Localization
Amogh Karmarkar · Jul 15, 2026 · 9 min read
You've narrowed it to three ways to get content into another language. Let a machine do it, pay a human, or run a machine draft past a human editor. The three don't cost the same, take the same time, or carry the same risk. Get it wrong and you pay either way, either on human hours you didn't need or on a fluent-looking error that goes out in front of a customer.
Neural machine translation got a lot better after 2020, which is what makes this call harder now rather than easier. The output reads fine. Whether "reads fine" clears your bar is the real question. This piece walks through the four things that decide it (accuracy, cost, speed, and risk), then says which kind of buyer should pick what.
Raw machine translation suits high-volume, internal, low-stakes text: usually fluent, occasionally critically wrong, almost nothing per word, back in seconds, effectively unlimited scale, and not native quality. Human translation suits high-stakes creative, legal, and brand-defining copy: native and culturally adapted, the highest cost, days to weeks, limited by people at roughly 2,000 words a day. AI + human post-editing suits the broad middle of docs, support, product, and most business content: near-native, with a human catching what the machine misses, at about 30% to 70% of full human rates, back in hours to a day, about 66% faster than human, at roughly 6,000 to 7,000 words a day per editor. Raw MT's biggest risk is fluent errors nobody catches, human's is cost and timeline on large volumes, and the hybrid's is an editor signing off on a weak draft.
Human wins on accuracy. The more useful question is how machine translation fails, because that tells you when you can trust it. A 2024 arXiv study, "Cyber Risks of Machine Translation Critical Errors," looked at Arabic mental-health tweets and found that of about 2,200 mistranslations Google Translate produced, roughly 800, about 36%, were critical errors that changed clinical meaning (arXiv, 2024). Raw MT rarely sounds broken. It sounds sure of itself, and that's the problem.
Neural systems also hallucinate. They produce grammatical, natural output that has drifted away from the source (MIT Press, TACL, 2023). A clumsy sentence gives itself away. A fluent one that happens to be wrong can get past a reader and past automatic quality scores. Idioms, honorifics, on-screen context, and regulated phrasing are the spots where a machine guesses and a human editor knows.
You can measure all of this. The industry uses the Multidimensional Quality Metrics framework, updated to MQM 2.0 in 2024, which scores a translation by counting errors and weighting them by severity (MQM Council / AMTA, 2024). Judged that way, machine and human output don't sit at two points on one line. They break in different places. AI + human is the setup that removes the machine's worst errors without paying a person to retype the easy majority.
Cost is where machine translation earns its keep, and where the hybrid model tends to win. Raw MT costs almost nothing per word. Full human sits at the top. Post-editing lands between them, and language companies usually price it at 30% to 70% of full human rates, for blended project savings around 30% to 50%. Those figures are directional, since they come from internal LSP pricing rather than one audited study, but they point the same way everywhere you look.
At enterprise scale the payoff shows up as ROI. A 2024 Forrester study commissioned by DeepL modeled a composite multinational and reported 345% ROI over three years and 2.79 million euros in efficiency savings from adopting machine translation (DeepL / Forrester, 2024). In a separate 2024 survey DeepL ran with Regina Corso Consulting, 96% of organizations said localization returned a positive ROI, and 65% put it at 3x or more (DeepL, 2024). Both are vendor-funded, so read them as motivated but pointing in a believable direction.
One caveat those ROI numbers skip. The savings only appear if the machine draft is good enough to edit instead of redo. On clean, repetitive business content, it usually is. On brand copy, it usually isn't, so price on its own can't settle the decision.
Speed is the hybrid model's clearest win, with one wrinkle worth knowing. A 2023 study published through the Chartered Institute of Linguists, one of the largest on the question, covered 90 million words, 879 translators, and 11 language pairs. Post-editing came out about 66% faster than translating from scratch (CIOL, 2023).
That average also hides a wide spread. It ran from 130% faster on English to French down to 7% slower on English to Swedish. Editing a poor draft can take longer than writing fresh. The gain is real, but it rides on the language pair and the quality of the machine output. In plain throughput, a human translator manages roughly 2,000 words a day, and an editor working from a decent machine draft manages about three times that.
Say a team needs a 40-page handbook in six languages by Friday. That gap decides the whole thing. Raw MT hands it back in seconds but risks the fluent errors above. Full human gives you native quality, though not by Friday. Getting it back in hours instead of weeks while keeping a human check is the reason the hybrid model caught on. If your content spans long documents and media, the same logic runs through document translation at scale and through dubbing and subtitling, where a person still directs tone and timing.
Each approach is risky somewhere different, so tie the risk to the content rather than to a slogan. Raw MT's exposure is the confident error. That's fine for an internal chat thread and unacceptable on a medication label or a signed contract, as the arXiv findings show. Human translation's exposure is operational, since the cost and timeline stop making sense at 500,000 words a month. AI + human carries a quieter risk, an editor who signs off on a weak draft instead of fixing it.
So "which is best" isn't really the question. The question is what a mistake costs you on this specific piece of content. A product description and a privacy policy live in the same company and belong in different lanes. The native-speaker check settles it. If a native reader can tell the text was translated, or catches an error, you bought less human than you needed. For anything carrying your brand voice or legal exposure, that resonance isn't optional, and machine-only won't reach it.
Match the approach to the content and the stakes, and if you're honest, most companies end up needing all three. Post-editing is already the default. Adoption across language companies climbed from 26% in 2022 to about 46% in 2024, and roughly 45% of them now use it on at least half their projects (Nimdzi Insights, 2025). It stopped being the compromise option a while ago.
That last lane is where Rian sits. The workflow drafts with a machine and reviews with a human across 60+ languages, so a six-language release comes back the same day rather than next month. You can add human review on the files that carry real risk, skip it on the ones that don't, and pay per use instead of committing to one mode for everything. Our customer stories show how that splits out across formats and industries. No platform removes the choice above, but once AI + human is your answer, you want a workflow built for that split.
For low-stakes, high-volume text, often yes. For anything customer-facing or regulated, no. A 2024 arXiv study found about 36% of Google Translate's mistranslations in sensitive content were critical (arXiv, 2024). The trouble with raw MT is that its fluency hides its errors, so review still matters.
Usually. Language companies price post-editing at roughly 30% to 70% of full human rates, and a 2024 Forrester study for DeepL reported 345% ROI over three years from adopting machine translation (DeepL / Forrester, 2024). The savings come from not paying a person to redo work a machine now drafts well.
Yes, and you probably should. One release might push its interface strings through raw MT, its help docs through post-editing, and its legal terms to a full human linguist. Sorting content by what a mistake would cost is the core of the job.
The numbers suggest otherwise. As machine use grew, post-editing adoption rose from 26% to 46% in two years (Nimdzi Insights, 2025). The work moved from typing to judging. Machines draft, and people decide what reads as native, which is not a job that disappears.
Through the MQM framework, updated to MQM 2.0 in 2024, which counts errors and weights them by severity instead of handing out a vague grade (MQM Council / AMTA, 2024). It lets you judge machine and human output on the same terms.
Human wins accuracy. Machine wins cost. AI + human wins speed. Machine and AI + human win scale. On risk-adjusted fit, and overall for most business content, AI + human wins.
Machine translation takes cost and raw speed, and human takes accuracy and nuance. AI + human takes the dimension most buyers weigh heaviest: quality they can trust, delivered fast enough to matter, at a price that scales. Pick raw MT when a mistake is cheap, full human when a mistake is expensive, and AI + human for the wide space between, which is where most of your content lives.
If that middle lane is yours, try Rian's AI + human workflow on a real file, watch where a human review changes the output, and decide which of your content needs it.