Translation Technology Since 2002What changed, and what we did each timeTranslation has changed several times since 2002: from fax to email, from typing to translation memory, from single translators to shared online teams, and now to AI. Each time, we changed how we work, kept what helped clients and dropped what did not. Below is what changed, what we did, and what it means for your project today. |
▮2002: Fax and Retyped Sources
Alafranga was founded in Istanbul in 2002, at the end of the fax era. Documents arrived by fax, poor transmission meant retyping by hand, and retyping meant errors. The translator's first job was often to rebuild the source text before translating it.
What we did. We made checking the source the first step of every project, before any translation starts.
What it means for you. Scans, photographed pages and PDFs exported from other systems are checked and, where needed, rebuilt before translation. → Document translation
▮Around 2005: Translation Memory
Computer-assisted translation tools removed the most wasteful part of the job: translating the same sentence twice. Translation memory stores approved sentences and reuses them across a project and across years.
The same applies if you translate with AI in-house: without a translation memory and glossary, every AI draft starts from zero and your terms change from one document to the next.
The same applies if you translate with AI in-house: without a translation memory and glossary, every AI draft starts from zero and your terms change from one document to the next.
What we did. We started working in Trados around 2005, when a licence still came on a USB dongle, and later added memoQ. From the first project, every client got its own translation memory and glossary.
What it means for you. When a manual is revised, you pay for what changed, not for the whole document again. Every sentence has a history, and the memory is yours: we return it in TMX on request.
▮2009: Managing Projects, Not Just Files
As projects grew to several languages and several translators, email and spreadsheets were no longer enough to track who was doing what.
What we did. In 2009 we moved project management to ]Project Open[, an open-source system we have configured to our own way of working ever since. The same year we introduced a structured application process for translators.
What it means for you. Every project since 2009 can be traced: who translated it, who reviewed it, what was quoted and what was delivered. → How we run projects
▮Around 2010: One Translation Memory for the Whole Team
Large projects need several translators at once.
What we did. We invested early in a Trados GroupShare server for ten users, so a team could work on one shared translation memory. In 2011 a multilingual programme for an international restaurant franchise group ran on it. Today we run most team projects in Smartcat, and we hold Phrase accounts for clients who work there.
What it means for you. A sentence approved by one translator is available to the others at once, so terminology stays the same across the team. Your own engineers or legal staff can be given access, and the terms they decide apply to every translator and every later project. If you already work in Trados, memoQ, Phrase or Smartcat, we work in your environment. If each one works on a separate copy of the translation memory, the same term ends up translated three different ways. If several people in your company translate with AI, the same rule holds: one shared memory and one glossary, or the same part ends up with three names.
▮2017: Machine Translation at Scale
Machine translation became good enough to produce a usable first draft for some content, and fluent enough to hide its errors.
What we did. We had post-edited smaller machine-translated jobs before. In 2017 we took on our first large programme: 2.2 million words of e-commerce product content in French and German for a translation agency partner, with teams coordinated across three continents.
What it means for you. We learned where machine output fails before most of our clients started using it: misread references, drifting terms, changed figures. Those are the points our reviewers check first. These are the same errors we now see in AI translations that companies produce in-house.
▮2018: SmartEdit, a Controlled AI Workflow
Faster drafts need stricter control, not less.
What we did. In 2018 we turned that experience into SmartEdit: a machine draft, edited by a specialist inside the CAT tool and signed off by a named person. When large language models arrived, the engines changed. The controls stayed. The same controls can be set up inside your company: who checks the AI draft, against which terms, and who approves it before it goes out.
How SmartEdit works today:
- The whole document, not one sentence at a time. A term chosen on page 2 is the same term on page 40, and a reference to an earlier section is translated knowing what it points to.
- The engine is chosen by testing. We run GPT, Claude and DeepL on your own sample files and use the one that performs best for that content and language pair.
- Editing happens in memoQ or Trados, against your translation memory and glossary, by a specialist in your subject.
- A named specialist signs off. We do not send unedited AI output to clients.
SmartEdit Pro adds an independent second reviewer, for customer-facing, regulated and complex technical content. Contracts, regulatory submissions and safety-critical documents stay fully human. → Compare the four workflows
▮Interpreting Moves Online
Conferences, works council meetings and training sessions moved from meeting rooms to video calls, and interpreting moved with them.
What we did. We moved from on-site to remote interpreting, in the meeting tools our clients already use, with the same interpreter teams.
What it means for you. Interpreting for an online or hybrid meeting without flying interpreters in. On-site interpreting is still available when the meeting needs it. → Interpretation
▮2021 to 2022: Subtitle Teams in the Cloud
Streaming services began launching in new markets with whole catalogues at once, on deadlines no single subtitler could meet.
What we did. For a streaming launch in Türkiye, more than two dozen subtitle translators worked in parallel in the client's cloud subtitling tool, on a three-stage workflow: translation, quality check and proofing. 647 episodes and films were delivered in five weeks.
What it means for you. High volume does not have to cost consistency, as long as the team works in one environment under one review routine. → Subtitling
▮Today: AI Agents for Repeat Checks
AI can now do more than draft. It can read files, compare terms and flag likely errors.
What we did. For larger or recurring programmes we can set up AI agents for repeat checks: reading incoming files for problems before translation starts, checking terminology across files, and flagging likely errors before the specialist's review.
What it means for you. Fewer errors reach the reviewer, and the reviewer spends time on language, not on missing tags. An agent flags. A person decides.
|
Your company is going through the same change Many companies now translate with AI in-house, much as agencies did in 2017: fast, fluent and not always checked. Every shift on this page was a change we went through ourselves, keeping control of quality each time. |
▮What Has Not Changed
Technology changes the speed, the file formats and the tools. It has not changed what good translation needs:
- A specialist who understands the subject
- A reviewer who checks against the source independently
- A project manager who knows the client's context
- A quality framework that holds all three to account
These four have been constant since 2002. The tools around them have changed. The discipline has not.
▮Questions We Hear Before Every Project
How has Alafranga adapted to changes in translation technology? Step by step since 2002: translation memory in Trados from around 2005, project management in ]Project Open[ from 2009, a shared Trados GroupShare server from around 2010, large-scale machine translation post-editing from 2017 and our controlled AI workflow, SmartEdit, from 2018. Each time we kept what helped clients and dropped what did not. The subject specialist, the independent review and the named sign-off have stayed the same.
How does AI translation work at Alafranga? AI produces a draft and a specialist decides what is delivered. We produce the draft with the whole document in view, not one sentence at a time, so terminology and cross-references stay consistent. The draft is imported into memoQ or Trados, checked against your translation memory and glossary, and edited segment by segment by a subject specialist. A named specialist signs off every file before delivery. We do not send unedited AI output to clients.
Which AI engines do you use? GPT, Claude and DeepL. We are not tied to one of them. Before a programme starts, we run them on your own sample files and use the one that performs best for that content type and language pair. In general, GPT suits general and narrative content, Claude technical and structured content, and DeepL high-fluency European language pairs.
Do you use AI for all translation projects? No. We assess content type, subject complexity and quality requirements before choosing a workflow, and we say so when AI is not the right tool. Contracts, regulatory submissions and safety-critical documentation go to HumanExpert Pro by default: one specialist translates and a second, independent specialist revises against the source, with no AI drafting.
How do you manage large projects with multiple translators? On one shared translation memory. We invested early in a Trados GroupShare server, and today we run most team projects in Smartcat, or in Phrase, Trados or memoQ when the client already works there. A segment approved by one translator is available to the others at once, and project managers see progress, reviewer assignments and delivery status in one view.
Can our in-house specialists contribute to the translation? Yes. Engineers, legal teams or medical staff on your side can be given access to the shared translation memory. When they decide how a term is translated, that decision applies to every translator on the project and to every later project. Your team's knowledge stays in the translation memory instead of a one-time briefing document.
Do you use AI agents? On larger or recurring programmes we can set up AI agents for repeat checks: reading incoming files for problems before translation, checking terminology across files and flagging likely errors before review. Agents flag issues; a named specialist decides what changes.