Our AI translation workflow, step by step

Step 1 · Runs at: all levels

What goes in: your files in your format (for example, DOCX, XLSX, JSON, or XLIFF), or a job in your translation management system (TMS), plus any references: glossary, translation memory (TM), style guide, and instructions.

• Agent: measures the files, counts new, matched, and repeated words, finds what changed since the last batch, and lists open questions.
• Linguist: your project lead confirms the scope and the service level for each content type.
• You: send files and references, and answer the open questions.

What comes out: a confirmed scope, a service level per content type, and a list of answered questions.

Step 2 · Runs at: all levels

What goes in: your glossary, your TM, and the content you have already published.

• Agent: proposes terms, each with its source. Your own glossary, TM, and published content come first. A form seen once is a candidate, not a rule. At AI + Edit + Review, separate agents research each term independently.
• Linguist: at AI + Edit and AI + Edit + Review, the reviewer approves new terms and settles any conflict between the researched options. At AI + Check, the checking linguist approves them.
• You: approve the initial glossary and every new key term (product, feature, and brand names, and legal or regulated terms). Other new terms are approved by your linguist and added to your glossary, which we share whenever you ask.

What comes out: an approved glossary in which every term carries its source.

What you gain: terms your team can trace and defend, used the same way in every batch.

No glossary or style guide yet? We create them with you before production starts.

Step 3 · Runs at: all levels

What goes in: the source text, the approved glossary, your TM, and your rules.

• Agent: translates against the glossary, the TM, and your rules, then runs automatic checks on the result.
• Linguist: no action at this step.
• You: no action at this step.

What comes out: AI output translated against your terminology and rules, with the automatic checks run and anything they catch flagged for the next step.

Your style rules and instructions go into your own rulebook. We turn the rules that can be tested reliably into automatic checks. Linguists judge meaning, tone, and anything else that needs judgment. Each client’s rules live in that client’s rulebook only.

Step 4 · Runs at: all levels, at different depths

What goes in: the AI output from step 3.

• Agent: a native-reader agent sees only the target text and asks one question: does this read like a translation? It flags anything that does. At AI + Check, an accuracy agent also compares source and target and flags meaning risks.
• Linguist at AI + Check: a native linguist checks every flagged segment plus a random sample, fixes errors, and settles every flag. The linguist’s time goes where the risk is.
• Linguist at AI + Edit and AI + Edit + Review: a native post-editor edits every segment.

What comes out: a checked translation at AI + Check, or a fully post-edited translation at the higher levels.

What you gain: text that has been read the way your customers will read it, before the linguist works on it.

Step 5 · Runs at: AI + Edit and AI + Edit + Review

What goes in: the post-edited text.

• Agent: quality assurance (QA) agents check accuracy, grammar, style, and terminology, each blind to the others’ results. Back-translation serves only as a check. It never becomes the target text. At AI + Edit + Review, a red team of agents tries to make the file fail, and it must back every claim with evidence.
• Linguist at AI + Edit: a reviewer runs a sampled quality check.
• Linguist at AI + Edit + Review: a second linguist reviews every segment.

What comes out: a list of findings. A finding counts only when it names a concrete defect and points to a reference: the source, your glossary, a rule, or your instructions. Preferences are logged separately and never block delivery.

What you gain: separate checks catch what a single post-editor can miss, and every finding comes with a reason your reviewers can verify.

Step 6 · Runs at: AI + Edit + Review. At the other levels, the linguist decides findings directly.

What goes in: the findings from step 5.

• Agent: an advocate agent and a devil’s-advocate agent argue each finding, and a judge agent recommends what to do.
• Linguist: the reviewer decides each finding and records the reason.
• You: your instructions outrank everyone.

What comes out: a decision on every finding, with its reason. A script applies the agreed changes.

Step 7 · Runs at: all levels

What goes in: the file as it will be delivered.

• Agent: at AI + Edit + Review, a regression check first re-reads everything that changed. Then an automated quality gate checks the file exactly as it will be delivered. If it fails, the file does not ship.
• Linguist at AI + Edit + Review: the reviewer does a final read of the file before the gate runs.
• You: receive the files, the gate result, the issue log, and your updated TM, with a Translation Memory eXchange (TMX) file on request. At AI + Edit and AI + Edit + Review, you also receive glossary changes and a query log, and quality reports: sampled on an agreed schedule at AI + Edit, full with every delivery at AI + Edit + Review.

What comes out: your delivered files and the evidence for them. Every delivery includes the quality-gate result and an issue log. At AI + Edit, sampled quality reports follow the schedule agreed for your program and can cover several small releases; at AI + Edit + Review, each delivery gets a full report.

If the gate flags something that is not an error, we fix the gate’s rule and record the case in your delivery note. We never edit your file just to make it pass.

What you gain: files that do not reach you with broken tags or placeholders, and a record you can show your stakeholders.