The Production Economics of Certified Translation
What a certified page actually costs an agency to produce, where the margin leaks, and how MTPE benchmarks changed the arithmetic. A cost model for LSPs selling into immigration work.
Certified translation looks like a healthy line of business until you model it properly. A page sells for $30 to $55 in the B2B channel. The linguist cost looks small against that. And yet certified work is where a lot of language service providers discover their thinnest real margins.
The reason is that the sell price is per page and the cost base is not. This piece models where the money actually goes.
The unit mismatch
Certified translation is quoted per page โ conventionally 250 words or fewer. Almost every input cost behaves differently.
Linguist time scales with complexity, not page count. A 180-word birth certificate with two seals, a marginal annotation, and handwriting takes longer than a 250-word letter.
Project management scales with documents, not words. A ten-document client file with ten short certificates carries ten times the coordination of one long document at the same total word count.
Desktop publishing scales with layout. A one-page household register with a nested table can consume more time than five pages of prose.
Certification and QA scale per document. Each item needs its own certification statement, its own check against the source, its own file.
So a file of twenty short civil documents โ the typical immigration bundle โ is the worst possible shape for a per-page price. It maximises the per-document costs and minimises the word count you get paid for.
That is the structural problem. Agencies that lose money on certified work are usually not underpricing the words; they are underpricing the document count.
Where the margin actually leaks
Formatting rebuild. This is the largest hidden cost and the least visible on a timesheet. Reconstructing a table, a bilingual column layout, or a right-to-left form is not translation, but it is billed to the same task and consumes the same hour. On structured civil documents it routinely exceeds the linguistic work.
PM administration. Industry observers put project manager time lost to manual updates, chasing, and data entry at 10โ20% at midsize LSPs. On high-document-count immigration files, that share runs at the top of the range.
Rework. A rejected translation is not a partial loss. It is the original cost, plus the correction, plus the relationship damage, plus โ where the client is a law firm โ the possibility of losing an account over one Request for Evidence.
Scope creep on completeness. The reverse side nobody scanned. The grading legend that turned up late. The bank statement quoted as three pages that turned out to be eleven billable pages of transaction rows.
Rush surcharges you absorbed. Accepting a 24-hour turnaround at standard rates because the client is important.
The MTPE benchmark
Machine translation post-editing stopped being controversial some time ago. Adoption moved from roughly a quarter of agencies in 2022 to close to half by late 2024, and it has kept climbing. It is now the default production model for a large share of commercial work.
The published rate bands make the arithmetic clear:
| Model | Typical rate |
|---|---|
| Full human translation | $0.09โ$0.40 / word |
| MTPE | $0.05โ$0.15 / word |
| Certified, sold per page | $30โ$55 / page |
MTPE lands roughly 30โ50% below full human translation. That is the benchmark buyers increasingly assume, and procurement teams at corporate clients now ask about it directly.
Certified immigration work has lagged this shift, for a defensible reason: generic MT engines destroy the thing certified documents depend on. They flatten tables into prose, skip seal text because it is an image, mirror right-to-left columns without remapping them to their headers, and silently drop blank form fields. On a marketing brochure that is a quality problem. On a household register it is a rejected filing.
So the sector kept paying full human rates for work that is substantially mechanical โ because the available automation solved the wrong half of the problem.
Splitting the page into two costs
The useful move is to stop treating a certified page as one unit of work and start treating it as two.
The mechanical half. Reading the scan, reconstructing the layout, rendering every field, capturing seal text, transcribing identifiers, preserving the table. This is most of the elapsed time on structured civil documents and it requires accuracy and patience rather than professional judgment.
The accountable half. A qualified translator reading the output against the source, correcting it, applying terminology decisions, and signing a certification in their own name. This is what the receiving authority is actually buying, and it cannot be automated โ see ai-in-certified-translation-workflows.
Once the two are separated, they can be costed separately. The mechanical half is the part where input cost has moved. Layout-preserving AI drafting runs at a per-page cost measured in cents rather than dollars โ ZenTrans volume pricing reaches $0.70 per page, around $0.004 per word at a 250-word page. Against an MTPE input band of $0.05โ$0.15 per word, that is a different order of magnitude, and it is spent on the specific failure mode generic MT does not address.
The accountable half does not get cheaper. It gets shorter, because a reviewer working from a complete, correctly structured draft spends their time on judgment rather than transcription.
Modelling your own number
Take your last ten certified files and pull four figures:
- Documents per file, not pages. This predicts PM and certification load better than anything else.
- Elapsed linguist hours split between formatting reconstruction and linguistic review. Most agencies have never measured this separately and are surprised by the ratio.
- Rework rate โ percentage of files with a client-reported correction or an authority rejection.
- Realised rate per page after rush absorption and scope creep, not the rate card.
The gap between figure 4 and your rate card is your actual pricing problem. The ratio inside figure 2 tells you how much of your cost base is addressable.
Repricing, not just re-costing
Lower input cost is only half the opportunity. The other half is fixing the quotation model:
- Quote per document, or add a per-document component, so twenty short certificates are not priced as if they were one long file.
- Define "page" explicitly in writing โ 250 words or part thereof โ and count transaction-row documents honestly at quote time rather than absorbing them.
- Price rush as a real tier, not a favour.
- Separate DTP for documents requiring genuine layout reconstruction, or ensure your production method removes that cost rather than hiding it.
Falling input cost tends to get competed away immediately unless the commercial model changes with it. Agencies that pass the entire saving into the rate card end up with the same margin and a lower price point.
What does not change
The certification. It remains a named, qualified human attesting in their own name that they reviewed the translation and vouch for it. Canada's IRCC refuses machine-generated translations outright; USCIS, UKVI, and Australian Home Affairs each require an accountable human by a different route. The tool is a drafting instrument; the certifier is the product.
Any cost model that assumes the certifier away is not modelling certified translation.
Frequently asked questions
Is per-page pricing wrong for certified work? Not wrong, but incomplete. It underprices high-document-count, low-word-count files โ which is most immigration work. Add a per-document component.
Will clients accept an AI-assisted production model? Corporate and law firm buyers increasingly assume MTPE already. What they want confirmed is that a qualified human certifies the output and carries the accountability. Lead with that.
Where should we measure first? The split between formatting reconstruction and linguistic review. It is the figure most agencies have never separated and the one that most often surprises.
Does lower input cost mean lower prices? Only if you let it. The saving is worth more as margin or as capacity for higher-value work than as an across-the-board rate cut.
Certified translation is priced per page and costed per document. Model those separately, measure how much of your linguist hour goes to rebuilding layout rather than translating, and fix the quotation model at the same time you fix the input cost.
Want to see what a production draft costs on your own source files? Open ZenTrans and run a sample batch โ the first 5 pages are free, and volume pricing runs to $0.70 per page.