MTPE for High-Volume Repetitive Jobs: A Scaling Playbook

The Hidden Cost of Scaling Human Translation¶

Your client calls Monday morning: “We need 500,000 words translated into 12 languages by Friday. Budget is tight.” You do the math. 500K words × 200 words/hour human translation = 2,500 translator-hours. At full-time rates, that’s 62 translators, full-time, for a week. Or you slip the deadline by 6 weeks.

This is where the translation industry has lived for decades: pick speed or sanity, not both.

According to Slator, human-only translation delivers 2,000-3,000 words daily per translator. MTPE (machine translation post-editing) flips that to roughly 7,000 words per day—while cutting cost per word from $0.20-0.30 down to $0.08-0.15. That same 500K-word job drops from 6 weeks to 10 days, and margins don’t collapse.

The playbook exists. Thousands of agencies run it. But most are leaving 40-60% of the efficiency gain on the table because they’re not routing segments intelligently, enforcing terminology, or pre-editing source text.

Here’s how to build a high-volume MTPE workflow that actually works.

What Is MTPE? (And Why High-Volume Changes Everything)¶

MTPE is not “let the AI do everything and call it done.” It’s a hybrid workflow: neural machine translation produces a first draft, human post-editors refine it to publishable quality, and the whole cycle repeats with intelligence—not brute force.

The key shift for high-volume work is differentiated editing: not all segments need the same effort. A repetitive product description that’s already 95% correct doesn’t need the same depth as a legal clause that came out 60% right. When you’re moving 100,000+ words per month, that difference is the difference between profit and burnout.

According to research from Phrase, the modern MTPE stack combines:

• Machine translation (neural models like DeepL, Google, or domain-adapted engines)
• Quality estimation (automated tools that score each segment’s confidence)
• Intelligent routing (high-quality segments skip editing, medium ones get light PE, low ones get full PE or retranslation)
• Terminology enforcement (glossaries and translation memory prevent inconsistency)
• Final review (human eyes on edge cases and style)

Without this orchestration, MTPE is just expensive raw MT. With it, you get 3x throughput, 50% cost savings, and quality that clients trust.

Which Content Is Worth the MTPE Setup?¶

MTPE doesn’t work equally well on everything. Before you invest in glossaries and routing, know your content type.

MTPE-Friendly Content (High ROI): - Product catalogs and e-commerce descriptions (repetitive phrasing, consistent terminology) - Technical documentation and manuals (structured, factual, domain-specific terms) - Knowledge bases and support articles (FAQ-style, recurring questions) - User-generated content and reviews (volume over perfection) - Release notes and changelog updates (boilerplate + new features) - Form translations and templates (heavy repetition, low variance)

These content types see edit distances (the % of text actually changed by post-editors) below 30%, meaning the MT engine is already doing 70% of the work correctly. Your post-editors spend time on polish, not reconstruction.

MTPE-Difficult Content (Low ROI): - Brand campaigns and creative copy (tone, idiom, and cultural nuance break MT) - Legal contracts and compliance documents (one wrong term costs millions—full PE mandatory) - Poetry and literary translation (idiom density too high) - Highly technical research papers (jargon-specific, context-dependent) - Formal interpersonal communication (HR policies, apologies, relationship management)

These see edit distances above 40-50%, which means post-editors are doing almost as much work as translators. MTPE saves you ~30% instead of 70%. Still worthwhile for 500K-word batches, but don’t expect the 172% productivity boost.

The Rule: If your content is repetitive and structured, MTPE is a force multiplier. If it’s creative or high-stakes, MTPE is a speed bump. Know which you have.

Step 1: Build Your Glossary Before Running MT¶

The single most expensive mistake in high-volume MTPE is running machine translation without a glossary.

As documented by Crowdin, inconsistent terminology is the #1 source of post-editing rework. Your client uses “workflow” in English, but one MT output says “workflow”, another says “working process”, and a third says “procedure”. The post-editor catches it and re-edits. The next batch does the same. You’ve just added 10% to your PE time.

This becomes catastrophic at scale. Imagine a 100,000-word product catalog where “dashboard”, “panel”, and “control center” all get used for the same UI element. By the time post-editors finish, they’ve spent an extra 20-30 hours fixing terminology alone. That’s $1,500-2,000 in wasted labor on a single project.

A glossary is a list of approved term pairs: English term + target-language equivalent(s), usually with context and notes. It’s not just a translation list—it’s a mandate that ensures the MT engine uses the same term every time, reducing downstream post-editing rework dramatically. Examples:

For repetitive content, a small glossary (50-200 terms) cuts post-editing effort by up to 43%, according to Phrase research. For product catalogs, a medium glossary (200-500 terms) can cut costs by up to 50% because the same terms reappear across hundreds of product descriptions.

Real case: A tech company translating a 250,000-word SaaS platform into German without a glossary paid for 500+ hours of PE (at $50/hr = $25,000). The same project with a pre-built 300-term glossary took 320 hours. The glossary cost $2,000 to build, but saved $12,000 in PE labor—a 6:1 ROI on the first project alone, and the glossary was reusable for all future versions and related products.

How to build it:

1. Extract recurring terms. Use terminology tools like Sketch Engine, or your CAT tool’s built-in term extractor. Look for terms that appear 10+ times in the source. Those are your high-impact candidates.

2. Have domain experts translate 50-100 key terms. This doesn’t need to be in-depth; 10-20 minutes with a subject-matter expert per term is enough. Include context: Is “filter” a UI element, a data concept, or a physical part? The engine needs to know.

3. Feed the glossary to your MT engine. Most modern systems support this: Trados has termbase integration, memoQ reads XML termbases, Phrase has built-in glossary management, Smartcat supports custom terminology, DeepL API accepts custom terminology in the request. If your engine doesn’t support glossaries natively, you’re using the wrong tool for high-volume MTPE.

4. Review post-edits from the first 5,000 words. After the first batch comes back from PE, scan for terminology misses. Did the glossary catch everything? Did the MT engine respect it, or did it output a term not in the glossary? Build a feedback loop: missing terms get added to the glossary before the next batch runs.

Most glossaries reach 80% maturity after the first 20K words. After that, the ROI comes in fast. By 100K words, you’ve amortized the glossary cost and are harvesting pure productivity gains.

Step 2: Use Quality Estimation to Route Segments, Not Edit Them All¶

This is where most agencies leave 40% of their efficiency gain on the table.

Quality estimation (QE) is an automated tool that scores each MT segment on a 0-100 scale. It doesn’t replace post-editing; it tells you which segments need what depth of PE. The math is simple:

• Score >80%: High-quality MT. Skip editing entirely (no human touch needed). These segments are “publish-ready.”
• Score 50-80%: Medium-quality. Light post-editing only (fix errors, don’t improve style). You’re correcting glitches, not rewriting.
• Score <50%: Low-quality. Full post-editing or retranslation. These segments need a linguist’s attention.

According to NIMDZI research, intelligent QE routing reduces post-editor workload by up to 40% without quality loss. High-quality segments skip editing (40-50% of your volume gets zero human touch). Medium-quality segments get light PE (target: 1,000 words/hour, not 500). This matters because the productivity gap between light and full PE is enormous—full PE at 500 wph vs light PE at 1,200 wph is a 2.4x difference in throughput on the same editor time investment.

Real example: 100,000-word project. - 45,000 words (45%) score >80% → skip editing → 0 hours → $0 cost. - 35,000 words (35%) score 50-80% → light PE at 1,000 wph → 35 hours → $1,750 cost (at $50/hr). - 20,000 words (20%) score <50% → full PE at 500 wph → 40 hours → $2,000 cost. - Total: 75 hours of human work, $3,750 cost instead of 200 hours / $10,000 if you edited everything uniformly.

That’s a 63% time savings and a 62% cost reduction on the PE portion alone. For a client’s $20,000 budget (full human translation), you’re now delivering the same output for $6,400 total (MT + PE + review).

Without QE, you’d likely assign all 100K to full PE (100 hours at 1,000 wph average), burning $5,000 and missing the deadline.

How to set up QE in practice:

QE tools are built into most modern TMS/CAT platforms (Phrase, Smartcat, Trados, memoQ with plugins). Some are free or low-cost (OpenKiwi, TER, BLEU). Start with your platform’s built-in QE, then run the first 5,000-word batch through it to calibrate your thresholds.

Set thresholds based on your content type and quality bar. For product catalogs (where terminology matters more than nuance), you can safely skip >85% quality. For legal/medical (where even small errors cost money), you might set >95% as “skip.” For technical docs, >80% is usually safe. The key is that you’re not guessing—you’re using data to route, not hunches.

After the first PE cycle, review the QE scores against what post-editors actually found. Did segments scoring 75% actually need PE, or were they publish-ready? Adjust your threshold. QE gets better when you tune it to your engine, language pair, and content type.

Step 3: Pre-Edit Source Text and Standardize Phrasing¶

Garbage in, garbage out. But garbage in source text means extra work in post-editing.

According to academic research, pre-editing (cleaning and standardizing source text before MT) cuts post-editing effort by up to 43%. This is one of the highest-ROI activities in MTPE and is often overlooked because it happens before MT—so the impact isn’t immediately visible.

Pre-editing includes:

• Fixing source grammar and clarity. Unclear phrasing, typos, and run-on sentences break MT worse than any language pair. MT engines can’t infer what you meant—they can only parse what’s written. Fix them first. A single typo in a product name can cascade across all 500 instances of that product in your catalog.

• Standardizing terminology in the source. If your source uses “workflow” 200 times and “working process” 50 times, unify to “workflow” before MT. Inconsistent source terminology forces post-editors to choose which translation is “correct”—extra cognitive load, slower PE.

• Breaking complex sentences. Long, nested sentences with multiple clauses produce worse MT. “We discovered that because the customer needed X, and the team implemented Y, the result was Z” → split into three sentences: “The customer needed X. Our team implemented Y. The result was Z.” Simpler syntax = better MT, lower edit distance.

• Removing redundancy. MTPE thrives on repetition, but over-repetition in source (saying the same thing 5 times in slightly different ways) just multiplies rework. If you say “You can use this tool to automate your workflow” and later “This tool automates your workflow,” consolidate to one phrasing.

• Adding context notes. For ambiguous terms, add XML tags or comments in the source: We process your <context>data privacy regulations</context> according to... tells MT “this ‘data’ is about privacy law, not general information.” Some MT engines respect these tags; others ignore them. But it never hurts to add clarity.

Pre-editing is 10-20% overhead (2-5 hours per 100K words, or $500-1,500 in labor), but it cuts PE time by nearly half. On high-volume jobs, that trade-off is almost always worth it. A tech company with 500K-word annual translation volume saves 1,000+ PE hours annually (~$50,000) by investing 50 hours (~$2,500) in pre-editing. That’s a 20:1 ROI.

Most agencies outsource pre-editing to junior linguists or source-language editors (often cheaper than post-editors). The workflow: source subject matter expert → pre-editor (fixes and standardizes) → MT engine → post-editor (polishes). It’s an extra step, but it halves the PE workload.

Step 4: Pick Your MT Engine (and Don’t Switch Mid-Project)¶

Your choice of MT engine affects everything: edit distance, terminology consistency, handling of complex syntax, and—critically—native integration with your CAT tool.

According to BLEU benchmarks from 2026, there’s no universal winner:

• DeepL: Best for European languages (EN→DE/FR/ES BLEU 60-65). Native integrations with memoQ, Trados (via connectors), Phrase. Consistent terminology handling, fewer hallucinations. Best choice for high-volume German/French/Spanish translation.
• Google Translate: Broadest language coverage (130+ languages). Good for lower-volume or rarer pairs. Weaker on European languages than DeepL, but solid fallback for non-Western languages. Native in many TMS platforms.
• ChatGPT / Claude: Excel on CJK (Chinese/Japanese/Korean), context-dependent content, and multilingual content. Slower (not ideal for 100K+ word batches), expensive per API call, but higher quality for creative or nuanced text. Best for smaller high-stakes projects.
• Domain-Adapted Custom MT: If you feed a custom engine with 100K+ words of your client’s translation memory, it outperforms generic engines on that specific client’s terminology and style (BLEU +5-10 points). Only worth it for retainer clients with 500K+ words/year.

The critical rule for high-volume: Pick one engine per language pair and stick with it for the entire job and all future jobs with that client. Switching MT engines mid-project doubles post-editing effort because post-editors are relearning the engine’s quirks and terminology patterns.

According to productivity research, post-editors hit peak productivity after 3 months with the same engine. If you switch, they start from zero.

Step 5: Optimize Post-Editing Workflow and Differentiate by Difficulty¶

Not all post-edits are equal.

Light post-editing (LPE): Fix errors, improve readability, but don’t rewrite for style or fluency. - Target: 1,000-1,200 words per hour. - Budget: 4,000-5,000 words per day per editor. - Best for: Product descriptions, FAQs, technical specs at 70%+ quality.

Full post-editing (FPE): Fix errors and improve fluency; make output sound native. - Target: 500-700 words per hour. - Budget: 2,500-3,500 words per day per editor. - Best for: Customer-facing content, brand material, anything <60% quality.

For high-volume projects, most of your work will be LPE. The 45% of segments that quality estimation flags as high-quality get zero editing (full skip). The next 35% are routed to LPE (your workhorse). Only 20% go to full PE.

This tri-level approach—skip / light / full—is what separates profitable high-volume MTPE operations from those that burn out editors and miss deadlines.

Step 6: Build a Terminology Enforcement and Review Gate¶

Once post-edits come back, they’re not done. A final review layer catches:

• Glossary compliance: Did the post-editor maintain terminology consistency? Automated term checking tools (built into Trados, memoQ, Phrase) flag misses.
• Style consistency: Does tone match your brand guidelines? This is the domain of style-checking tools (some built-in, some third-party).
• Number and name accuracy: Edit distance doesn’t catch “€25 EUR” changed to “€250 EUR” or “John” to “Joan”. Automated QA flags these.
• Missing segments or obvious errors: Spot checks of 10% of the PE output, especially segments marked as <50% quality before PE.

This review gate is usually 2-5 hours per 100,000 words (2-5% overhead). For high-volume, it’s automated mostly: run QA checkers, then 5-10% manual sampling of risky segments.

If you’re running 1M+ words per month, this gate is the difference between “repeatable process” and “chaos and client complaints.”

The Economics: Pricing and ROI¶

Let’s do the math on a real 100,000-word project, EN→DE, product catalog (MTPE-ideal content).

Human-Only Translation¶

• Rate: $0.25/word (fair market for German)
• Cost: 100,000 words × $0.25 = $25,000
• Time: 100,000 words ÷ 250 wph = 400 hours = 10 weeks (5 translators)

MTPE Workflow with Quality Estimation and Glossary¶

• MT engine cost: $0 (assuming you use DeepL or Google with credit)
• 45,000 words (45%) skip editing: 0 cost
• 35,000 words (35%) light PE at $0.07/word: $2,450
• 20,000 words (20%) full PE at $0.12/word: $2,400
• Subtotal PE: $4,850
• Glossary setup (first project only): $500
• Pre-editing (2,000 words worth of source cleanup): $300
• QE tool / TMS add-on: $0-200/month (amortized)
• Final review (5 hours at $50/hr): $250
• Total Cost: $6,400 (or $5,900 if amortized across 3 projects)

Savings: $25,000 − $6,400 = $18,600 (74% cost reduction)

Time: 100,000 words ÷ 1,200 wph avg = 83 hours = 2 weeks (1-2 editors)

The Client Angle¶

You have two ways to price this:

1. Pass the savings: Offer $0.10/word instead of $0.25. Client saves 60%, you make $10,000 on the project (vs. $25,000 if you were doing it full human). They’re thrilled because they got human-quality output at 2/5 the cost and half the turnaround time.

2. Margin expansion: Keep your price at $0.20/word ($20,000), cut your cost to $6,400, and pocket $13,600 (68% margin). Clients who value speed see it as a win (half the turnaround). Clients who budgeted at $0.25 see it as a win (10% savings). You win either way.

Most agencies split the difference: price at $0.15/word (40% discount from human-only), cost is $6,400, profit is $9,600, and the client saves $10,000 compared to their $25,000 budget.

Common Failures at Scale (And How to Avoid Them)¶

Failure 1: Running Raw MT Without Glossaries¶

Symptom: Post-editors are making the same corrections over and over. “Dashboard” is translated three different ways in one 20,000-word batch. Root: No glossary fed to MT engine. Fix: Mandatory glossary review before any MT run. Even a 50-term glossary cuts rework by 30%.

Failure 2: Editing Everything Uniformly¶

Symptom: Burnout. Editors are marking time, work quality drifts, deadlines slip. Root: No quality estimation routing. All segments go to full PE, even 90%-correct ones. Fix: Deploy QE scoring and assign light PE to >60% quality segments. Throughput jumps 50%, margins survive.

Failure 3: Switching MT Engines Mid-Project¶

Symptom: Editors suddenly are slower and errors increase. “This MT is different from yesterday.” Root: You switched from DeepL to Google mid-job to save money, or tried a new engine. Fix: Lock in one engine per language pair for the full project and all related jobs. Don’t optimize engine per batch.

Failure 4: No Pre-Editing of Messy Source¶

Symptom: Edit distance is 50%+. You’re doing almost as much work as human translation. Root: Source text is unclear, terminology is chaotic, sentences are nested and long. Fix: Invest 2-5 hours in pre-editing per 100K words. Cuts PE effort by 43%. ROI is immediate.

Failure 5: Ignoring Terminology Feedback Loop¶

Symptom: Same wrong terms keep appearing in PE output. Root: Post-editors catch glossary misses (MT ignores a term, or the glossary is incomplete), but feedback isn’t fed back to the glossary. Fix: Weekly glossary reviews. Post-editors flag terms, you add them. Next batch uses the updated glossary. Glossary matures fast.

FAQ¶

Q: At what project size does MTPE break even vs. human translation? A: Projects above 10,000 words almost always save money with MTPE. A 100,000-word project at $0.20/word human translation ($20,000) drops to $10,000 with MTPE at $0.10/word—50% savings. The real gain is speed: the same 100K words that take 40 days with human-only takes 15 days with MTPE.

Q: What if our source text is terrible quality? A: Bad source breaks MTPE. Poor grammar, unclear terminology, and inconsistent style in the source create more post-editing work than starting fresh. Pre-editing (cleaning source text, enforcing terminology, standardizing phrasing) can cut PE effort by up to 43%. Start with source quality first.

Q: How do I route segments between light and full post-editing automatically? A: Quality estimation (QE) tools score each MT segment on a 0-100 scale. Set thresholds: scores >80% → no editing, 50-80% → light PE (fix errors only), <50% → full PE or retranslate. This intelligent routing cuts editor workload by up to 40% without quality loss.

Q: Do I need a translation memory for MTPE to work? A: Not required, but TM + glossary combo cuts post-editing effort by 43%. A well-fed TM ensures consistent terminology and captures approved phrases that reappear across jobs. For repetitive content (product catalogs, manuals), TM leverage can cut costs by up to 50%.

Q: Which MT engine is best for high-volume MTPE: DeepL, Google, or ChatGPT? A: No one winner. DeepL leads on European languages (EN→DE/FR/ES BLEU 60-65); ChatGPT/Claude excel on Asian languages and context-dependent content. For high-volume scaling, use the engine your CAT tool integrates natively with—Trados, memoQ, Phrase. Switching engines mid-project doubles PE work.

Q: What’s the most common MTPE mistake on high-volume jobs? A: Running MT raw without glossaries or quality estimation. Inconsistent terminology, poor segmentation, and editing everything uniformly (no routing) kills margins. Fix: glossary first, QE second, segmentation third.

Q: Can MTPE scale to 1M+ words monthly? A: Yes, if you fix these three things: (1) glossary-driven MT consistency, (2) quality estimation to skip high-quality segments, (3) light PE for medium-quality, full PE only for <50% quality. Agencies running 1M+ monthly use this exact stack. Without it, you’ll have burnout and quality drift.

Q: How do I price MTPE to clients without them rejecting it as inferior? A: Position as ‘faster, not cheaper.’ MTPE delivers 50% cost savings and 2-3x speed (same output in 1/3 the time). Clients buying on turnaround (not just price) see it as a win. Frame it: ‘Certified in 3 days instead of 2 weeks’ beats ‘$10K instead of $20K.’