Choosing between AI translation, machine translation with human editing, and professional human translation is easier when you separate speed, risk, cost, and brand sensitivity. This guide provides a repeatable way to assess a project, estimate its total effort, and select a workflow that fits the content rather than relying on a blanket rule about human versus machine translation.
Overview
AI translation tools can produce a useful first version quickly, but output quality depends on the language pair, subject matter, source-text quality, and review process. Human translation remains important when wording must be precise, persuasive, legally reliable, culturally appropriate, or consistent with a defined brand voice. Between these options is a practical middle path: machine translation followed by structured human editing.
Think of the three approaches as workflow levels:
- AI-only translation: Best suited to low-risk understanding, internal notes, rough drafts, and high-volume material where minor imperfections are acceptable.
- AI or machine translation with human editing: Useful for business content, support articles, product information, and recurring document translation that needs review but does not require every sentence to be newly translated by a specialist.
- Professional human translation: Appropriate for legal, medical, safety-critical, highly creative, regulated, or brand-defining content, especially when terminology and cultural nuance carry significant consequences.
The right choice is not determined by word count alone. A short contract clause may deserve more control than a long internal report. Likewise, a simple product description may become high-risk if it affects search visibility, customer trust, or a purchase decision.
For website translation and multilingual SEO, the process should also include keyword research, search intent checks, metadata review, and local-language editing. A translated page can be grammatically correct and still fail to match how people search in the target market. See the website translation SEO checklist and guide to translating a website without hurting SEO for related planning considerations.
How to estimate
Use a simple decision score before selecting a workflow. Rate each factor from 1 to 5, where 1 is low and 5 is high:
- Risk: Could an error create legal, financial, medical, safety, or reputational problems?
- Brand sensitivity: Does the text need a distinctive voice, careful persuasion, or culturally specific wording?
- Audience visibility: Will customers, regulators, partners, or the public read the final version?
- Terminology difficulty: Does the content contain technical terms, product names, abbreviations, or an approved glossary?
- Deadline pressure: Would a slower process create a meaningful operational cost?
- Review capacity: Is a qualified bilingual reviewer available to check the result?
Add the first four scores to create a quality-control score. A low score may support AI-only translation for an appropriate use case. A middle score usually points to AI-assisted translation with human review. A high score calls for professional human translation or specialist review. The exact thresholds are not universal; the value of the exercise is making assumptions visible and documenting why a workflow was chosen.
You can also estimate effort with this formula:
Total effort = source words × translation effort per word + review effort + formatting or implementation effort
Use your own internal rates or quotes rather than assuming that a tool's subscription cost represents the full project cost. A low-cost first draft may still require terminology research, fact checking, layout repair, SEO adaptation, or a second review. For recurring work, record actual time and error types from the first few projects, then use those observations to improve the estimate.
Inputs and assumptions
A useful estimate begins with a clear project brief. Capture the following inputs:
- Language pair and direction: State the source and target language, including regional variants where relevant.
- Content type: Separate web pages, email campaigns, product interfaces, subtitles, contracts, support content, and internal documents. Each has different expectations.
- Volume and repetition: Count words, pages, screens, or minutes of audio. Repeated segments may reduce editing effort when translation memory or a controlled glossary is available.
- Source quality: Identify unclear sentences, inconsistent terminology, formatting problems, jokes, idioms, and text that has not yet been approved in the source language.
- Required output: Decide whether you need a rough translation, publication-ready copy, a formatted document, subtitles, voice output, or content prepared for a content management system.
- Review standard: Define who checks meaning, terminology, grammar, tone, numbers, links, layout, and cultural suitability.
- Data handling: Establish which content may be placed into an AI translation tool and which material must remain within an approved environment. This is a project-specific governance decision, not something to leave implicit.
Before translating, clean the source text. Remove duplicate spaces, resolve contradictory terms, expand ambiguous abbreviations, and separate instructions from copy intended for readers. A text cleaner tool or compare-text-differences utility can support this preparation, but neither replaces editorial judgment.
For a consistent AI-assisted translation workflow, use a terminology list containing preferred terms, prohibited alternatives, product names, measurements, capitalization rules, and examples of the desired tone. Give the tool a narrow task: translate only the approved text, preserve placeholders, do not alter URLs or code, and flag uncertain passages instead of guessing. Then have a reviewer check meaning against the source before polishing style.
Worked examples
Example 1: Internal research notes
A team needs to understand a long article in another language before deciding whether it is relevant. The audience is internal, the material will not be published, and speed matters. The risk and brand-sensitivity scores are low. AI-only translation may be reasonable, provided that names, numbers, quotations, and important conclusions are checked against the source. The output should be labeled as a working translation rather than treated as final copy.
Example 2: A support knowledge base
A company is translating recurring help-center articles. The content is public and terminology must remain consistent, but most articles use established language. A suitable estimate includes AI or machine translation, a bilingual edit, terminology review, link checks, and a sample of user-facing pages tested in context. Repeated headings and approved phrases can make the workflow more efficient over time. The explanation of translation memory covers how reusable segments can affect recurring projects.
Example 3: A legal or medical document
A short document contains obligations, warnings, diagnoses, or instructions. Even if the word count is small, the quality-control score is high. A professional human translation with relevant subject expertise and a defined review process is the safer default. AI may assist with terminology searches or an internal comparison, but it should not be treated as the final authority. For more planning guidance, use the guide to choosing translation services for specialized documents.
Example 4: A marketing landing page
A landing page must sound natural, motivate action, and reflect local expectations. Direct translation may preserve meaning while weakening the headline, call to action, or cultural relevance. A practical workflow is to use AI for a draft, give a human editor the brand brief and target audience, and then review keywords, layout, claims, and conversion elements in the target language. For email campaigns, subject lines and calls to action deserve the same attention as the body copy; the email translation best practices guide provides a useful checklist.
In each example, the estimate changes when the purpose changes. Recalculate the work if the content moves from internal use to publication, if a new market is added, or if the reviewer discovers recurring errors in the AI output.
When to recalculate
Review your translation decision whenever one of the project inputs changes. Recalculate when the word count grows, a new language or regional variant is introduced, the deadline moves forward, or the target audience becomes more public or commercially important. Also reassess after a tool changes its output behavior, your approved glossary is updated, or a reviewer reports repeated problems with names, numbers, tone, or terminology.
Pricing assumptions should be revisited when internal rates, vendor quotes, tool plans, review time, or formatting requirements change. Do not compare only the visible translation cost. Compare the estimated total effort, including corrections, publishing, quality checks, and the cost of fixing an error after release.
At the end of each project, save a short review record:
- What workflow was used?
- How many passages required meaningful correction?
- Which error types occurred most often?
- Did the glossary and instructions work?
- Was the final effort lower or higher than estimated?
- Would the same content type justify the same workflow next time?
This feedback turns human versus AI translation from a one-time opinion into an operating method. Use AI translation tools where they provide a suitable first pass, add human control where meaning or reputation requires it, and reserve full professional translation for content whose consequences justify the additional expertise. The best workflow is the one whose review standard matches the real risk of the text.
