Every AI feature in wxrks runs as an operation, and each operation sends a defined set of content to the model that serves it. This article lists, for each operation, what it does and exactly what wxrks sends, so you can judge what your content exposes before you enable a feature or route it to your own engine.

💡 Who is this for? Account Admins who review what data reaches an AI model, and those who route operations to their own engine in Model Management.

How to read this article

Operation: the names below are the values of the Operations field in Settings > AI > Models > Add configuration.
Which model receives the content: by default, the wxrks default models. If one of your configurations on the Models page covers the operation, your engine receives it instead. Routing works the same way for every operation, except Text to Speech, Embedding, TMill QA and TMill Label Classification, which have their own rules. See Operations you can route to your engine in Connect your own AI engine.
What wxrks sends to the model: the content wxrks puts into each request. Nothing else from the project or account is added. For which provider receives the data and what it commits to about retention and training, see wxrks AI Privacy: Data Shared with AI Providers.

What you can do as an Account Admin

Use the tables to decide, operation by operation, where your content may go:

Keep a feature from sending anything. Turn off its action in Settings > AI > Augmented actions. An action that is off sends no content. See wxrks Augmented Actions for which actions you can turn off and at which level.
Send the content to your own engine instead of the wxrks default models. In Settings > AI > Models, click Add configuration, turn off Default model for all operations, and pick the operation in Operations. Model Management requires the wxrks FUSION plan or higher.
See which operations ran. Open the Usage tab on the Models page. Tokens by operation lists each operation that ran and the tokens it used.

ℹ️ Note: Some operations start without anyone choosing an AI action. They run on a schedule or as part of another step:

In the background: QA Smells, Infer Translation Context, Message Feed, Home Insights and Alert Outbound (when a project alert is created).
During pre-translation: CS Translation and Add Missing Tags.
While a linguist works: Spell Check (as they type) and Term Extraction (when they confirm a segment).
On upload: Detect Language (portal uploads) and Embedding (reference set documents).

Check these first if you want to know what leaves wxrks without a deliberate AI action.

Translation and quality evaluation

Operation	What it does	What wxrks sends to the model
CS Translation	Writes the final translation of a segment: Context Sensitive Translate in the Editor, AI pre-translation, and translation in Free Flow (the editor where you translate or edit a whole document as running text).	Source segment with tags and the language pair; the top TM matches; glossary terms, including forbidden terms; the MT output when there is no TM or glossary match; Sous Chef instructions and profile (domain, audience, formality, tone, creativity); the segment comment and character limit, if any. Neighboring segments are not sent.
Smells	The Smells check a linguist runs in the Editor. Lists the quality problems in a translated segment, each with a type, a severity and an explanation.	Source and target segment with tags; the language pair; glossary terms found in the source; TM matches of 75% or more; the Sous Chef profile; the reasoning of the CS Translation and fixes already applied.
QA Smells	The batch review that produces Confidence Scores, including the score used by Auto-Confirming Segments by Confidence Score.	Source and target segment with tags and the language pair, in batches of 50 segments. Depending on the account's configuration, also the glossary, TM matches and Sous Chef context that Smells sends.
Fix Smells	The automatic fix of a finding: Fix or Fix all in the Editor's Smells panel. Changes only the smallest span needed.	Source and target segment with tags; the language pair; the finding, or every open finding by severity; glossary terms, TM matches, the Sous Chef profile and previous fixes.

ℹ️ Note: The Confidence Score comes from Smells. If you route QA Smells to your own engine, your engine scores the translations. If you route CS Translation but not QA Smells, your engine writes the translation and the wxrks model scores it.

Editing assists

Operation	What it does	What wxrks sends to the model
Suggestion	Augmented actions > Alternative suggestion in the Editor: an improved version of the translation, with an explanation.	Source and target segment with tags; the language pair; the character limit, if any. No TM, glossary or MT.
Proofreading	Augmented actions > Proofread in the Editor: rewrites the target to read better while staying faithful to the source.	Source and target segment with tags and the language pair. Nothing else.
Proof Detect	Improve on selected text in Free Flow: fixes grammar and spelling of the selection.	The selected text only.
Free Flow Edit	Free Flow in Proofread or Rewrite mode: edits text that is already in the target language.	The paragraph's text and formatting; the target language and mode; the Sous Chef profile; glossary terms, including forbidden terms; the original text as a reference, for variants. No TM matches.
Fix Tags	Augmented actions > Fix tags in the Editor, and tag placement during Context Sensitive Translate: puts the source tags into the target without changing the text.	Source segment with tags and the target text without tags. No language pair.
Add Missing Tags	Fix all in the tag mismatch check before delivery or download, and after pre-translation: adds only the missing tags.	Source and target segment with tags and the list of missing tags, for segments with missing tags only.
Spell Check	Active SpellCheck: spelling, grammar and punctuation errors as the linguist types and in QA.	The target text and target language. The source text is not sent.

Review, terminology and context

Operation	What it does	What wxrks sends to the model
Assess Review	Review Assessment in the Editor: classifies a reviewer's change with the account's error categories and scores the translation and the review. Also suggests a severity when a reviewer logs an error without one.	Source, the previous step's translation and the reviewer's text; the language pair; the account's error categories and severities.
Term Extraction	Learns terms when a linguist confirms a segment, if the project glossary learns terms.	Source and target segment; the language pair; the file's term list.
Glossary Term Inflection	Adapts glossary target terms (gender, number, tense) to fit the segment, without translating them again, when Augmented Term Lookup is on for the project glossary.	Source segment and current target draft; the language pair; up to 10 glossary pairs already matched on the segment.
Part of Speech	Suggests the part of speech of a term being added to the glossary in the Editor.	The term, the segment's source text (up to 2,000 characters each) and the source language.
Sous Chef	Three translations of a segment adapted to the Sous Chef profile (domain, tone, formality, audience and instructions), from the Editor's Sous Chef panel.	Source segment with tags; the language pair and target country; the stored text before and after the segment; the Sous Chef profile and instructions; the character limit, if any. No TM, MT or glossary.
Infer Translation Context	Works out a Sous Chef profile from the file when none with instructions applies. Runs about once per file.	The file's source text, up to 25,000 characters, and the source language.
Detect Language	Identifies the language of a text: portal uploads, Free Flow, The Aligner, reference sets and Text to Speech.	A sample of the text: the first 50 to 2,000 characters, or up to 200 paragraphs cut to 200 characters each.

Assistant, reference sets, TMill and audio

Operation	What it does	What wxrks sends to the model
Agent	The Assistant chat, and Ask in the translation portal. Answers questions about the workspace by looking up projects, glossaries, languages and Drive files. Changes, such as new glossary terms, are only proposed until the user approves them. Requires the Use the Assistant permission.	The user's message and the chat history; documents uploaded to the chat; what the lookups return (project names, status, languages, progress, people, dates and cost; glossary terms; Drive file names), limited to what the user may see. In the portal, also the language pair, the Org Unit's Sous Chef profile and the quoted translation.
Embedding	Makes reference sets (collections of reference documents, such as style guides) searchable by meaning.	The text of each uploaded reference document, in chunks, and each search query. Only numbers come back.
Reference Summary	Writes the brief of a reference set from its documents, and drafts a Sous Chef profile from that brief.	Each document's text; the notes and file names when they are merged; the set's summary instructions, if any.
TMill QA	The Translation Smells option of TMill, the Labs tool that cleans translation memories: reviews each TM pair, scores it from 1 to 5 and suggests a fix.	Source and target of each TM pair; the language pair; the What is this content? text, if filled in.
TMill Label Classification	Maps the problems TMill QA finds onto the account's quality dimensions, to score the run.	The account's quality dimensions and the problem labels, each with up to 3 short example explanations.
Text to Speech	Generate audio in Free Flow, video voice-over and custom voices.	The text to read, the voice, a style description (tone and pace) and, optionally, the language.

Feeds, alerts and other

Operation	What it does	What wxrks sends to the model
Message Feed	Classifies each chat message (sentiment, category) and drafts a reply for the Feed. Also writes the project alerts: late, at risk of delay, ready for delivery and not approved.	The message text and the channel type, once per recipient. For alerts, the alert type and, for each project, its name, reference, dates and Org Unit.
Home Insights	Writes the one-line reason for each item in Needs a decision on the Account Admin's Home (on accounts with the new Home).	The type of each item and its counts, amount and currency. No segment, file or message text.
Alert Outbound	Drafts a short message about a late or at-risk project to the vendor with the open task or to the client contact.	The alert type, the project name and due date, and the recipient's first name and role.
Alignment	Matches source and target segments in Tools > The Aligner.	The language pair and the numbered source and target segments, 50 of each at a time.
Web Preview	The AI option in the Editor's web preview: places translations into the live page where exact matching fails.	The source language, text from the page and candidate source segments. Translations are not sent.

Quick reference: how much content each operation sends

Amount of content	Operations
One segment, paragraph, term, TM pair or selection per request	CS Translation, Smells, Fix Smells, Suggestion, Proofreading, Proof Detect, Free Flow Edit, Fix Tags, Add Missing Tags, Spell Check, Assess Review, Term Extraction, Glossary Term Inflection, Part of Speech, Sous Chef, TMill QA
Batches of segments	QA Smells (50 segments), Alignment (50 source and 50 target segments), Web Preview (page text and candidate segments)
Whole documents	Infer Translation Context (up to 25,000 characters of the file), Embedding and Reference Summary (reference set documents), Agent (documents uploaded to the chat), Text to Speech (the text to read)
A text sample only	Detect Language
Messages	Message Feed
No translatable content	Home Insights (counts and amounts), Alert Outbound (project name, due date, recipient's first name and role), TMill Label Classification (problem labels and short explanations)

Connect your own AI engine for automatic post-editing (Context Sensitive) and quality evaluation (Smells): which operations you can route, and the routing rules.
Configure Custom AI Providers with Model Management: the full Model Management walkthrough.
wxrks AI Privacy: Data Shared with AI Providers: providers, retention and training commitments.

What wxrks sends to the AI model for each operation