AI workflows are built around your own reference material: for every batch of content, they retrieve the matching translation memory entries and terms and hand them to the AI as context, together with your instructions. The settings on this page put you in charge of what counts as good context for your content: how close a match must be to qualify, which approval statuses count, and which segments the AI must leave alone. And if you are unsure, you do not need to decide anything: every setting ships with a default chosen to perform well out of the box, so a workflow with untouched settings is already set up for success.
Where Parameters Appear
|
Workflow |
Where you configure it |
|---|---|
|
AI translation |
Settings > Machine translation systems > Wordbee AI connector |
|
AI post-editing |
Supplier pricelist service, AI Agent tab |
|
AI quality assurance, run from a QA profile |
QA profile, AI Analysis rule |
|
AI quality assurance, run as a supplier service |
Supplier pricelist service, AI Agent tab, the same as post-editing |
One workflow, two configurations. A quality assurance workflow you use in both places is one AI workflow with two independent setups, like the same app installed on your phone and your tablet: settings changed on one device never carry over to the other. The configuration in the QA profile applies only when QA runs from the profile; the configuration in the supplier service applies only when the workflow runs as a supplier service on a job. Setting up one never configures the other, so if you use both, configure both, even when you want the settings identical.
Defaults and Overrides
Each parameter has a checkbox in front of it:
-
Unchecked: the recommended default applies. The field is greyed out and shows the default value, so you always see what will be used. When Wordbee improves a default in a future release, your configuration picks it up automatically.
-
Checked: your own value applies and is kept as you set it, regardless of future default changes.
A parameter that only does something with a concrete value (a number, a selection, a checkbox) cannot be enabled and left empty: the configuration will not save until you provide a value or untick the parameter. Free-text instructions can be enabled and left blank.
Parameter Reference
|
Parameter |
What it does |
Default |
|---|---|---|
|
Custom Instructions |
Free-text guidance added to the AI's built-in instructions. Multi-step workflows show one instructions field per step or agent. |
Empty |
|
TM Minimum Similarity (%) |
Only translation memory matches at or above this similarity are given to the AI as reference. Range 50 to 110. This governs the AI's reference material only; it is unrelated to the pre-translation threshold in your word count profile. |
100 (post-editing: 90, so the AI also sees close fuzzy matches as context) |
|
TM Match Statuses |
Only memory matches with a selected status (Neutral, Confirmed, Error) are used as reference. |
Confirmed |
|
TB Minimum Similarity (%) |
Only term base matches at or above this similarity are given to the AI. Range 50 to 100. |
100 |
|
TB Match Statuses |
Only terms with a selected status are used. |
Neutral, Confirmed |
|
Skip Segments By Origin |
Segments whose latest content came from a selected origin are left untouched by the AI. |
Varies per workflow, see below |
|
Skip Locked Segments |
When enabled, locked segments are not sent to the AI. |
Enabled |
Skip Segments By Origin: defaults per workflow
The available origins are: User-edited, 100% pre-translation, Fuzzy pre-translation, Rule-based pre-translation, MT translation, 100% pretranslation (prev. version), 110% pretranslation (prev. version), Machine correction, and 110% pre-translation.
|
Workflow |
Default |
Why |
|---|---|---|
|
AI translation |
Skips every origin except machine correction |
Translation fills only what has no usable content yet. |
|
AI post-editing |
Skips user-edited, 100% and 110% pre-translations, and previous-version pre-translations |
Reliable content is protected; fuzzy matches and machine translation output are exactly what post-editing is for. |
|
AI quality assurance |
Skips nothing |
Quality checks must evaluate everything, whoever produced it. |
Writing Custom Instructions
Custom instructions are the most flexible setting. They are written in plain language and added to the AI's built-in instructions for that step. Things that work well:
-
Tone of voice: "Use a formal register. Address the reader as 'vous' in French."
-
Brand rules: "Never translate the product name Wordbee. Use 'Connector' instead of 'Plugin'."
-
Locale conventions: "When translating to fr-CA, prefer Canadian variants."
-
Domain rules: "This is medical content; preserve all clinical abbreviations exactly."
-
What to ignore: "Do not flag informal register; it is intentional in this content."
Keep instructions short and specific. They travel to the AI with every batch, alongside the segments and the reference material: a few precise rules are followed reliably, while a long list of rules makes each individual rule less certain to be applied and adds processing time to every batch. A useful test: if your instructions try to cover several kinds of content at once, split them into separately configured services instead (for example one for marketing and one for legal), each with its own short instructions.
In multi-agent workflows, give each agent instructions about its own task: the first agent translates or post-edits, so it takes style and terminology guidance; a validating agent checks, so it takes guidance on what to verify or explicitly ignore.
For quality assurance workflows, custom instructions refine how the existing checks behave: what to treat as intentional, domain conventions, what to leave alone. They are not a place to request new check types or redefine what the QA should catch. The QA agents are engineered to flag only their defined categories dependably, which is what keeps false positives low; instructions that ask for new checks produce unreliable results. When you need the QA to cover something the built-in checks do not, or want to apply an extensive rule set, talk to your Wordbee contact: new checks are added as engineered agents, which is how they stay reliable.
If your needs outgrow what instructions can reliably carry, for example elaborate rule sets, behavior that changes per content type, or logic the workflow was not designed for, a customized workflow is often the better answer: the rules move from free text into the workflow's own design, where they are applied dependably. Contact your Wordbee customer success manager to discuss it.
In post-editing and quality assurance workflows, your instructions can also use the values of segment custom fields: see Using Segment Custom Fields in AI Instructions.
Learn More
-
AI Translation, AI Post-Editing, AI Quality Assurance: What each workflow does with these settings.
-
AI-powered suppliers: The AI Agent tab where post-editing services are configured.