ISO 18587 Finally Catches Up with AI: Translation Is Now an AI Workflow

Translation standards have always arrived after practice. That is understandable: standards are meant to stabilize good practice, not chase every model release. But the gap between the language industry’s formal process model and the way modern translation is actually produced has become too large to ignore.

ISO 18587:2017 was written for a recognizable sequence: a machine translation system produces output, a qualified human post-editor corrects it, and the result is delivered. In 2017 that was already useful. In 2026 it is no longer a sufficient description of serious multilingual production.

Today a document may be ingested automatically, matched against previous approved translations, conditioned by terminology and style rules, translated by an NMT engine or an LLM, evaluated by a quality-estimation model, checked by a second evaluator or “judge” model, routed according to risk, reviewed by a human only where necessary, and fed back into the next cycle. Calling all of that “machine translation post-editing” obscures more than it explains.

What has actually changed in ISO 18587?

The most visible change is in the title of the draft itself: “Translation services — Post-editing of non-human translation output — Requirements.” ISO’s public abstract says the document covers the process of human post-editing of non-human translation output and post-editors’ competences, and is applicable to content generated by non-human translation systems.

That phrasing may sound slightly awkward, but it is strategically important. It avoids tying a process standard to one generation technology. Conventional NMT, generative AI and large language models can evolve rapidly; the standard can instead focus on the production process, human competence, customer requirements, feasibility and quality.

Industry reporting on the revision also points to a broader vocabulary that includes concepts such as non-human translation, artificial intelligence translation and automated post-editing, stronger alignment with ISO 17100, explicit attention to customer specifications, and a feasibility/risk perspective when deciding how post-editing should be applied.

The old workflow is becoming the real risk

Much of the translation industry still operates a workflow whose basic economics were established twenty or thirty years ago: files arrive, project managers prepare them, translators work through them, reviewers check them, files are moved back through the chain, and the language service provider absorbs coordination cost, staffing risk and often a substantial financing gap between paying linguists and being paid by the client.

Adding an LLM at the beginning of that chain does not make it an AI workflow. It simply makes the first step faster while preserving most of the old operating cost.

This is where the revision of ISO 18587 becomes more interesting than a terminology update. A standard that pays more attention to customer requirements, feasibility, different levels of post-editing and hybrid workflows is much closer to the decisions enterprises actually need to make.

It also creates an uncomfortable question for the industry: if quality can be measured earlier, if terminology and approved translations can be applied automatically, and if human intervention can be directed to the segments or documents where it is genuinely required, why preserve a workflow whose main virtue is that everyone already knows how to operate it?

The human is not removed. The human moves up the stack.

The weakest argument in the AI-versus-human debate is that automation necessarily means eliminating professional translators. It confuses the number of manual touches with the amount of human expertise in the system.

In a mature AI translation workflow, expert linguists remain essential, but their role changes. Human expertise is increasingly valuable for defining what good output means, approving terminology and style, creating reference data, identifying high-risk error types, designing evaluation rubrics, adjudicating ambiguous cases, validating sensitive content and improving the system when it fails.

This distinction is particularly important with LLMs. A fluent error is often harder to detect than an obviously broken machine translation. The answer is not therefore “remove the human”. The answer is to use humans where their judgement changes the risk profile, rather than using human labour as the only available quality-control mechanism.

A modern translation workflow should measure before it routes

Pangeanic’s view is that enterprise translation should be treated as a controlled production system. The generation model matters, but the system around it matters more. A strong workflow connects organizational language assets, AI generation, automated evaluation, human review and production feedback.

This is the operating logic behind Pangeanic’s Deep Adaptive AI Translation : apply the organization’s own linguistic evidence during generation, then use Machine Translation Quality Estimation (MTQE) as a release and routing layer rather than treating human review as an undifferentiated final step.

The standard should govern outcomes, not fossilize labour

Standards are valuable when they protect the buyer from uncontrolled variation. They are less useful when organizations interpret them as instructions to preserve an old sequence of tasks simply because those tasks were historically associated with quality.

The question procurement teams should ask is not whether a provider has inserted AI into its workflow. Almost everyone has. The question is whether the provider can explain, measure and reproduce the decision process around that AI.

What ISO 18587 does not mean

There is a risk that a standards update becomes marketing shorthand. The current draft does not justify several claims that are already tempting to make.

• It does not certify an LLM or NMT engine. ISO 18587 is a process-oriented standard for post-editing and competence, not a leaderboard for generation technologies.
• It does not say that LLMs are better than NMT. Different systems remain appropriate for different languages, domains, latency constraints and deployment environments.
• It does not make raw AI output “ISO compliant” by itself. The scope is human post-editing of non-human translation output.
• It does not remove the need to evaluate quality. In fact, a more flexible workflow makes measurement more important because routing decisions must be defensible.
• It does not mean every AI-assisted translation must be publicly labelled as AI-generated under the EU AI Act. Article 50 contains specific scopes and exceptions; translation use cases need to be assessed according to the actual function, degree of transformation, publication context and human editorial control.

From post-editing to AI Data Operations

The deeper significance of ISO 18587’s revision is that translation is becoming easier to describe as a data and evaluation operation rather than a sequence of linguistic handoffs. The workflow has inputs, approved evidence, generation models, evaluation signals, human decisions, release rules and feedback. That is structurally much closer to modern AI Data Operations than to the project-management model on which much of the language industry was built.

This does not erase the translation profession’s history. It confirms its technological continuity. Translation memories were structured evidence. Terminology databases were governance. Machine translation introduced automated generation. Quality estimation made confidence measurable. LLMs added flexible reasoning and context. The next step is not another isolated tool. It is to connect these pieces into one controlled system.

For a technical comparison of generation approaches, see NMT vs LLM translation: which is better for my use case? . The more useful enterprise conclusion is increasingly that the model choice should sit inside a governed architecture rather than define the architecture itself.

The economics are part of quality

One of the most persistent mistakes in translation procurement is to separate quality from workflow economics. A process that achieves excellent output but requires unnecessary manual handoffs, duplicated checks, dedicated coordination and avoidable financing pressure is not automatically the safest process. It may simply be an expensive process whose inefficiencies have become familiar.

Good automation does not mean “accept the model output”. It means making every expensive human action earn its place. If automatic evaluation can identify where confidence is low, human expertise should concentrate there. If terminology can be enforced before delivery, it should not be rediscovered manually in every job. If approved translations can condition new output, they should not sit passively in a database waiting for a fuzzy match.

What buyers should do now

1. Map the real workflow. Document every point where content is ingested, generated, transformed, evaluated, reviewed, approved and delivered.
2. Separate quality requirements by risk. Publication, legal, safety-critical, internal and exploratory content should not automatically follow the same review path.
3. Treat terminology and approved translations as active assets. They should influence generation and validation, not merely exist as reference files.
4. Introduce measurable quality gates. Use MTQE and other validation methods before deciding where human review is required.
5. Define the human role explicitly. Specify when bilingual review, subject-matter expertise, full post-editing, light post-editing or adjudication is required.
6. Preserve provenance. Record enough information about models, assets, checks and reviewers to reconstruct important decisions.
7. Measure the whole outcome. Track quality, rework, human effort, turnaround and cost per successful task rather than optimizing a single unit price.

ISO is catching up. The industry should do the same.

ISO 18587 is not yet a final revised international standard, and the precise text may still change. But the draft has already made one thing difficult to deny: “machine translation” is too narrow a category for the way professional translation is now produced.

The important division is no longer human versus machine. It is controlled versus uncontrolled production. Can the organization apply its own linguistic knowledge? Can it measure output quality? Can it route risk? Can it explain where humans intervene and why? Can corrections improve the next result? Can the entire process be audited without recreating it manually?

That is the standard enterprise buyers should increasingly demand. ISO 18587 is beginning to adopt the language required to describe it.