On this page
The model in one tableCalculatorReading the resultGetting real ratesCosts that hide, and savings that hideMore on thisA finance team asks localization a simple question: if we move to AI-assisted translation, what do we save? The answer they get on AI translation cost is a single percentage from a vendor's deck. Six months later the Korean budget has barely moved, because most of the company's Korean volume was contracts and regulated labels, which still need full human translation. The saving was real, but only on the content that was never the expensive part.
A hybrid pipeline mixes machine translation, post-editing, human translation and review, routed by content risk. Whether it saves money, and how much, depends on how your volume splits across those routes. This post gives a simple cost model and a calculator you can fill in with your own figures.
The model in one table
| Cost line | Driver | Notes |
|---|---|---|
| Human translation + revision | Words in the critical tier × rate | Contracts, regulated, safety, high-visibility marketing |
| Full post-editing | Words in the customer-facing tier × post-editing rate | UI, website, onboarding; ISO 18587's full level |
| Light post-editing or sampling | Words in the support/internal tier × lower rate | Help center, internal docs |
| Machine translation | Characters processed × engine price | Usually small next to human costs |
| Translation memory savings | Repetitions and matches × discount grid | Applies to every tier; use a real analysis |
| Setup and run costs | Hours for glossary, configuration, QA, sampling, lead linguist | Front-loaded; often forgotten |
Translation memory discounts deserve a note: vendors analyse new text against your memory and charge each band (repetitions, exact matches, fuzzy matches, new words) at a different share of the full rate, using a grid agreed in the contract. Tools such as Phrase implement these net rate grids directly. Budget from a real analysis of your files, not from a guessed discount.
Calculator
All default values are illustrative placeholders, not market rates. Replace them with your volumes and your vendor's quoted rates. The calculator compares the hybrid model with sending every word to human translation and revision.
"Fixed costs" covers tooling, engine fees and the money value of setup hours; the hours field is shown for planning only. Memory discounts are not included: apply them to each tier from your vendor's analysis.
Try the two presets. With the same rates, a support-heavy content mix shows a large difference; a regulated-heavy mix shows a small one, and with higher fixed costs the hybrid model can even cost more. That is the point of modelling before committing.
Reading the result
Take the support-heavy preset, with its placeholder rates. Two hundred thousand words at the human rate cost 40,000 in the calculator's units. In the hybrid model, the 20,000 critical words still cost 4,000, the customer-facing tier costs 6,000 at the full post-editing rate, the support tier 6,000 at the light rate, and fixed costs add 5,000: 21,000 in total. Almost the entire difference comes from the support tier, which is 60% of the words. Now look at your own split. If your largest tier is critical content, the honest conclusion may be that a hybrid pipeline is worth it for speed and consistency, but not for budget. That is still a useful finding, and much better made before a vendor switch than after.
Getting real rates
- Pilot on your content. Run a few thousand words per tier through the engine with your glossary, and ask vendors to quote post-editing on that real output.
- Measure effort. Record post-editing time per thousand words and edit distance. If post-editors rewrite most segments, the post-editing rate will converge with the translation rate.
- Price review separately. In-context review, sampling and lead-linguist time are separate lines; do not let them vanish into a per-word rate.
- Revisit after three months. Glossary and prompt improvements usually reduce effort; renegotiate on data.
Costs that hide, and savings that hide
Rework
Light post-editing on content that needed fullSupport tickets, corrections, re-releasesRouting too much to the cheap tier moves cost to support and engineering.Ownership
No lead linguistThe same errors fixed again and againA few hours a week of ownership often pays for itself in fewer edits.Speed
Korean shipping with EnglishRevenue and support savings from simultaneous launchNot in the translation budget, but real. Count it separately.Quality estimation can move words between tiers automatically: Phrase's Quality Performance Score, for example, can be used with thresholds to decide which segments skip or receive human editing. Set thresholds conservatively at first, check them against human sampling, and only then let them drive cost.
For the roles behind the pipeline, see the hybrid human-AI operating model; for QA that keeps cheaper tiers safe, sampling-based and risk-based translation QA; for post-editing levels, light versus full post-editing; and for a word-count budget by content type, budgeting localization with translation memory.
Want a Korean cost model built on your files?
We analyse your content by tier and match band, pilot post-editing on your real machine output, and give you per-tier rates to plug into the model. Request an analysis →