The client was a German medical-device maker, and the call came on a Tuesday: their distributor in three new markets would not accept the localized instructions-for-use unless the work was done "to ISO 17100." The project manager who took the call had been booking translation jobs for six years and had never once read the standard. She knew the number. Everyone in localization knows the number, the way you know a road sign you have driven past a thousand times without ever stopping to read the fine print. So she did what most people do: she promised the client it would be done to ISO 17100, hung up, and then went looking for what she had just promised. What she found was not a quality test you run at the end. It was a description of how a translation is supposed to be produced from the first minute to the last: who is allowed to touch it, what they must be qualified to do, and the one step almost no rushed pipeline actually performs, a second qualified linguist checking the first one's work against the source. This lesson is about that standard, what it actually requires in the room where the work happens, and why a document written for human translators became the silent foundation that the machine-translation standards stand on top of.

The Standard Everyone Cites and Few Read

Start with what ISO 17100 is, in the plainest possible terms, because the name carries a weight out of proportion to how many people have read it. ISO 17100, titled "Translation services: Requirements for translation services," is the international standard that defines what a professional, conformant translation service looks like as a process. It was first published in 2015, replacing an older European standard (EN 15038), and it remains the baseline reference that clients, language-service providers, and certifiers point to when they say a translation was done "properly." When a procurement department writes "ISO 17100-certified" into a tender, when an LSP puts the certification badge on its website, when a regulated client demands conformance, this is the document they all mean.

Before we go further, define the actors, because the standard is precise about them and the precision is the point. An LSP is a language-service provider, the company or individual that takes the client's content and returns the translated content; it can be a thousand-person agency or one freelancer working alone. A linguist is the human who works on the language: translating, revising, reviewing. Localization (l10n) is the broader practice of adapting content for a target market, of which translation is the core; the "10" is the number of letters between the l and the n. And the words this lesson turns on, translator, reviser, and the revision step itself, are defined by the standard in ways that working linguists use loosely and the standard uses exactly. We will pin them down as we go.

Here is the first thing that surprises people who finally read it. ISO 17100 says almost nothing about the words on the page. It does not tell you how to translate a sentence, what makes a good rendering, or how to score the output. It is not a quality metric. It is a process standard: a description of the steps, the people, and the records a translation service must put in place so that quality is produced reliably rather than accidentally. The standard's bet is that if you control the process, who does the work, what they are qualified to do, what checks the work passes through, then quality follows. It governs the assembly line, not the inspection of the final part. That distinction matters enormously once we reach the machine-translation standards, because they were built to bolt onto exactly this process skeleton.

ISO 17100 does not judge the translation. It governs how the translation is made: who is qualified, what steps it passes through, and what record proves it.

Why Process, Not Prose

It is worth pausing on why a standard would deliberately refuse to judge the words. The reason is that quality in translation is not a property you can fully specify in advance and then test against, the way you can test whether a bolt is 10 millimeters. Two excellent translators will produce two different, equally correct translations of the same paragraph. Quality is contextual, it depends on the audience, the purpose, the register, the brand. What you can specify in advance is the process that reliably produces a competent result: a qualified person doing the translation, a second qualified person checking it, agreed terminology, a clear brief, and a record of what happened. ISO 17100 is the recognition that in a craft where you cannot pre-define the perfect output, the next best guarantee you can offer a client is a disciplined, qualified, checked process. That insight is exactly why the standard survived into the machine-translation era. The engine changed who, or what, produces the first draft. It did not change the value of a qualified human checking that draft against the source before it ships.

The TEP Process: What the Standard Actually Requires

The operational heart of ISO 17100 is a production process, and once you see it laid out you will recognize it as the spine of every serious translation workflow you have ever worked inside, whether or not anyone named the standard. The industry shorthand for it is TEP, which stands for Translation, Editing, Proofreading, though the standard's own vocabulary is slightly more precise, and the precision is where the value lives. Let us walk it step by step, in the order the content moves.

Step One: Translation

A qualified translator renders the source content into the target language. The standard's word "translation" carries an obligation the casual use does not: the translator is required to handle meaning, terminology, grammar, register, locale conventions, and the client's specifications, and to do so working from the source, not from their imagination of what the source probably said. The translator also does their own self-check before passing the work on. This is the step the machine now contests. In an MT-first pipeline, an engine produces this first draft, and the question the later standards exist to answer is: if a machine did the translating, what has to happen next for the result to still count as professional? Hold that question. ISO 17100 answers it for the all-human case first, and the answer becomes the template.

Step Two: Revision, The Step Almost Everyone Skips

This is the clause that separates ISO 17100 conformance from "we translated it and it looked fine." The standard requires revision: a comparison of the source and target content by a second person who is not the translator, to assess whether the translation is fit for purpose. Read that twice, because every word is load-bearing. It is a comparison against the source, not a read-through of the target. It is bilingual, the reviser reads both languages and checks the target against the source segment by segment. And critically, it is performed by a different qualified person than the one who did the translation. The translator cannot revise their own work and call it conformant. The standard is built on the recognition that the person who produced a rendering is the worst-placed person to catch their own error, because they will re-read their own intention into the words instead of reading what is actually there.

The standard calls this second linguist a reviser, and it requires the reviser to hold translation-level competence, not merely the ability to read smoothly in the target language. Revision is a bilingual, comparative, source-versus-target check by a qualified second human. It is the mandatory four-eyes principle of professional translation. If you have ever worked a job where a colleague's translation came to you "for review" and you were expected to read it against the original and flag anything that did not match, you were performing ISO 17100 revision, whether the purchase order named the standard or not. And if you have ever worked a job where the translation went straight from the translator to the client with no second linguist in the loop, that job was not ISO 17100-conformant, no matter how good the translator was.

The revision step is the standard's beating heart: a second qualified linguist compares the target against the source. Skip it, and the work is not conformant, however good the translator was.

Step Three and Beyond: Review, Proofread, Verification, Sign-Off

After revision, the standard allows for further optional steps depending on the client's specification. Review is a monolingual, domain-focused check: a subject-matter expert reads the target alone to confirm it is appropriate for the field and the audience, for example a practicing physician reading a translated medical text for clinical sense. Proofreading is a final monolingual check of the target for surface errors, typos, formatting, spacing, the things you catch reading the finished product. Then there is a final verification and release by the project manager, who confirms the specifications were met before the content goes back to the client. Note the layering: revision is bilingual and mandatory; review and proofreading are narrower, often monolingual, and applied as the content and the contract require. The full chain, translation, revision, review, proofreading, verification, is what "the TEP process" really expands to once you look closely, and it is why a serious translation job has more than one name attached to it before it ships.

Competence: Who Is Allowed to Touch the Work

A checked process is only as good as the people inside it, and ISO 17100 knows it. The standard does not let just anyone be the translator or the reviser. It specifies professional competences the linguists must hold, and it specifies how a provider demonstrates that a given person holds them. This clause is the one that quietly turns "I speak two languages" into "I am a qualified translator," and it is the clause the machine-translation standards reach straight back to.

The Competences the Standard Names

ISO 17100 enumerates the competences a translator must possess. In plain working terms, they are:

• Translation competence: the ability to render meaning from source to target, handling problems of content and language, and producing output that meets the agreed specifications.
• Linguistic and textual competence in both the source and the target language: real command of both, not just fluency in one and a dictionary for the other.
• Competence in research, information acquisition, and processing: knowing how to find, evaluate, and use the information and reference material a translation requires.
• Cultural competence: command of the locale conventions, value systems, and regional specifics of the target market.
• Technical competence: the ability to use the tools of the trade, the CAT (computer-assisted translation) tool, the translation-management system, the termbases and translation memories, to carry out the work.
• Domain competence: understanding of the subject area enough to translate it correctly, which is why a legal translator and a medical translator are not interchangeable.

These are not abstractions. They are the difference between a fluent bilingual who produces plausible-sounding text and a professional who produces correct text, and the gap between those two is exactly the gap a fluent machine exploits. Hold this list, because the revised machine-translation standard, as we will see, demands that the human post-editing the engine's output holds this same set.

How Competence Is Proven

The standard does not leave competence as a feeling. It requires the provider to obtain evidence that a translator holds these competences, through one of several qualifying routes: a recognized degree in translation, an equivalent degree in another field plus translation experience, or sufficient years of documented professional translation experience. The reviser must additionally have translating and revising experience in the relevant domain. The provider has to keep records of these qualifications. This is the part procurement departments are really buying when they demand ISO 17100: not a promise that the words will be perfect, but documented proof that qualified people, with verified competences, did the work and checked each other. In a world where anyone can ask an LLM to translate a contract in four seconds, "a qualified human, demonstrably competent, was accountable for this" is precisely the assurance that the machine cannot self-issue. The qualifications clause is where the standard's value concentrates in the MT era.

Everything Around the Words: Brief, Assets, and Records

It would be a mistake to think ISO 17100 is only about the three or four people who touch the text. A large part of the standard governs the wrapper around the translation, the steps before the first word is rendered and after the last one is checked, and this wrapper is exactly where MT-first pipelines most often fail quietly. Three pieces of it deserve attention.

The Brief and the Specifications

The standard requires that the work be governed by agreed specifications: a clear understanding, fixed before translation starts, of the source and target languages, the purpose and audience of the content, the terminology to be used, the style guide, the locale, the file formats, and the deadline. This is not bureaucracy. A translation cannot be correct in a vacuum; it can only be correct for a stated purpose and audience. The same German sentence is translated differently for a legal filing and a marketing brochure, and the only way the translator and reviser know which is which is the specification. When a client hands a pile of strings to an MT engine with no brief, no glossary, and no statement of audience, the engine fills the vacuum with its own statistical guess about register and terminology, and the result reads fluently while quietly missing the purpose the client never wrote down. ISO 17100 forbids that vacuum.

Terminology, Translation Memory, and Linguistic Assets

The standard expects the provider to manage terminology and to use the client's termbase (the controlled list of approved terms) and translation memory (the TM, the database of previously translated and approved segments) where they exist. This matters because consistency is a quality dimension a single read-through will never catch: a term translated correctly on page one and differently on page forty is an error of the whole document, not of either sentence. The standard's insistence on terminology management is the human-process ancestor of every "the engine must honor the approved term" rule in the AI-assisted workflow. The discipline existed before the machine; the machine just made the discipline harder to enforce, because an engine will happily prefer a common synonym over the client's approved term unless something forces it not to.

Records and Traceability

Finally, ISO 17100 requires records: documentation of the project, the people who worked on it, their qualifications, the steps performed, and client communications, kept so that conformance can be demonstrated after the fact. This is the unglamorous clause that turns "trust us" into "here is the evidence." When a Critical error does ship, the record is what lets everyone reconstruct who did what, where the check was supposed to catch it, and how to fix the process. The provenance discipline that the advanced MT-pipeline lessons build, the segment-level record of source, edit, and decision, is ISO 17100's records requirement, grown up and made granular for a world where a machine wrote the first draft.

The Baseline the Machine-Translation Standards Build On

Now the payoff, and the reason this lesson sits where it does in the program. ISO 17100 was written for human translators, and yet it is the document the entire machine-translation standards landscape is anchored to. To see why, line up the three standards a working linguist hears about and notice what each one assumes.

ISO 17100 is the human-translation baseline: qualified translator, qualified second-person revision against the source, specifications, terminology management, records. ISO 18587, "Translation services: Post-editing of machine translation output," governs what happens when a machine produced the first draft instead of a human translator, and a human, the post-editor, brings it up to quality. ISO 5060:2024, "Translation services: Evaluation of translation output," formalizes how a human evaluator scores the result, classifying each error by dimension (accuracy, terminology, locale, fluency) and severity (Critical, Major, Minor), with the rule that one Critical error fails the file. Three standards, one shared skeleton, and that skeleton is ISO 17100's.

How the Post-Editing Standard Inherits the Baseline

Look at what ISO 18587 actually does. It takes the ISO 17100 process and asks: what changes when an engine replaces the human translator in step one? Its answer is that the translation step is now machine-produced, but everything that made step one professional, the competence, the comparison against the source, the terminology discipline, the records, has to be supplied by the human post-editor instead. The post-editor reads the machine's fluent draft against the source the way an ISO 17100 reviser reads a colleague's translation against the source. The skill is the same skill. This is why the revised ISO 18587, the version expanded to cover AI and LLM "non-human translation output" and in DIS (Draft International Standard) ballot with publication targeted for late 2025 into 2026, makes the inheritance explicit: it aligns directly with ISO 17100 and insists that the post-editor hold the same full linguistic competence as a professional translator, the exact competences enumerated in the ISO 17100 qualifications clause. The machine took over the typing. The standard responded by demanding that the human in the loop be no less qualified than the translator the machine replaced.

ISO 18587 did not invent a new, lighter kind of linguist for the machine age. It reached back to ISO 17100 and said the post-editor must be a fully qualified translator, because catching the machine's fluent error is a translator's job.

The Revision Step Becomes the Post-Edit

There is a clean way to hold the whole relationship in your head. In ISO 17100, a human translates and a different qualified human revises against the source. In the MT-first world, the machine "translates" and a qualified human post-edits against the source, and then, on high-stakes content, another qualified human revises or evaluates that. The bilingual, comparative, source-versus-target check, the four-eyes principle, did not disappear when the engine arrived. It moved. What used to be the revision of a human's draft is now the post-editing of a machine's draft, performed by a person who must hold the same competence the standard always required. And the reason the check matters more, not less, in the machine era is the silent critical error: a fluent machine draft hides its mistakes inside competent-sounding prose more effectively than a struggling human translator ever did, so the qualified second read that ISO 17100 mandated becomes the one thing standing between a fluent hallucination and a shipped liability. The baseline standard was right about the four-eyes principle before anyone knew how badly the machine would need it.

What Conformance Looks Like on a Real Job

Bring it back to the German project manager from the opening, because abstractions about standards mean nothing until they touch a delivery date. When she promised the distributor "ISO 17100," here is what she actually committed her LSP to do, translated from the clauses into the work.

First, she had to assign a qualified translator whose competences her company could document: the right languages, the right domain (medical-device instructions are not generic text), and the technical ability to work in the CAT tool against the client's TM and termbase. Second, she had to build in the revision step: a second, different, qualified linguist who would read the target against the German source segment by segment, not a quick proofread of the English, but a true bilingual comparison, because that is the clause that makes the work conformant. Third, she needed a specification agreed up front: purpose (regulatory submission and end-user safety), audience, terminology, locale, format. Fourth, terminology had to be managed against the approved termbase so the device's component names stayed consistent across forty pages. Fifth, the whole thing had to leave a record: who translated, who revised, their qualifications, the steps, so that if the distributor or a regulator ever asked, the conformance could be proven rather than asserted.

Notice what that list is not. It is not a promise that the words will be flawless, no process can promise that, and it is not a score of the output. It is a promise about how the work was produced: by qualified people, through a checked process, with the four-eyes revision in place and a record to prove it. That is the entire offer ISO 17100 lets an LSP make, and in 2026 it is a more valuable offer than ever, because it is precisely the assurance a client cannot get by typing their content into a free translation box. When the same project manager later starts running MT-first jobs, she will find that the machine changes step one and nothing else: she still needs the qualified human, the revision against the source, the terminology discipline, and the record. ISO 18587 will tell her the post-editor must be as qualified as the translator. ISO 5060 will tell her how to score what comes out. But the shape of the obligation, qualified people, a real check, a provable process, was set by ISO 17100, and it is the shape every one of those later standards builds on.

Why This Matters for the AI-Aware Linguist

If you take one thing from this lesson into an MT-first pipeline, take this. The standards are not three unrelated acronyms to memorize for a certification badge. They are one continuous idea, expressed first for humans and then extended to machines: a competent linguist, comparing target against source, accountable for the result, with a record to prove the process. When a client asks whether your AI-assisted workflow is "to standard," they are asking whether you have preserved that idea through the engine, whether a qualified human still revises against the source, whether terminology still holds, whether accountability still rests with a person, and whether you can prove it. ISO 17100 is where that idea is defined in its simplest, human form. Understand it here, in plain terms, and ISO 18587 and ISO 5060 stop being intimidating new rules and become what they actually are: the same baseline, carried forward into a world where the first draft is written by a machine that is fluent, confident, and, on the sentence that matters most, sometimes silently wrong.

Key Takeaways

• ISO 17100 is the international standard for professional translation services, first published in 2015. It is a process standard, not a quality metric: it governs who does the work, what they must be qualified to do, what checks the work passes through, and what records prove it, rather than judging the words on the page.
• Its operational core is the TEP process (Translation, Editing, Proofreading), which expands to translation, revision, optional review, proofreading, and final verification. A serious translation job therefore has more than one qualified name attached to it before it ships.
• The mandatory, load-bearing step is revision: a bilingual, comparative, source-versus-target check performed by a second qualified linguist who is not the translator. This is the four-eyes principle. Skip it and the work is not conformant, no matter how good the translator was.
• The standard requires documented competence: translation, linguistic and textual, research, cultural, technical, and domain competence, proven through a translation degree, an equivalent degree plus experience, or sufficient documented professional experience. Procurement is buying proof of qualified, accountable humans, not a guarantee of perfect prose.
• ISO 17100 also governs the wrapper around the words: agreed specifications and a project brief fixed before translation starts, terminology and translation-memory management, and records for traceability. MT-first pipelines most often fail in this wrapper, by handing the engine content with no brief, no glossary, and no statement of audience.
• ISO 17100 is the human-translation baseline that the machine-translation standards build on. ISO 18587 (post-editing) and ISO 5060 (evaluation) share its skeleton: a competent linguist, comparing target against source, accountable for the result, with a record to prove it.
• The revised ISO 18587 (covering AI and LLM "non-human translation output," in DIS ballot, publication targeted late 2025 into 2026) makes the inheritance explicit: it aligns with ISO 17100 and requires the post-editor to hold the same full professional-translator competence. The machine took over the typing; the standard demands the human in the loop be no less qualified than the translator the machine replaced.
• The bilingual source-versus-target check did not disappear in the MT era, it moved: what was the revision of a human's draft is now the post-editing of a machine's draft. It matters more, not less, because a fluent machine hides the silent critical error inside competent-sounding prose better than any struggling human translator ever did.