Skip to content

Human or AI-Assisted Translation? Choosing the Right Workflow

For direct clients and translation agencies: how our four workflows compare

 

The question we are asked most often before a project starts is some version of this: do you use AI, and if so, where exactly?

The answer depends on what you send us. We run four workflows: two fully human, two with an AI draft. The choice between them is not a matter of budget alone. It is a matter of what the document is for, who signs off on it, and what happens if a sentence is wrong.

This page sets out all four, so that you can select a workflow before you request a quote, and so that your own compliance and procurement records can be completed without a further round of questions.

 

▮Choosing in One Minute

  • The document will be filed with a regulator, signed by a party, submitted as a sworn translation, or read by someone operating a machine. Choose human translation with revision (HumanExpert Pro). No generative AI, and a second specialist checks the translation against the source.

  • The content must not reach any AI provider, and it will be used internally or reviewed on your side before use. Choose human translation (HumanExpert). No generative AI, one specialist.

  • The content is technical, recurring, or an update to material we have translated before, and you want a translation memory and glossary that serve the next project too. Choose AI draft with post-editing (SmartEdit).

  • The content is high volume but carries real consequences if it is wrong, and you want a second human stage without the cost of two full human translations. Choose AI draft with post-editing and revision (SmartEdit Pro).

  • You already have a translation, from an AI tool, a machine translation engine or another supplier. None of the four applies: what you need is a check against the source, independent review.

If you send us content that does not match the workflow you selected, we tell you before translation begins rather than after delivery.

 

▮The Four Workflows Compared

  Human translation (HumanExpert) Human translation + revision (HumanExpert Pro) AI draft + post-editing (SmartEdit) AI draft + post-editing + revision (SmartEdit Pro)
Best for Confidential content that must not reach an AI provider, used internally or reviewed on your side. Contracts, regulatory filings, medical, safety-critical and sworn content. Technical, recurring content and updates, where the memory and glossary will serve later projects. High-volume content that carries consequences if it is wrong.
Generative AI None, at any stage. None, at any stage. At draft stage only. We produce the draft ourselves, with the whole project or document in the AI's view. At draft stage only. Same draft stage as SmartEdit.
Human stages One. A specialist translator, who checks their own work against the source. Two. A specialist translator, then revision by a second specialist against the source. One. A specialist post-edits every segment of the draft. Two. A specialist post-edits every segment, then a second specialist revises the result against the source.
Working environment memoQ or Trados, configured per client. memoQ or Trados, configured per client. memoQ, Trados or an online translation platform, with your translation memory and glossary from the first segment. Same as SmartEdit.
Content sent to AI providers None. None. Source text with document context, through business accounts that exclude model training. Same as SmartEdit.
Standard Produced under our quality management system, which operates to ISO 17100:2015. With a single human stage, conformity is not claimed. Follows the translation and revision steps of ISO 17100:2015. ISO 17100 does not cover post-edited machine output, so conformity is not claimed. Qualification, terminology and QA are those of our ISO 17100:2015 system. Same position as SmartEdit. The second human stage raises the level of checking, but the text starts from a machine draft.
Speed Standard. One full human stage. Slowest. Two full human stages. Fastest on volume. The draft removes most of the first pass. Between SmartEdit and HumanExpert Pro.
Cost One specialist fee. Highest. Two specialist fees. Lowest on recurring and high-volume content. Between SmartEdit and HumanExpert Pro.
Sworn translation Not available. Sworn work requires revision. Available. Required for sworn translation. Not available. Not available.
Regulatory filings Not recommended. Regulatory content goes to HumanExpert Pro by default. Default workflow. CE technical files, medical device documentation, pharmaceutical files, KVKK, BDDK, SPK and EPDK submissions. Not recommended. Only where you have carried out and recorded your own assessment.
Safety-critical documentation Not recommended. Default workflow. Machinery manuals, operating and safety procedures, hazardous materials documentation. Not recommended. Assessed case by case, in writing, before we accept the project.
Language pairs Every pair in our active network. Every pair in our active network. Pairs where draft quality meets our standard. Where it does not, we switch to a human workflow and tell you why. Same pairs as SmartEdit.
When we propose a change If the content turns out to be legal, regulatory or safety-critical, we recommend adding revision. If volume and terminology stability make an AI draft sensible, we say so. If the content becomes legal, regulatory or safety-critical, we stop and tell you before translation starts. Same as SmartEdit.
 
Already have a translation? That is not one of these workflows but a separate service: we check the existing text against the source and return a corrected file with a report of changes. See independent review.

 

▮When Revision Is Required

Revision by a second specialist is on by default for contracts, regulatory submissions, safety-critical documentation, and medical and pharmaceutical content. It can be removed only on your written instruction. For sworn translation it cannot be removed.

 

▮How We Prepare the AI Draft

In both AI-assisted workflows, the draft is not a raw machine output. Before anything is generated, we read the full text, build the glossary and prompts around your content and terminology, and test the engines (GPT, Claude or DeepL) to apply the one that performs best on that content and language pair. The AI works with the whole project or document in view, not one sentence at a time. The model used and the prompt version are recorded in the project file, and prompts are versioned per account.

 

▮The Same in All Four Workflows

Some things do not change with the workflow you choose. These apply to every project we accept.
  • Your own experts contribute directly. Your engineers, lawyers or product specialists can record terminology decisions with us. Once recorded, those decisions apply automatically to every later project on the account.

  • Repetition and match discounts. Applied against the human stages in all four workflows. There is no separate per-word charge for AI output.

  • Project records. All work runs on our own installation of ]Project Open[, hosted inside the EU and in continuous use since 2009. Records show which specialist handled which file and when, and for the AI-assisted workflows they record the model used.

  • Named specialists. The identity and qualifications of everyone who worked on your project are available to you on request.

  • Written declaration of AI involvement. A signed statement naming the workflow used, and therefore whether generative AI was involved, is issued with delivery on request. It can be set as standard for your account.

  • Confidentiality. A non-disclosure agreement is standard on every project and applies to every specialist assigned to it.

  • Working inside your systems. Where you require us to work in your own tools, we do so under your access controls and your retention rules. Tell us at quotation stage.

  • Work for other language companies. All four workflows are available under your brand, inside your tools, with a non-disclosure agreement as standard.

▮For Your Supplier Records

Corporate buyers now keep a written record of where AI sits in their supply chain. The four statements below are written to be entered directly into that record. If your form asks a question these statements do not answer, send it to us and we will answer it in writing.

 

▮Human translation (HumanExpert)

  • Generative AI used: No, at no stage.
  • Human stages: One. A specialist translator, who checks their own work against the source before delivery.
  • Content sent to AI providers: None.
  • Records: Held in Alafranga's own installation of ]Project Open[, hosted inside the EU, in use since 2009.
  • Standard: Produced under Alafranga's quality management system, which operates to ISO 17100:2015. Conformity with ISO 17100 is not claimed for this single-stage workflow.
  • Responsibility for the delivered text: Alafranga Europe Ltd, registered in England and Wales, company number 16711244.

 

▮Human translation with revision (HumanExpert Pro)

  • Generative AI used: No, at no stage.
  • Human stages: Two. A specialist translator, then revision by a second specialist against the source text.
  • Content sent to AI providers: None.
  • Records: Held in Alafranga's own installation of ]Project Open[, hosted inside the EU, in use since 2009. Both specialists recorded per project.
  • Standard: Follows the translation and revision steps of ISO 17100:2015, within Alafranga's quality management system, which operates to that standard.
  • Responsibility for the delivered text: Alafranga Europe Ltd, registered in England and Wales, company number 16711244.

 

▮AI draft with post-editing (SmartEdit)

  • Generative AI used: Yes, at draft stage only. Model recorded per project. No segment is delivered without post-editing by a specialist.
  • Human stages: One. A specialist post-edits every segment.
  • Content sent to AI providers: Source text of the project, with document context, through business accounts under which the content is not used to train the provider's models. Sub-processor list and retention terms available on request.
  • Records: Held in Alafranga's own installation of ]Project Open[, hosted inside the EU, in use since 2009. Model and specialist recorded per project.
  • Standard: Produced under Alafranga's quality management system, which operates to ISO 17100:2015. ISO 17100 does not cover post-edited machine output, and conformity is not claimed for this workflow.
  • Responsibility for the delivered text: Alafranga Europe Ltd, registered in England and Wales, company number 16711244.

 

▮AI draft with post-editing and revision (SmartEdit Pro)

  • Generative AI used: Yes, at draft stage only. Model recorded per project.
  • Human stages: Two. A specialist post-edits every segment, then a second specialist revises the result against the source.
  • Content sent to AI providers: As SmartEdit.
  • Records: As SmartEdit, with both specialists recorded per project.
  • Standard: As SmartEdit.
  • Responsibility for the delivered text: Alafranga Europe Ltd, registered in England and Wales, company number 16711244.

 

▮Questions We Hear Before Every Project

Can you guarantee in writing that no AI was used?
Yes, for HumanExpert and HumanExpert Pro projects. We issue a signed statement naming the workflow and confirming that no generative AI was involved. This can be set as standard for your account so that it arrives with every delivery.

What is the difference between HumanExpert and HumanExpert Pro?
The second specialist. In HumanExpert, one specialist translates and checks their own work against the source. In HumanExpert Pro, a second specialist revises the translation against the source text. Choose HumanExpert for confidential content you will review yourselves, and HumanExpert Pro wherever the text will be filed, signed or relied on without further review.

Which workflows follow ISO 17100?
Only HumanExpert Pro follows the translation and revision steps that ISO 17100:2015 requires. HumanExpert has a single human stage, and SmartEdit and SmartEdit Pro start from a machine draft, which the standard does not cover, so conformity is not claimed for them. All four run under the same quality management system, with the same specialists and terminology controls.

Which AI models do you use, and can I exclude one?
GPT, Claude and DeepL, selected per project after testing on your content and language pair. You can exclude any provider on your account, and we will record the restriction so it applies to every future project without you having to repeat it.

Does my content train anyone's AI model?
No. AI drafts are generated through business accounts under which submitted content is not used for model training. The current list of AI sub-processors and their retention terms is available on request.

We already have a translation. Which workflow applies?
None of the four. A translation produced elsewhere, by an AI tool or another supplier, is checked against the source in a separate service: independent review. You receive the corrected file and a report of changes.

Can I switch workflows during a project?
Yes, before translation starts, and we will tell you what changes in price and delivery date. If the content turns out to be legal, regulatory or safety-critical, we raise it ourselves rather than continuing on the workflow you first selected.

Do you offer these workflows to other language companies?
Yes. All four run under your brand, inside your tools, with a non-disclosure agreement as standard. We work in more than 40 languages; Turkish is our founding language, active since 2002. Working with agencies

 

Not sure which one your content needs?

Send us a sample file and tell us what it will be used for. We will tell you which workflow applies, and if it is the more expensive one, we will explain why.
→ Send a sample file