Download Translation Options Compared (PDF)
Agency vs In-House vs Machine Translation
There is no single right answer to “how should we handle translation?” The right approach depends on what you’re translating, who will read it, what the consequences of errors are, and how much volume you produce.
This guide compares the three main options — professional translation agency, in-house or bilingual staff, and machine translation — honestly. Each has situations where it makes sense and situations where it doesn’t. The goal is not to argue for one approach but to give you a framework for deciding which approach fits which content.
We are a translation agency. We have an obvious interest. But we also know that agencies are not the right answer for every translation need, and pretending otherwise would undermine the credibility of everything else in this guide.
Option 1: Professional Translation Agency
How it works
You send source documentation to the agency. A project manager assesses the scope, assigns qualified native-speaking translators with relevant subject matter expertise, and manages the process through translation, review, formatting, and delivery.
Professional agencies use translation memory technology to leverage previous translations, maintain terminology databases for consistency, and apply a second-linguist review before delivery.
True costs
Visible costs:
– Per-word rate (typically £0.08-£0.15 per word for European languages, higher for rare languages or specialist content)
– Project management (usually included in the per-word rate)
– Desktop publishing or formatting (typically quoted separately)
– Certification (if required, quoted separately)
Hidden benefits that reduce cost over time:
– Translation memory leverage — repeated content from previous projects doesn’t need retranslating. By the third or fourth project, 20-40% of content may already be in the translation memory
– Terminology consistency — maintained glossaries mean fewer corrections and less review time
– Reduced internal time — the agency manages the process, not your team
Typical cost for a 10,000-word technical manual into 3 European languages: £3,000-£5,000 including translation, review, and project management. Subsequent editions of the same manual cost significantly less due to translation memory leverage.
When it makes sense
- Technical, regulatory, or safety-critical content where accuracy has legal or commercial consequences
- Customer-facing documentation that represents your brand
- Content in multiple languages, where coordination and consistency across languages matters
- Ongoing translation programmes where translation memory builds value over time
- Content types requiring specialist domain knowledge (medical, legal, engineering, financial)
- When you need DTP, formatting, or certification capabilities
- When you lack internal language capability for quality review
When it might not
- Low-stakes internal communications where approximate understanding is sufficient
- Very small volumes where the overhead of briefing exceeds the value
- Content that changes so rapidly that formal translation cannot keep pace (though MT + post-editing may work here)
- Situations where immediate, real-time translation is needed (interpreting, not translation)
Option 2: In-House or Bilingual Staff
How it works
A member of your team who speaks the target language translates the content. This might be a bilingual colleague, a native speaker in a local office, a sales agent in the target market, or a member of staff whose role is not translation but who has the language skills.
True costs
Visible costs:
– Apparently free — the person is already on the payroll
Hidden costs that are rarely tracked:
– Opportunity cost — time spent translating is time not spent on their primary role. A bilingual engineer translating a manual is not engineering
– No quality assurance — there is typically no second review by another qualified linguist. Errors go undetected until a customer or regulator finds them
– No translation memory — the work doesn’t build any reusable asset. The next similar document starts from scratch
– No terminology management — different staff members make different terminology choices, creating inconsistency across your documentation
– Speed — translation is not their primary skill. A bilingual staff member typically translates at a fraction of the speed of a professional translator
– Availability — they have other responsibilities. Translation is fitted around their main role, creating delays and bottlenecks
– Risk of burnout or resentment — being asked to translate on top of a full-time role is a burden, not a benefit
Typical true cost for a 10,000-word technical manual into 3 languages: If a bilingual engineer earning £55,000/year spends 15 hours per language on translation (at roughly half the speed of a professional), the labour cost alone is approximately £1,200 per language — plus the value of the engineering work they didn’t do during those hours. No translation memory is created. No terminology database is maintained. Quality is unverified.
When it makes sense
- Internal communications where approximate accuracy is acceptable
- Very short, non-critical content (a brief internal email, a quick product description for internal use)
- Situations where the bilingual staff member genuinely has translation expertise (some do — but bilingualism and translation skill are not the same thing)
- Initial review of translations produced by a professional agency (in-market review by a knowledgeable native speaker is valuable)
When it doesn’t
- Customer-facing documentation of any kind
- Regulatory or compliance documentation
- Safety-critical content
- Large volumes or ongoing programmes
- Content where terminology consistency matters
- Any situation where the consequences of error are significant
- When the bilingual staff member’s primary role is more valuable than the translation they’re being asked to do
Option 3: Machine Translation
How it works
Machine translation (MT) uses AI to translate text automatically. Current systems (neural machine translation) produce significantly better output than the statistical and rule-based systems of a decade ago.
Machine translation exists on a spectrum:
Raw machine translation: Text goes in, translation comes out. No human involvement. Free or very cheap.
Machine translation with post-editing (MTPE): Machine produces the first draft, a human linguist reviews and corrects it. Faster and cheaper than full human translation, but quality depends on the content type and the post-editor’s skill.
AI-assisted translation: Professional translators use MT suggestions as a starting point, accepting, rejecting, or modifying each segment. Similar to MTPE but with the translator in control throughout.
True costs
Visible costs:
– Raw MT: free (Google Translate, DeepL free tier) or very low cost (API pricing)
– MTPE: typically 40-60% of full human translation rates
– AI-assisted: similar to full human translation, with potential efficiency gains
Hidden costs:
– Review time — someone must review MT output for accuracy. For technical content, this review takes significant time and requires subject matter knowledge
– Error risk — MT systems make confident-sounding errors. They don’t flag uncertainty. A measurement translated incorrectly looks identical to one translated correctly
– No domain adaptation — generic MT systems don’t know your product, your terminology, or your industry conventions (though custom-trained models can improve this)
– Inconsistency — MT may translate the same term differently in different contexts. Without translation memory, there is no mechanism for enforcing consistency
– Liability — who is responsible if a machine-translated safety instruction causes harm? The question is not theoretical
Typical cost for a 10,000-word technical manual into 3 languages with MTPE: £1,500-£2,500 — cheaper than full human translation, but with variable quality depending on content complexity.
When it makes sense
- Internal comprehension — understanding the gist of a foreign-language document for internal purposes
- High-volume, repetitive, non-critical content — product descriptions, catalogue entries, user reviews
- First draft for human review — where speed matters more than perfection and a qualified post-editor will review everything
- Content with a short lifespan — material that will be replaced or updated within weeks
- Languages where professional translators are scarce and expensive
When it doesn’t
- Safety-critical documentation of any kind
- Regulatory submissions or compliance documentation
- Legal contracts, terms and conditions, or patent documentation
- Customer-facing content that represents your brand
- Content where a confident-sounding error could have consequences (medical, pharmaceutical, engineering)
- Marketing content that requires cultural adaptation, not just linguistic conversion
- Content requiring certified or sworn translation
The quality question
Machine translation quality varies enormously depending on the language pair, the content type, and the specific MT engine. Some observations:
- MT performs best on well-structured, formal content in common European languages
- MT struggles with ambiguity, idiom, cultural context, and creative language
- MT is weakest for rare languages, specialist terminology, and content requiring cultural adaptation
- MT quality is improving rapidly — but the gap between “impressive” and “reliable enough for safety-critical content” remains significant
- The best use of MT is as a tool for professional translators, not as a replacement for them
Decision Framework
Choosing by content type
| Content Type | Recommended Approach | Why |
|---|---|---|
| Safety-critical documentation | Professional agency | Legal liability, accuracy non-negotiable |
| Regulatory submissions | Professional agency | Compliance requirement, notified body scrutiny |
| Customer-facing manuals | Professional agency | Brand representation, user safety |
| Marketing and brand content | Professional agency (transcreation) | Cultural adaptation required |
| Legal documents | Professional agency | Legal precision required |
| Technical datasheets | Professional agency or MTPE | Depends on complexity and stakes |
| Internal communications | In-house or raw MT | Low stakes, approximate accuracy sufficient |
| Product catalogue descriptions | MTPE or agency | Depends on volume and visibility |
| Support knowledge base | MTPE | High volume, updateable, reviewable |
| Understanding foreign documents | Raw MT | Comprehension, not publication |
Questions to ask yourself
- What happens if there’s an error? If the answer involves regulatory consequences, safety risk, financial loss, or brand damage — invest in quality.
- Who will read this? Customers, regulators, and partners deserve professional quality. Internal teams can work with approximations.
- Will this content be reused? If yes, professional translation with translation memory builds a compounding asset. MT builds nothing.
- How many languages? Multiple languages amplify both the value of professional translation (consistency) and the risk of MT (inconsistent errors across languages).
- What’s the true comparison? Don’t compare the per-word cost of MT against an agency. Compare the total cost including review time, error correction, risk, and the value of translation memory.
A Blended Approach
Most organisations benefit from a blended approach:
- Professional agency for high-stakes, customer-facing, regulatory, and safety-critical content
- MTPE for high-volume, moderate-stakes content where speed matters
- Raw MT for internal comprehension and low-stakes communication
- In-house review of professionally translated content by knowledgeable native speakers
The key is matching the approach to the content — not applying one approach to everything.
Key Takeaways
- Translation is not a commodity — different approaches suit different content types
- The cheapest option per word is rarely the cheapest option when total costs are considered
- Translation memory is a compounding asset that only professional translation builds
- Machine translation is a powerful tool but not a replacement for human expertise on high-stakes content
- The right question is not “which is cheapest?” but “what are the consequences of getting this wrong?”
- A blended approach — matching investment to risk — is usually the most cost-effective strategy
Need Help?
Download more resources: bubblestranslation.com
Questions? info@bubblestranslation.com | 0870 777 7750
Bubbles Translation Services — Honest advice on translation options since 2003
