How to Cut Localization Costs with AI Post-Editing
You cut localization costs by routing content to three quality tiers instead of treating every word the same: MT-only for low-risk strings, MT plus light post-editing for support content, and full human translation for legal and marketing. Done correctly, blended cost per word drops 30-50% while the strings users actually read stay clean.
The trap is applying one tier to everything. Full review on 40,000 UI strings is waste. No review on your pricing page is a lawsuit.
Why does MTPE often fail to save money?
Because post-editing can cost more than translating from scratch when the MT output is bad. A linguist fixing scrambled machine output spends time unlearning before rewriting.
The fix is measurement. Track post-editing effort per segment using edit distance or time-on-segment, not just the invoice. When a content type consistently needs heavy editing, that content type should move up a tier, not stay in the cheap lane with a permanent cost overrun.
Rules that hold in practice:
- If a linguist edits more than 60% of a segment, MT added nothing. Move the content type up.
- If a linguist edits under 15%, the content can probably drop a tier.
- Watch short strings. UI labels under five words have the worst MT quality ratio because there is no context.
How do you build the three tiers?
Map every content type to a tier once, in a table, and put that table in your localization brief.
| Tier | Content | Process | Relative cost |
|---|---|---|---|
| 1 | Legal, pricing, privacy, brand headlines | Human translation or transcreation | Highest |
| 2 | Help center, onboarding, release notes | MT plus full post-edit | Medium |
| 3 | UI strings with glossary locked, internal tools | MT only plus spot check | Lowest |
Two-thirds of a typical SaaS company’s word volume lives in tier 3. That is where the savings are.
How do you make tier 3 safe?
MT-only is acceptable when the failure mode is cosmetic. It is not acceptable when a wrong translation changes what a user does.
Guardrails for tier 3:
- Lock the glossary. Product names, feature names, and UI verbs must be do-not-translate or fixed translations. Unlocked glossary is what produces a settings menu in three different languages.
- Exclude interpolation placeholders from translation. Broken variables in translated strings are a common production bug.
- Spot check 5% of every batch in each language, chosen randomly. Catch systematic errors, not one-offs.
- Pilot with one language for four releases before widening. A tier-3 policy that works in French may fail in Japanese.
How much does this actually save?
Worked example: 100,000 words a quarter.
- All tier 2 at $0.15 a word: $15,000.
- Rerouted as 20k tier 1 at $0.22, 30k tier 2 at $0.15, 50k tier 3 at $0.06: $4,400 + $4,500 + $3,000 = $11,900.
That is a 21% reduction, not 50%. The bigger savings come from the second effect: fewer review cycles and fewer post-launch fixes in languages you can barely audit. Teams that measure total cost including rework typically report 30-45% reductions.
Be skeptical of anyone quoting 70% savings. Those figures almost always come from comparing MT to full human translation while ignoring the review labor that keeps MT publishable.
What breaks when you push too cheap
Support content in tier 3. Users reading a badly translated help article file the same ticket twice and lose trust in the product. Help content deserves tier 2.
Screenshot and video captions. Translating captions without updating the screenshots produces instructions that reference UI that no longer matches. Localization is not only text.
Right-to-left languages. Arabic and Hebrew need layout work, not just string translation. A tier-3 string that overflows a button is a design bug MT cannot fix.
How do you keep quality visible?
Report two numbers quarterly: cost per word and user-reported localization issues per language. If cost drops and issues stay flat, the routing is working. If cost drops and one language’s complaint count climbs, that tier assignment was wrong.
That pairing is the whole governance model. Cost alone is a metric that always looks good right before quality collapses.
Key Takeaways
- Route content into three quality tiers instead of reviewing everything identically.
- Tier 3 MT-only is safe when the glossary is locked and placeholders are excluded.
- Measure post-editing effort per segment; above 60% edits means MT added nothing.
- Realistic savings are 30-45% blended, not the 70% vendors quote.
- Track localization complaints per language alongside cost per word. Cost alone hides quality collapse.
Translation workflow options are covered on the Welocalize page at iAIFeed.
