Skip to main content

The AI review, the AI check and the second opinion

The AI is the second layer. It does two jobs on a check, and a third when you ask for it.

It reads every segment​

Every translated segment with a target goes to the AI, because a reversed meaning or a dropped sentence usually sits in a segment where no rule fired. It looks for what rules cannot describe: a mistranslation, an omission or an addition, a register your style guide forbids, a grammar agreement error with the segment before. Its findings appear in the report under AI findings, each with the AI's reason and a suggested fix. Grammar (AI) is an agreement error in gender, number, pronoun or case, often between a segment and the one before it. Punctuation (AI) is a mark that changes the meaning or breaks the sentence, a missing or doubled mark, a wrong quote or bracket; an optional stylistic mark is not reported. When a rule check already found the same mark, the AI's finding joins that row, which keeps the rule's label and reads Found by rule check, AI agrees.

Source check (AI) means the source may hold a typo or a word that looks like another one. The finding names the word as the source writes it and the word it looks like. It is minor and has no fix to accept: check the source before you change the translation. A German heading, label or button that ends in an infinitive (Filter reinigen) may be translated as a noun phrase or as an instruction, so the AI does not report its form as an error unless your style guide asks for one form. The What was checked lines count the findings it set aside.

It reads each segment with its neighbours: the source and the translation of the segment before and after it, in the order of the file. A translator often moves a clause into the next segment because the target language builds the sentence in another order. The AI sees where those words went, so a moved clause is not reported as an addition in one segment and an omission in the other. Under each finding the report has a shut Context row with the same two neighbours, so you can check what it saw.

A profile can narrow the review to the segments the rules flagged, which is cheaper; the report then says reviewed flagged segments instead of reviewed every segment. Locked segments (units the file locks, a setting apart from 100% and ICE matches) follow the profile's Locked setting in Findings: not checked by default, and the report says how many were left out and why. Untick Locked (or pick the Delivery role) and the rules, the AI review and the spelling pass all read them, and their findings count; fixes on locked units are not applied, and the report says so. Text the file never opened for translation is never sent. When the review did not finish, the report says how far it got and why, and never a quiet pass.

The AI check of rule findings​

The rules that need a reading (terms, numbers, length, consistency, untranslated text, checklist hits and custom rules) get a second look at each critical or major finding. The AI reads the segment with its tags and what the rule saw: the term and its list, the numbers on each side, the rule as written. Then the finding carries AI agrees or AI disputes with the AI's reason; when it cannot tell, there is no label. A finding that reads Found by rule check is a fact and gets no opinion.

Nothing is removed from the report. A disputed finding stays in the list, counts at half weight in the score until you ignore it or mark it fixed. The reviewer decides. The report's AI check line is one sentence whose numbers add up to the findings, for example "Of 11 findings, the AI agrees with 5, disputes 1, left 4 without a label and did not check 1".

Standard or Premium​

A profile asks one question about review quality: Standard or Premium. Premium uses stronger models and costs more AI words per word; the profile shows what each costs. Premium is coming to Business and Enterprise.

What the AI did, on every report​

Every row of the Reports list carries a tag for what the AI did on that run: Cached, Fallback, No AI check, AI failed or AI skipped. The run writes it from what happened, never from the settings. Inside the report a banner at the top says the same in words and the What was checked lines say how many segments the AI read.

If the AI provider is down or out of quota, the rule layer still runs and the report is still produced: the AI status reads failed or partial in the banner, and rule findings are unaffected. The result the integrations read stays the result of the rules, so a provider outage never turns a clean file into a failed delivery.

The settings​

In Settings, AI review (a workspace admin), top to bottom:

  • AI checks, on by default. Off, only the rule checks run and nothing is sent to an AI provider, for checks from the app, the API, email, Dropbox, the extension and the connectors alike. On a server with no AI provider set up, every check runs the rules only and the report says so.
  • Providers: which AI providers may read your workspace's text, fallbacks included. A provider your plan does not offer is greyed out. AI calls run on our own accounts with each provider, never on keys you supply. Your segments are never used to train models. Security and privacy.
  • Primary model: the model that does the AI review and the AI check of rule findings, among the models of the ticked providers. The list follows the ticks; when you untick the provider of the model you picked, the page says so and picks another. Premium stays a choice per profile, under Review quality. When your account manager has set the models of your workspace, they are shown and cannot be changed.
  • Second opinion, off by default: a second AI model from another provider rechecks each critical and major finding of the AI review; a finding it agrees with is listed as Confirmed. Nothing is removed.
  • Second opinion model, shown whenever a feature uses it: it settles a rule finding the AI disputed on a mechanical check, such as a number or a term, and rechecks the AI findings while the second opinion is on. It is another model than the primary one.
  • Daily AI runs per person: how many AI reviews a linguist or a translator may start in a day.
  • AI check of rule findings, on by default: the AI agrees with or disputes each critical and major rule finding that needs a reading and confirms the extra meaning checks. Off, rule findings stand as found.

In a profile's AI review section, the controls are the profile's own:

  • AI review for this profile: on, every job matched to the profile runs the AI review, panel uploads, API, Dropbox, email and assistant jobs alike. Off, the profile runs the rule checks only, a manual upload included. The Run page says so. Not set, a panel upload follows the Run page box and the automatic jobs run the rule checks only.
  • AI check of rule findings and Second opinion, each a three-way choice: Inherit follows the workspace setting named in the option, On and Off decide for this profile whatever the workspace says. A shared account or division scope can carry the same choice beneath a profile.
  • Critical only, off by default: the AI review also flags register, tone and word choice until you turn it on. The profile's AI guidance is what the AI reads besides the segment.

AI checks off for the workspace stops every one of them; the profile's section is greyed and keeps its values for when they are back on.

For one check, the Settings for this check on the Run page override the defaults, except that a profile with its AI review switched off runs the rule checks only.

In the API​

The fields keep their names: secondOpinion (and second_opinion in profiles, MCP and the checks of a report) is the AI check of rule findings; ai_second_opinion is the second opinion on AI findings.

The second opinion​

When the workspace or the profile turns it on, a second model from another provider rechecks the AI review's critical and major findings. A finding it agrees with is listed as Confirmed; one it does not confirm stays, without the label. It never removes or downgrades a finding.

What it costs​

The AI review counts against the workspace's monthly allowance of AI words. The other AI passes (the AI check of rule findings, the second opinion, the spelling check, the confirm lane, fix completion, the whole-file review and the transcreation check) are recorded on the run and not counted. Rules are not metered. The AI words chip in the top bar shows this month's use. Plans and allowances.