Cast your mind back to the 1990s when Translation Management Systems (TMS) first strutted onstage. Suddenly linguists traded spiral notebooks for digital dashboards and discovered there was, in fact, life after copy‑and‑paste. Today, Large Language Models (LLMs) are repeating the trick—only this time the productivity dial is snapped clean off. Think jumping from dial‑up to 5G overnight.
Yet the narrative isn’t “bots versus humans”. It’s “bots plus humans equals bigger, better, faster”. The machine chews the repetitive bits; the linguist brings the nuance, wit, and cultural flair that keeps audiences hooked. Below, we unpack the full workflow—pre, mid, and post‑translation—plus the ROI maths that has CFOs smiling like they’ve found an extra quarter under the sofa.
1. Pre‑Translation: Glossary Autopilot
The pain
Traditionally, compiling glossaries felt like hand‑knitting a scarf—valuable but painfully slow. Miss a key term and your brand voice drifts faster than a Premier League striker looking for a better contract.
The AI twist
Feed your source files through an LLM‑powered frequency‑analysis routine and watch it spot candidate terms in seconds. It then checks existing termbases, pings style guidance, and surfaces the gaps—all before the first line of Machine Translation (MT) fires. Amazon Translate’s Custom Terminology, Google’s Glossaries API, or open‑source libraries like termextract can slot in here nicely.
Result
A bespoke, on‑brand glossary that nudges MT toward higher quality on draft one. More seasoning, less salty feedback later.
2. Mid‑Translation: Traffic‑Light QA in Real Time
Old routine
MT churns out a first pass; humans read every segment anyway “just in case”. Productivity gains? Meh.
New routine
LLM‑driven quality evaluation grades each segment—green (ship it), amber (needs polish), red (get the humans). Under the bonnet you’ll find COMETKiwi or MQM‑aligned models scoring fluency, adequacy, and dreaded hallucinations. Reviewers dive only where nuance matters, like switching from London dry humour to Tokyo formality without losing face.
3. Post‑Translation: Adaptive MT & Continuous Feedback
What happens
Linguists finesse the reds and ambers, feeding those edits straight back into the engine. Adaptive MT (think ModernMT or SDL’s Language Cloud) hoovers up corrections to retrain on the fly. Tomorrow’s system starts smarter, and PEM‑quality curves inch ever closer to “publish‑ready”.
Quality proof
In pilots we’ve seen LQA (Linguistic Quality Assessment) scores jump from low‑80s to mid‑90s within three feedback cycles, trimming delivery times by ~25 %. Bonus: fewer rounds of client review means fewer midnight emails marked “URGENT!!!”.
4. ROI: Your New Favourite Acronym
| Metric | Pre‑AI Workflow | AI‑Augmented Workflow | Delta |
|---|---|---|---|
| Average words/day/editor | 4 000 | 7 500 | +87 % |
| LQA score threshold met first time | 62 % | 91 % | +29 pp |
| Cost per 1 k words (mixed resourcing) | £45 | £32 | –29 % |
| Time‑to‑market (avg software sprint) | 10 days | 7 days | –30 % |
Numbers based on blended data from SaaS, gaming, and e‑commerce clients across five language pairs.
The savings don’t just shrink invoices; they free budgets for new markets. A mid‑tier retailer we worked with funnelled the surplus into Thai and Arabic roll‑outs—markets they’d shelved for “next year”.
5. Implementation Cheat‑Sheet
- Choose your AI stack wisely – Open‑source (e.g., Marian‑NMT + Hugging Face LLMs) versus enterprise SaaS (OpenAI, Google, DeepL). Check data‑privacy clauses—especially if you handle legal or medical text.
- Automate term extraction first – Glossaries are the cheapest, highest‑leverage entry point.
- Integrate traffic‑light scoring into CAT/TMS – Most modern platforms have API hooks; if not, a webhook and a shared JSON does the trick.
- Set an LQA baseline – Use MQM or your favourite flavour. Measure before bragging.
- Close the loop – Push post‑edit data back to MT/LLM models at least weekly.
6. Human Creativity: The Irreplaceable Variable
AI can mimic style, but it can’t feel the cultural heartbeat that decides whether a slogan sings or stumbles. Translators now pivot from sentence‑level carpentry to brand guardianship, transcreation, and strategy—roles that machines can’t quite grasp, much like teaching a toaster to appreciate Shakespeare.
Final Thoughts
Large Language Models aren’t the end of localisation jobs; they’re the turbo upgrade we’ve craved since the first TMS login screen. Embrace the glossary autopilot, let traffic‑light QA keep you honest, and feed your edits back into the machine daily. The result? Happier linguists, faster launches, and a balance sheet that suddenly looks like it’s had a double espresso.