you do not understand the target language. Even in this case, you can mitigate the risk of conceptual errors or missing information in the translation. Ask the agent to translate the content back into the source language and compare it against the original. Or ask it to review the target-language files and report the problems it finds without translating anything. Both are described in Reviewing Translations Before You Commit.

The workflow is yours to steer

One advantage of having an intelligent agent handling the translation is that it can adapt the workflow or do a whole bunch of related tasks. Here are some capabilities that you can add via custom instructions:

Import a project with existing translations: The agent skips the translation part and adds the files to the index.
Rebuild a conventional workflow if you want one: You can ask the agent to put all the untranslated files in a folder such that you can download and send to a translation agency. When you get the files back it can ingest them and update the translation index.
Ask it questions instead of translations: A job does not have to translate. It can report the translation status of each publication per language, audit terminology across the languages you already have, or build your glossary as it goes.
Discover terminology: For a legacy project, discovering the terms is a chore. You can ask the agent to find them either by markup or by simply reading the content.
Implement introspection: After translating a topic, have a subagent re-read the output against the source and report grammar or fluency problems, and another agent to fix the findings. You can even loop until everything is settled.

Another benefit of this steerability is that, on a first run, the agent does more that translate files, it sets up everything: it creates the target-language folders and maps, copies the binaries, and mirrors the deliverables from the source language and can create multilingual publications.

Try it in five minutes

Try this workflow for free in Content Fusion Cloud. Create an organization and a project from the Hydraulic Gear Pump sample. It arrives already set up with a source folder, per-language folders, a translation index, and a multilingual PDF deliverable. Edit an English topic and commit. Start a translation job for a couple of languages. Watch the card, bring the result into your workspace, read the diff in the Changes view, and build the multilingual PDF. The full walkthrough is in Getting Started with Translation.