There is a line on the quote in front of you that reads something like “AI-assisted translation with human review.” It is priced well under the human line. Further down the same document, a logo block advertises the vendor’s ISO 17100 certification. You are a procurement lead, or a compliance officer, or simply the person who got handed the language budget, and the implication seems clear: same standard, same protection, less money.
Here is the thing most buyers have wrong, and it is checkable in about ninety seconds. ISO 17100 does not cover post-edited machine translation at all. Raw machine output plus post-editing sits expressly outside that standard’s scope; it is governed by a different standard entirely, ISO 18587. A vendor can hold a genuine ISO 17100 certificate and still be selling you a service that certificate says nothing about. Nobody is necessarily lying. The certificate is simply pointing at the wrong line item.
What post-editing actually is, and what the standard requires
ISO 18587:2017, Translation services — Post-editing of machine translation output — Requirements, is the only international standard that addresses machine translation directly. It defines the process, sets out the competences and qualifications a post-editor must hold, and structures the work across pre-production, production, and post-production stages, including final verification and feedback into the engine.
Two details in that standard matter enormously to a buyer, and almost no vendor marketing mentions either.
First: the requirements are for full post-editing only. ISO 18587 provides the requirements against which a conformity claim can be made for full post-editing — output intended to be comparable to the product of human translation. Light post-editing, the cheaper tier that aims only at comprehensible text without trying to read like human translation, is described in an annex. It is not the normative requirement set. So “we post-edit to ISO 18587” means full post-editing or it means very little; if the quote is priced at light-touch rates, ask which one you are actually buying and get the answer into the scope of work.
Second: ISO 18587 applies only to content processed by a machine translation system. It is a process standard for a specific production method. It is not a quality grade you can sprinkle onto a project, and it does not make post-edited output interchangeable with human translation for regulatory purposes.
The certification claim that does not transfer
ISO 17100:2015 sets requirements for human translation services. Its defining production control is that a translation must be revised by a person other than the translator — the translator’s own self-check does not satisfy it. That independent revision step is a large part of why the standard carries weight with regulated buyers.
And that standard explicitly excludes the use of raw machine translation output with post-editing from its scope. This is not an interpretation; it is how the scope is drawn, and it is why ISO 18587 had to be written as a separate document in the first place.
The practical consequence for your procurement file: an ISO 17100 certificate cannot be the quality evidence for an MTPE line item. If your specification invoked ISO 17100 and the delivered work was post-edited machine output, you did not get what you specified, and you may not discover it until something goes wrong downstream. Put the question in writing at bid stage: for each deliverable, which standard governs, and is the human step revision by a second linguist or post-editing by one?
A vendor that cannot tell you, per deliverable, whether you are buying ISO 17100 human translation or ISO 18587 post-editing is not running either process deliberately.
Not sure which process your current work is actually going through? The fastest tell is the revision record. Human translation under ISO 17100 produces evidence of a second linguist’s revision; post-editing does not, because there is only one human in the chain. Ask for the record, not the certificate.
Where MTPE fits, and where it does not
The honest answer is not “machine translation is bad.” Used on the right content, post-editing is a legitimate, standards-backed production method that gets far more material into more languages than a human-only budget ever could. The discipline is in matching the process to the risk the content carries.
| Content type | Appropriate process | Why |
|---|---|---|
| Internal knowledge base, IT and HR help articles | MTPE | High volume, short shelf life, low consequence if phrasing is imperfect |
| Support ticket and chat archives, user reviews | MTPE, or raw MT for gisting | Comprehension is the whole goal; nobody relies on it as a record |
| Large product catalogs, spec sheets, inventory copy | MTPE, full | Repetitive and template-driven; terminology control does the heavy lifting |
| Marketing, brand, and campaign copy | Human translation or transcreation | Machines optimize for fidelity, not persuasion; tone does not survive the round trip |
| Certified translations for USCIS, courts, licensing boards | Human translation | The certification statement is a personal attestation — see below |
| Contracts, filings, evidence, discovery for the record | Human translation, ISO 17100 | Ambiguity is adversarial; one mistranslated clause is the whole exposure |
| Informed consent, patient instructions, clinical trial documents | Human translation with documented review | Comprehension by the subject is a regulatory requirement, not a quality preference |
| Medical device IFUs, labels, safety and warning content | Human translation | Harm is physical, and the language obligation is set by regulation |
| Public notices where language access is legally required | Human translation | The obligation is meaningful access; unreliable output does not discharge it |
Notice what separates the two halves of that table. It is not difficulty, and it is not length. It is who bears the cost of an error. When the reader absorbs a bad sentence as mild confusion, MTPE is a reasonable trade. When the reader absorbs it as a wrong dose, a waived right, a signature on something they did not understand, or a filing that gets rejected, the trade does not exist at any price.
Certified translation is a different thing entirely
This is where “AI-assisted” quotes cross from cost-saving into misrepresentation. A certified translation is not a quality tier. It is a signed statement, attached to the translation, in which a competent person attests that the translation is complete and accurate and that they are competent to translate from the source language. For immigration filings, that requirement sits at 8 CFR 103.2(b)(3): any foreign-language document submitted must be accompanied by a full English translation the translator has certified as complete and accurate, together with the translator’s certification that they are competent to translate from the foreign language into English.
Read that as what it is — a personal attestation about work a person did. Raw machine output has nobody standing behind it. If a certification statement is affixed to output that no one with the required competence actually translated or fully revised, the statement is not merely weak; it is inaccurate on its face, and the person whose name is on it carries that. We covered the downstream version of this problem in what USCIS, courts, and employers actually accept, and rejection is the mild outcome. Re-filing after a rejection costs more than doing it correctly would have.
The same logic runs through regulated life sciences work. Where a regulation requires that information be understandable to a patient or trial subject, the sponsor has to be able to evidence how that was achieved. “An engine produced it and someone skimmed it” is not a defensible answer to a monitor, an IRB, or a notified body.
A three-step way to get this right
1. Sort your content by consequence, not by volume
Take an inventory and put every content type into one of three buckets: record (certified, regulated, contractual, safety), reliance (customer-facing, brand, instructional), and reference (internal, archival, gisting). The record bucket is human translation, full stop. The reference bucket is where MTPE pays for itself. The reliance bucket is the judgment call, and it usually turns on whether the content is instructional or persuasive.
2. Name the standard per deliverable in your scope of work
Not per contract — per deliverable. Write “ISO 17100, including independent revision by a second qualified linguist” against the record bucket and “ISO 18587 full post-editing” against the reference bucket. This single change makes bids comparable and makes substitution visible instead of invisible.
3. Require disclosure of the production method, in writing
Ask every vendor to state, per deliverable, whether machine translation was used in production and at what post-editing level. A provider running a deliberate process answers in a sentence. A broker reselling to the lowest bidder cannot, because they do not know.
What it costs you to get this wrong in either direction
Over-specify and the damage is bounded and financial: you paid human rates for an internal wiki nobody will read twice. That is a budget conversation.
Under-specify and the damage is unbounded. A rejected immigration filing restarts a clock a family is living inside. A mistranslated consent form can invalidate enrollment and force re-consent across a site. An inaccurate warning in device instructions is a safety event before it is a compliance event. A contract clause that says something different in the other language gets discovered by the party it favors. None of that is recovered by the savings on the line item that caused it — and in every case, the record will show a process the organization chose.
Why work with Taika
Language Access Hub, powered by Taika Translations, is a veteran-owned (VOSB), SAM-registered, GSA- and NASPO ValuePoint-contracted language services provider. We run both processes deliberately, and we tell you which one you are buying: certified human translation by ATA-certified linguists with independent revision for anything that becomes a record, and disciplined machine translation post-editing where volume and risk profile genuinely support it. For organizations running both at once, our managed translation services keep terminology, translation memory, and routing rules consistent across the two, so content does not quietly drift into the wrong pipeline.
We will also tell you when post-editing is the wrong answer for a given deliverable. That conversation is cheaper to have before the work than after the rejection.