Breadcrumbs

AI Quality Assurance

AI quality assurance adds AI-driven checks on top of Wordbee Translator's built-in QA rules. The built-in rules are pattern-based; the AI agents read the segment, its context, and your reference material, so they catch what patterns cannot: a term that exists but is wrong for this context, a date format that does not match the target locale, or a qualifier silently dropped from the translation.

What the AI Checks

The checks are performed by two kinds of agents.

The form and terminology agent runs five checks:

Check

What it flags

Terminology

The translation renders a source term differently from your term base or memory references. Term base entries count as the authority, and grammatical variations required by the target language are not flagged.

Identifiers and codes

Usernames, model numbers, file names, and version strings that were altered, translated, or re-cased.

Dates and numbers

Dates and numbers not adapted to the target locale's conventions, or whose value changed. The value itself is never questioned, only its form.

Bracketed references

Reference codes such as ticket numbers whose content or brackets changed.

Tags and placeholders

Markup or placeholders that are missing, added, mangled, or translated.

The meaning agent runs two checks:

Check

What it flags

Additions

Content in the translation that is not in the source and changes what the reader would understand.

Omissions

Source content missing from the translation, with time expressions such as "daily" or "tomorrow" always treated as significant.

The single-agent workflow (for example LQA workflow: 1 Agent) runs the form and terminology checks. The two-agent workflow (for example LQA workflow: 2 Parallel Agents) runs both agents side by side over the same content, adding the meaning checks.

All agents are tuned to flag only what clearly matches these categories, because a false QA flag costs reviewer time. Politeness words, articles, and grammar the target language requires are never flagged. Where possible, each issue comes with a suggested correction.

The catalogue of AI checks keeps growing: new checks are added to this page as they become available, so it is worth checking back here from time to time.

Issues appear in the Editor's QA panel alongside the built-in check results, with a description of what was found. Unlike other workflows, quality assurance evaluates every segment regardless of who produced it, and skips only locked segments. Both behaviors come from the workflow's parameters (Skip Segments By Origin and Skip Locked Segments) and can be changed per configuration (see Configuring AI Workflow Parameters).

AI Checks and Built-In Rules

AI analysis complements the built-in QA rules; it does not replace them. The built-in rules are deterministic: they run instantly, cost nothing, and return the same result for the same input every time. For anything a pattern can decide, such as missing tags, number values, punctuation, forbidden characters, or length limits, the built-in rule is the better and quicker choice.

Reserve the AI checks for what patterns cannot judge: whether a term is right for its context, whether formatting matches the target locale's conventions, and whether the translation says more or less than the source. AI checks take time to run and consume AI usage, and their findings are judgments rather than rule matches, so a focused AI configuration on top of a solid built-in profile gives better results than moving everything to the AI.

How to Run It

The two ways to run AI quality assurance behave differently:

  • From a QA profile: the workflow behaves as a regular QA analysis. It examines the content, reports its findings in the QA panel, and changes nothing else. The AI checks run wherever the profile runs: Run QA in the Editor, Run QA on the job's details page (the Check translations popup), or an automated QA step between jobs.

  • As a supplier service on a job: the workflow behaves as a supplier. The job starts on its own, the QA analysis runs, and when it finishes the findings are attached to the content and the job is completed automatically, so the workflow can move to whatever comes next.

From a QA profile

Add an AI Analysis rule to a QA profile:

  1. Go to Settings > Quality Assurance > QA Profiles and edit a profile.

  2. Add the AI Analysis rule and select a workflow. The rule offers two slots, so you can run two workflows in one QA pass, each with its own configuration.

  3. Configure the parameters, including custom instructions per agent (see Configuring AI Workflow Parameters).

The AI checks then run whenever that profile runs: when someone selects Run QA in the Editor or on the job's details page, or as an automated QA step between jobs in your workflow.

Two QA run options narrow what the entire run checks, built-in rules and AI alike: Do not check segments in green status and Do not check locked segments. For locked segments, the workflow parameter Skip Locked Segments is an additional, independent filter: if either it or the run option excludes locked segments, the AI does not check them. To have the AI check locked segments, untick the run option and disable Skip Locked Segments in the AI Analysis rule. For confirmed (green) segments, the run option is the only control; there is no workflow parameter for them.

As a supplier service on a job

Quality assurance can also be a workflow step of its own: define a supplier service with an AI agent using a quality assurance workflow, and assign it to a job, exactly as described in AI Post-Editing. The job runs automatically and the findings are attached to the content.

Note that the QA profile and the supplier service hold separate configurations, even for the same workflow: see Configuring AI Workflow Parameters.

Learn More