Human translation, MTPE, or raw AI output: Which quality tier does your content actually need?

One of the most expensive mistakes a business makes when commissioning translation is choosing the wrong quality tier. Commissioning full human translation for internal FAQs that only need to be understood (not published) wastes a budget that could fund higher-stakes work. Sending a regulatory submission through machine translation without proper human review exposes an organization to rejection, liability, and the cost of starting over.

The question is rarely "should we use AI or humans?" That framing is outdated and unhelpful. The real question is: which content type requires which level of human expertise, and what does the cost of getting it wrong actually look like?

According to a Nimdzi survey, adoption of MTPE has jumped from 26% in 2022 to nearly 46% by late 2024, reflecting a genuine shift in how organizations approach translation workflows. But adoption growth does not mean universal applicability. The surge in MTPE use has also produced a surge in misapplication – content that needed full human expertise routed through post-editing workflows designed for technical documentation, and raw AI output published in contexts that required human certification.

This guide maps every major content type to the tier that actually fits it. Tomedes has applied these criteria across thousands of projects since 2007, across legal, medical, technical, marketing, and e-learning content in more than 270 languages. The framework below reflects that accumulated experience.

Table of Contents

• What are the three translation quality tiers?

• What are the three translation quality tiers?

• What is the difference between full MTPE and light MTPE?

• What does each tier actually cost?

• Which content types require full human translation?

• Which content types are well-suited to MTPE?

• When is raw AI output acceptable without human review?

• What factors determine the right tier beyond content type?

• How does a pre-translation check improve tier decisions?

• What does a real mixed-tier project look like?

• FAQs

What are the three translation quality tiers?

There are three distinct tiers in professional translation workflows. Each has a defined place; the differences between them are not primarily about quality in isolation but about the level of human expertise applied, the risk tolerance appropriate for the content, and the intended audience.

• Tier 1: Full human translation (ISO 17100:2015) 

A certified human linguist translates the content from scratch, applying subject-matter expertise, cultural judgment, and stylistic precision. A second linguist reviews the output. This tier meets the requirements of ISO 17100:2015 (the international standard for professional translation services) and produces work that is publication-ready, legally defensible, and culturally accurate. It is appropriate wherever the content will be read by external audiences, carries legal weight, or requires the translator to exercise independent professional judgment.

• Tier 2: Machine translation post-editing / MTPE (ISO 18587:2017) 

A machine translation engine produces a draft, which a qualified human post-editor then reviews and corrects. ISO 18587:2017 (the international standard for MTPE) defines two sub-tiers: full post-editing (output suitable for publication, meeting the same quality bar as human translation) and light post-editing (output suitable for internal or informational use, where minor stylistic imperfections are acceptable). MTPE is appropriate for high-volume, repetitive, or time-sensitive content where cost and speed are primary drivers and the content does not require original creative or legal judgment.

• Tier 3: Raw AI / machine translation output (no human review) 

Machine translation output published or used without any human review. This is appropriate only in narrowly defined scenarios: internal-use comprehension (gisting), real-time communication monitoring, or preliminary research on foreign-language documents. It is not appropriate for any external publication, any legally significant document, any patient-facing or safety-critical content, or any content where brand voice or cultural resonance matters.

The critical principle is this: the tier is a function of the content's risk profile and audience, not of budget alone. Choosing a lower tier to save money on a high-risk document does not reduce cost – it transfers cost to rework, liability, or reputational damage downstream.

What is the difference between full MTPE and light MTPE?

Full and light MTPE are often treated as interchangeable but they are not. The distinction matters when specifying a project.

Full post-editing produces output that meets the same quality standard as human translation. The post-editor corrects all errors (linguistic, terminological, stylistic, and cultural) and the result is publication-ready. This is the appropriate sub-tier for any MTPE content that will be read by external audiences: customer-facing documentation, product user guides, e-learning modules, and software interfaces.

Light post-editing produces output that is intelligible and factually accurate, but does not guarantee stylistic polish or tonal consistency. The post-editor addresses critical errors and mistranslations but does not refine the text to read as naturally as human translation. This is appropriate for internal content, temporary use, research gisting, or situations where the organization needs to understand the content – not publish it.

A common error is commissioning light MTPE for content that needs full MTPE. This happens most often with technical documentation that is technically internal but is forwarded to clients, or with e-learning content that is treated as infrastructure but is read by learners as finished product. The Tomedes project management process flags this at intake – the intended end use of the translated content determines the sub-tier, not the content type alone.

What does each tier actually cost?

Cost comparisons between tiers are widely cited but frequently misleading because they ignore the cost of errors. The honest framing is not "how much does each tier cost?" but "what is the total cost of the project, including the cost of any mistakes the tier permits?"

Indicative rates (industry range, 2025–2026):

The math on tier selection changes sharply when the cost of error is included. A legal contract translated at the raw AI tier to save $500 in per-word fees can generate tens of thousands of dollars in costs if a mistranslated clause is discovered after signature. A pharmaceutical label translated at light MTPE rather than full human translation can trigger a regulatory rejection that delays a product launch by months.

Which content types require full human translation?

Full human translation is the only professionally defensible tier for the following content categories. No MTPE workflow (however well-managed) provides the same level of accountability, subject-matter judgment, or legal validity for these content types.

• Legal documents 

Contracts, court filings, affidavits, patents, immigration documents, regulatory submissions, and terms and conditions all require full human translation by a linguist with legal domain expertise. A meaning error in a contract clause can alter obligations, trigger disputes, or render an agreement unenforceable. Certified translation (which requires the translator to attest personally to the accuracy of the translation) is a legal requirement for many of these documents and cannot be issued under an MTPE workflow that lacks full post-editor accountability.

• Medical and clinical documents 

Patient-facing medical content (informed consent forms, discharge instructions, pharmaceutical labeling, and IFUs) require full human translation because the consequence of error is patient harm, not just document rejection. A 2021 study published in the Journal of General Internal Medicine found that machine translation of medical discharge instructions had a 40% error rate in certain languages. That figure should be sufficient to establish why no clinical content should move through a raw AI or light MTPE workflow. Full human translation under ISO 17100:2015, with mandatory human review at every stage, is the only appropriate tier for patient-facing medical content. Regulatory submissions (FDA, EMA, and equivalents) require ISO 18587:2017-certified MTPE at minimum, and human translation for all primary clinical sections.

• Marketing, brand, and creative content 

Marketing copy, campaign headlines, brand narratives, and transcreated advertising require the original creative judgment that no post-editing process can replicate. The goal of this content is not accuracy – it is persuasion, emotional resonance, and cultural connection. A post-editor correcting machine output is constrained by what the machine produced; a human translator begins from the source intent and produces work that achieves it in the target language and culture. Marketing content for a new market entry that reads as machine-generated is not just imperfect, it actively undermines the brand impression it was supposed to create.

• Regulated financial content 

Annual reports, prospectuses, financial disclosures, and regulatory filings in financial services require full human translation by linguists with financial domain expertise. These documents carry legal liability for the issuing organization and must be accurate to the letter. MTPE is not appropriate as the primary tier, though it may be used for supporting documentation and internal financial reporting.

• Any content requiring certified translation 

Certification (the formal attestation by a qualified translator that the translation is accurate and complete) cannot be attached to a document translated by a machine without human assumption of professional accountability. Certified translation requires full human translation under ISO 17100:2015. If an institution requires certified translation, that requirement itself dictates the tier.

Which content types are well-suited to MTPE?

MTPE is not a compromise – for the right content types, it is the most intelligent workflow available. It delivers the consistency and speed advantages of machine translation while retaining the human quality gate that eliminates the significant errors raw AI output cannot catch.

The content types best suited to full MTPE are:

• Technical documentation 

User manuals, product specifications, engineering documentation, software help content, and API documentation are ideal for full MTPE. This content is typically high in repetition (where machine translation excels), low in creative ambiguity, and moderately high in terminology specificity (where a post-editor's domain expertise adds value). Technical manuals, product datasheets, support documentation, and knowledge base content tend to repeat structures and terminology – machine translation engines handle this type of text well, especially when supplied with translation memories and client glossaries, and MTPE can reduce turnaround times significantly without compromising quality.

• E-learning content 

Structured e-learning modules with consistent terminology, quiz-style content, and instructional text are well-suited to full MTPE. The repetitive structure of learning content makes MT output more reliable, and a post-editor with subject-matter knowledge can efficiently verify accuracy and adapt culturally sensitive examples. Volume discounts from MTPE savings can be reinvested in additional language coverage, reaching more learners in more markets at the same budget.

• Software localization (UI strings) 

Application interfaces, error messages, and system notifications are high-volume, short-segment content where MT performs well and full MTPE produces consistent, accurate results. The primary post-editing task for UI content is terminology consistency and character-length constraints, tasks well-suited to structured post-editing workflows.

• E-commerce product descriptions (standard) 

Standard product descriptions (dimensions, materials, features, care instructions) are suitable for full MTPE. They require accuracy but not originality, and the repetitive nature of product catalog content makes MT output more reliable. Brand-voice-sensitive descriptions (flagship product narratives, campaign-tied product copy) should move to Tier 1.

Light MTPE is appropriate for:

• Internal reports and operational documents for internal comprehension only

• Knowledge base articles used by internal support teams

• High-volume background research documents that need to be understood, not published

• Inter-team communications across language barriers

Light MTPE should never be the tier for content that will be read by customers, partners, patients, or regulators – regardless of how "internal" the content appears at the time of commissioning.

When is raw AI output acceptable without human review?

Raw AI output (machine translation published or used without any human review) is appropriate in a narrow and specific set of scenarios.

• Gisting and comprehension 

When the goal is to understand the general meaning of a foreign-language document (not to publish a translation), raw AI output is sufficient. A procurement team reviewing foreign supplier contracts in draft stages, a research team scanning foreign-language academic papers for relevance, or a customer service team understanding the substance of an inbound message in another language can all use raw AI output for this purpose. The content never leaves the internal context in which it was generated.

• Real-time multilingual monitoring 

Social media monitoring, news scanning, and customer sentiment analysis at scale require raw AI output. The volume is too high for human review, the purpose is pattern detection rather than precise understanding, and no output is published or acted upon directly without a human decision layer above it.

• Preliminary research on large document sets 

Due diligence processes that require reviewing large volumes of foreign-language documents to identify which ones warrant professional translation are a legitimate use case for raw AI output – as a filtering mechanism, not as a deliverable.

• What raw AI output is not acceptable for 

Includes any content that will be published externally, shared with a customer or partner, presented to a regulator, used in legal proceedings, or relied upon for any decision where accuracy materially affects the outcome. The absence of human review means the absence of professional accountability. No quality guarantee can be attached to raw AI output.

What factors determine the right tier beyond content type?

Content type is the primary factor in tier selection, but not the only one. Four additional variables should be assessed for every project.

• Language pair 

Machine translation quality varies significantly by language pair. English-Spanish, English-French, English-German, and English-Portuguese produce substantially better MT output than English-Arabic, English-Japanese, English-Korean, or English-Swahili – because the latter pairs have fewer digital resources and different grammatical structures that current MT models handle less reliably. For languages with fewer digital resources, the quality of the automatic draft degrades, and the post-editing effort increases to the point where it can match or exceed the cost of human translation. For lower-resource language pairs, the economics of MTPE deteriorate and full human translation often produces better value.

• Audience and visibility 

The more people who will read the content, and the more consequential their reaction to it, the higher the tier required. An internal document read by ten employees is a different risk profile from a customer-facing website page read by tens of thousands. Visibility and audience affect both the probability and the magnitude of harm from a translation error.

• Regulatory and legal environment 

Some jurisdictions and sectors impose mandatory requirements on translation quality that override cost or speed preferences. USCIS requires certified translation for immigration documents. FDA and EMA regulate translation of pharmaceutical labeling. Courts specify translation standards for admissible evidence. In these contexts, the regulatory requirement determines the tier, and no business case for a lower tier can override a legal obligation.

• Source text quality 

Machine translation quality is directly limited by source text quality. Ambiguous phrasing, inconsistent terminology, long and complex sentence structures, and idiomatic language in the source document all degrade MT output – and increase the post-editing effort required to correct it. A source document that requires significant revision before translation will perform poorly at the MTPE tier regardless of the language pair or MT engine used.

How does a pre-translation check improve tier decisions?

The single most practical step a buyer can take before commissioning translation is to assess the source text before selecting a tier. Most tier-selection errors are not caused by ignorance of the framework, they are caused by inaccurate assumptions about the quality of the source material and the MT engine's likely performance on it.

Tomedes' Pre-Translation Toolkit was designed specifically to address this problem. Before a project is costed or assigned, the toolkit analyzes the source text and surfaces:

• Potential quality concerns - grammar issues, punctuation errors, cultural sensitivity flags, and inconsistent terminology that will degrade MT output if left uncorrected

• Text complexity assessment - a readability and complexity score that indicates how well the content will perform at the MTPE tier versus requiring full human translation

• Key terms for review - industry-specific or non-translatable terms that require consistency decisions before translation begins

This analysis directly informs tier selection. A source text flagged as high-complexity with significant ambiguity is a signal that full human translation will be more efficient than MTPE on that document, because the post-editing effort on a poor MT draft of complex text can exceed the effort of translating from scratch. A source text assessed as low-complexity, highly repetitive, and terminologically consistent is a clear candidate for MTPE.

Running a pre-translation check takes minutes and eliminates the most common source of tier-selection error: commissioning MTPE for content the MT engine cannot reliably handle, then discovering the problem during post-editing when changing the tier is costly.

What does a real mixed-tier project look like?

Most real-world translation projects contain content that appropriately belongs in more than one tier. Treating a complex project as a single-tier problem is itself a source of unnecessary cost and quality risk.

Consider a software company launching a product in Germany, France, Japan, and Brazil. The project includes:

A single per-word rate across this entire project would either overpay for internal documents or underpay (and underprotect) for legal and brand content. The right approach is a tiered quote that applies the appropriate workflow to each content category.

Tomedes provides this kind of tiered project structure as standard. A dedicated project manager assesses each content category at intake, applies the appropriate tier, and manages the workflow – so buyers are not left making these decisions in isolation. Every deliverable, regardless of tier, is covered by the 1-Year Quality Guarantee.

FAQs

Q: Is MTPE the same quality as human translation?
A: Full MTPE, when conducted to ISO 18587:2017 standards, produces output that is equivalent in accuracy and publication-readiness to human translation – for content types where machine translation performs reliably. The distinction is not in the output quality of a well-executed full MTPE project; it is in the content types for which MTPE is appropriate. Legal content, creative content, and content in lower-resource language pairs are not suitable for MTPE regardless of how rigorously the post-editing is conducted, because the MT draft introduces structural and cultural errors that post-editing cannot fully correct without effectively retranslating the text.

Q: How much cheaper is MTPE than human translation?
A: For content that is genuinely suited to MTPE (high-volume, repetitive, technically structured), full MTPE typically costs 30-50% less per word than full human translation, with faster turnaround. Light MTPE can reduce costs further for internal-use content. However, these savings only hold when the content type is well-matched to the tier. When MTPE is applied to content that performs poorly at machine translation (complex creative copy, ambiguous legal language, lower-resource language pairs), the post-editing effort increases until the cost advantage narrows or disappears entirely.

Q: Can I use machine translation for a certified translation?
A: No. Certified translation requires a qualified human translator to personally attest to the accuracy and completeness of the translation. This attestation represents professional accountability that cannot be delegated to a machine. USCIS, courts, academic institutions, and most official bodies that require certified translation will not accept machine-translated documents – with or without a certification statement attached. Certified translation requires full human translation under ISO 17100:2015.

Q: Does the language pair affect which tier is appropriate?
A: Yes, significantly. High-resource language pairs (English-Spanish, English-French, English-German, English-Portuguese) produce more reliable MT output and are better candidates for MTPE. Lower-resource pairs (many African languages, indigenous languages, less widely digitized Asian languages) produce MT output that requires substantially more post-editing – to the point where the cost and time advantages of MTPE over human translation are reduced or eliminated. For any language pair outside the top tier of MT coverage, tier selection should be validated against the expected MT output quality before project start.

Q: What happens if we choose the wrong tier?
A: The most common consequences of tier mismatch are: rework costs (the translation must be redone at the correct tier after errors are discovered); rejection (a regulatory body, court, or institution rejects a document that did not meet the required translation standard); reputational damage (published content that reads poorly in the target language or contains cultural errors); and in the worst cases, legal liability or patient safety incidents. The cost of rework consistently exceeds the initial saving from the lower tier. The most cost-effective tier selection happens at the outset, informed by a pre-translation assessment of the source text and content type.

Q: How does Tomedes determine which tier to recommend for a project?
A: Every Tomedes project begins with a brief and a source text assessment. The dedicated project manager reviews the content type, end use, audience, language pair, and source text quality to identify complexity and quality concerns. Based on that assessment, the project manager recommends a tier and explains the rationale. For mixed projects, the recommendation covers each content category separately. Buyers are not expected to arrive with a tier decision already made; the project manager's role is to guide that decision based on the content, not to apply a default workflow regardless of fit.