AI Translation vs Human Translators: What Documents Really Cost

Yes, by a wide margin on price. Machine translating a 2,000-word document through an AI model costs a few cents, while human translation at Gengo's published rates starts at $180 (Standard, $0.09 a word) or $360 (Advanced, $0.18 a word). The fair comparison is AI plus a bilingual reviewer, which usually lands in between.

The decision turns on what a mistake would cost. An internal note can go out machine-translated. A price list, booking terms or a safety briefing that customers rely on needs someone who knows both languages to check it, and anything legal or official may need a certified translator whatever the price. So price each kind of document separately rather than choosing one route for everything.

Follow me on Instagram@sagnikteaches

Prices per 2,000-word document, route by route

A 2,000-word document is roughly a four-page itinerary, a set of booking terms or a long proposal. Here's what it costs by each route, using prices published by the vendors in September 2026:

Connect on LinkedInSagnik Bhattacharya

The review line is an estimate built on two assumptions you should test: a reviewer checking about 1,000 words an hour, and Upwork's 2026 rates guide, which lists translators at $15 to $25 an hour and proofreaders at $18 to $35. Time your own reviewer on a sample page before you rely on it.

Subscribe on YouTube@codingliquids

Why machine translation costs cents: the token sum

AI models charge per token, and a useful rule of thumb is that a million tokens is roughly 750,000 English words. A 2,000-word document is therefore about 2,700 tokens of input. The translated output often takes more tokens than the English, because many other languages and scripts break into more tokens per word, so budget generously: say 4,000 output tokens.

• gpt-6-luna ($0.10 per million input tokens, $0.50 per million output): about $0.0003 in plus $0.002 out, well under a cent.
• Claude Haiku 4.5 ($1 in, $5 out): about $0.003 plus $0.02, roughly 2 cents.
• Claude Sonnet 5 ($2 in, $10 out): about $0.005 plus $0.04, roughly 5 cents.

Even a year of heavy use rarely reaches a noticeable bill at these rates. That's why the machine cost almost never decides the question; the review time does. API access is billed separately from chat subscriptions, so a ChatGPT Plus or Claude Pro plan doesn't cover automated translation through an API, and the API needs its own account and spending limit.

DeepL's plans and limits in September 2026

DeepL is a dedicated translation service rather than a general chat assistant, and it keeps a document's formatting when it translates a file. Its plans changed in 2026, so older price comparisons are out of date. From its own pricing page, in its USD pricing, all billed annually (annual billing saves about 16%; prices vary by country, so check the figure shown for yours):

Watch the file limits more than the character limits. A business translating 30 brochures a month would run out of files on the Individual plan in the first week, although 30 brochures come nowhere near 300,000 characters.

DeepL's API has changed too. According to its help pages, the old API Free and API Pro plans can no longer be bought. New developers get a Developer plan capped at 1,000,000 characters in total, which doesn't reset, and then move to a Growth plan: a fixed monthly or yearly price with an included character allowance, usage above it billed on top, a monthly cost-control limit, and texts deleted immediately after translation. Check the current Growth price on DeepL's site before building anything on it.

What a bilingual review adds, and what it costs

Post-editing, the translation industry's name for reviewing and correcting a machine draft, is where most of the quality comes from. A reviewer who knows both languages reads the draft against the original and fixes meaning, tone and terminology. The tutorial on a translator's machine translation post-editing workflow describes the craft from the reviewer's side. For a business, the practical questions are who reviews and how much of each document.

Not every page needs the same scrutiny. A sensible split for customer-facing documents:

• Read every word: prices, dates, times, cancellation terms, safety and allergy information, anything a customer acts on.
• Read closely but faster: descriptive text such as tour descriptions or room details.
• Spot-check: repeated templates once they've been approved, such as a confirmation email that only changes names and dates.

Here's what a reviewer's notes on one 900-word AI-drafted itinerary might look like, which is a fair picture of how much a good draft still needs:

1. Times and pickup points: all correct.
2. Cancellation line: "at least 48 hours before" had become "within 48 hours of". Corrected.
3. Tour names: "Castle Ridge Trail" translated literally on page 2 despite the glossary. Restored to English.
4. Dietary note: "vegetarian options on request" had become "vegetarian menu included". Corrected.
5. Form of address: switched between formal and informal "you" across paragraphs. Standardised to the formal form the company uses.
6. Time taken: 55 minutes.

Two of those five issues would have caused real trouble with customers, and neither was visible to anyone who doesn't read the language.

A reviewer can also be a member of staff who speaks the language well, provided they have time and a checklist. Checking AI translations before customers see them has a fuller checking routine.

A tour operator's year of translation, costed three ways

Imagine a tour operator running 24 small-group tours, an illustrative business, that translates its customer documents into Spanish and Italian. Per language, the year's writing looks like this: 24 tour descriptions of 300 words (7,200 words), 24 itineraries of 900 words (21,600), booking terms and cancellation policy (2,500), a safety briefing (1,200) and ten pre-trip email templates of 150 words (1,500). That's 34,000 words per language, 68,000 in total.

Route 1: all human, Gengo Advanced. 68,000 words at $0.18 is $12,240. At Gengo Standard ($0.09) it would be $6,120, though Gengo recommends Advanced for content that needs more accuracy, which itineraries and terms do.

Route 2: AI drafts, bilingual review, humans for the high-stakes pieces. The booking terms and safety briefing (3,700 words per language, 7,400 in total) go to a human translator at the Advanced rate: $1,332. The other 60,600 words are drafted by AI for a few dollars and reviewed at about 1,000 words an hour: roughly 61 hours at $25, or $1,525. Add a $20-a-month AI subscription for the staff member preparing drafts ($240 a year) and the total is about $3,100.

Route 3: AI only. About $240 a year for the subscription, or less through an API. Cheapest by far, and the riskiest, because nobody who reads the language checks what customers see.

Route 2 costs about a quarter of Route 1 and keeps a qualified person on everything that carries legal or safety weight. The operator used Route 3 only for internal notes to its local guides, where a clumsy phrase costs nothing. For how the translated itineraries fit into the wider booking process, see how small tour operators use AI for bookings and itineraries.

When the AI saving is too small to matter

The sums change at low volumes. Consider a boutique hotel that translates one document a year: a 1,500-word guest information pack, into Spanish, Italian and Japanese. At Gengo's Advanced rate that's 4,500 words at $0.18, or $810 a year. The AI-plus-review route might cost roughly $80 to $160 in reviewer time at the rates above, but finding reviewers for three languages is its own small project, and the pack includes fire procedures and allergy information that deserve a professional. Saving $650 or so a year isn't worth the risk or the organising. The hotel pays a translator for the pack and uses AI for the things that change weekly: replies to reviews and messages from guests who write in another language, which a member of staff checks before sending.

A rough rule: the AI route earns its keep when you translate often, when documents repeat with small changes, or when volume makes human rates painful. For a single annual document, the translator's fee is usually money well spent.

Getting human quotes you can compare

Translators quote per word, per hour or per page, and include different things. Send every one the same short brief so the quotes line up:

Documents: 24 itineraries (21,600 words), English to Spanish
Subject: guided walking tours; some safety wording
Files: Word documents; please keep formatting
Glossary: attached (tour names stay in English)
Deadline: 3 weeks
Please quote:
  - price per word, or total, and whether proofreading by a second
    linguist is included
  - price to review and correct AI drafts instead (post-editing)
  - rush fees, formatting charges, and how revisions are handled

Asking for a post-editing price alongside a full translation price shows you the true gap between the two routes for your documents, from someone who has actually looked at them.

Where machine translation goes wrong in customer documents

Modern AI translation is fluent, which is exactly why its mistakes are easy to miss. The errors that cost money are rarely grammar; they're meaning. Realistic examples, shown here by translating the AI's version back into English:

• A flipped condition. Source: "Full refund if you cancel at least 48 hours before departure." The draft, read back: "Full refund if you cancel within 48 hours of departure." A reader would reasonably expect a refund the day before.
• A time that moved. "Meet at 7.30 for a 7.45 departure" came back with both times written in a way local readers interpret as evening, because of the format conventions in the target language.
• A name translated. A tour called "The Old Harbour Walk" was translated word for word in one itinerary and left in English in another, so customers thought they were two different tours.
• A softened warning. "Sturdy footwear is essential" became the equivalent of "comfortable shoes are recommended".

None of these would be caught by someone who doesn't read the language, and all of them would be caught by a reviewer reading prices, times, conditions and warnings word by word. Whether machine translation is good enough for client-facing documents goes further into where the line sits.

A glossary prompt and a back-translation check

If you translate with a chat assistant or an API, a fixed prompt with a glossary prevents most consistency problems:

Translate the document below from English into Spanish for
customers of a small tour company.
Rules:
- Keep these names in English: The Old Harbour Walk, Castle Ridge Trail
- Always translate "meeting point" as [your approved term]
- Keep all prices, dates and times exactly as written, in 24-hour
  format, and do not convert currencies
- Keep warnings as strong as the original: "must" stays "must"
- If a sentence is ambiguous, keep the ambiguity and mark it [CHECK]
Return only the translation.
[paste document, with customer names removed]

Then run a second, separate check that doesn't depend on reading the target language: ask a fresh chat to translate the result back into English, and compare the two English versions for conditions, numbers and warnings. In an illustrative run on a set of booking terms, the back-translation showed the flipped 48-hour condition above and a missing sentence about deposits. It didn't catch a slightly too casual tone, which only the bilingual reviewer noticed. A back-translation is a cheap first screen, not a substitute for a person who reads the language. For keeping terminology consistent over many documents, how translators use AI to keep terminology consistent has techniques you can borrow.

Documents that should stay with a human translator

• Contracts and terms you'll enforce. The wording is the product; a qualified translator, and possibly a lawyer, should own it.
• Anything official. Documents for courts, licensing bodies or public authorities often need certified translation.
• Safety, health and allergy information. The cost of a softened warning is far higher than a translator's fee.
• Slogans and headline marketing. Adapting a slogan so it works in another culture, sometimes called transcreation, is a creative job, not a literal one.
• Anything regulated in your sector. Financial promotions, medical claims and similar wording carry their own rules in each market.

Hidden costs on both sides of the comparison

AI translation looks almost free until you count the rest: reviewer time, reformatting when a PDF's layout breaks (a reason to use a tool that translates files with formatting kept), building and maintaining the glossary, plan limits on files, and the data question. Before pasting customer details into any tool, check its terms; business plans from the main AI vendors don't train on your content by default, but consumer plans rely on a switch in the privacy settings. Removing names first is simpler. Whether it's safe to put client documents into AI translation tools covers that in detail.

Human translation has hidden costs too: rush fees, charges for formatting or layout work, rounds of revisions, and turnaround measured in days. Gengo, for example, says it has no minimum spend or set-up fees for website orders, but others do, so ask. And whichever route you choose, someone in the business still has to brief the translator or the AI, answer questions and publish the result. Count that time on both sides and the comparison becomes honest.

Translation costs: follow-up questions

Do I need a certified translator for contracts and official documents?

Often, yes. Courts, government offices and some institutions ask for certified or sworn translations, which only a qualified translator can provide, and no AI tool can certify its own output. For contracts you sign in another language, a qualified translator or a lawyer who works in both languages is the safe route. Ask the receiving organisation what it requires before you order anything.

Can AI keep our brand and product names consistent in every language?

Yes, if you tell it to. Give the tool a glossary of names and terms that must stay in English or must always be translated a certain way, and include it in every prompt or in the tool's glossary feature. Then spot-check a few pages, because models sometimes translate a name that looks like an ordinary word.

Is it safe to paste guest or customer details into a translation tool?

Only on a plan whose terms you've checked. Business plans from the main AI vendors don't train on your content by default, and OpenAI's API doesn't use data for training by default; consumer plans have a model-training switch in their privacy settings. Better still, remove names and contact details before translating, and add them back afterwards.

Further reads

• How to Offer Multilingual Customer Support With AI Translation — Using AI translation for live customer messages.
• How Small Translation Agencies Build an AI-Assisted Workflow — How translation agencies combine AI and human work.
• What Does the ChatGPT API Cost for a Business Automation? — Price API usage for your own document volumes.
• How to Stop AI Tools Training on Your Business Data — Settings to check before translating business documents.
• How to Read AI Software Pricing: Seats, Credits, and Usage Fees — Seat, character and usage pricing decoded.
• When Not to Use AI in Your Business: 9 Tasks to Keep Human — Nine tasks a person should always decide, what AI can safely prepare for each, and a three-question score for any task that isn't on the list.
• Does Your Translation Business Need an AI Usage Policy? — Turn client restrictions, tool choices and reviewer responsibilities into a short policy your translators can use on each job.
• AI vs Human Transcription: Which Is Worth Paying For? — When AI transcription is good enough, when a person is worth $1.99 a minute, how to test accuracy on your own recordings, and the hybrid in between.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Gengo pricing page; Upwork hourly rates guide (2026); DeepL plan comparison and DeepL help article on API plans (checked September 2026); Anthropic and OpenAI API pricing; OpenAI and Anthropic plan pricing.