Two languages, one publishing date.

Plenty of organizations are required to publish in two languages, by regulation, by charter, or simply because half the audience reads the other one. In practice the first language goes out on schedule and the second goes out when somebody finds time, which is usually about three weeks later, by which point it is no longer news and the audience reading it has learned that they are the afterthought.

The alternative failure is faster and worse. Somebody runs it through machine translation to catch up, and the specialist vocabulary comes out technically correct and professionally wrong. In a field with its own language, a term that is nearly right is a signal to every expert reader that nobody who understood the subject looked at this.

There is almost always a glossary. It is almost always a spreadsheet, it is usually two years out of date, and nobody owns it.

There is a translator, internal or contracted, who is a bottleneck by capacity rather than by skill. They are not slow. There is simply one of them and a publishing calendar built as though there were three.

And there is an unstated hierarchy in which the second language is treated as a translation of the first rather than a version of it, which is the actual root of the three-week gap. Everything downstream follows from where the second language sits in the process.

The scarce resource here is not typing, it is judgment about terminology and register. So the pipeline should hand the reviewer a draft to correct rather than a blank page to fill, and should have already flagged every term where an approved equivalent exists.

That reframes the glossary from a reference document into the most important system in the pipeline. It gets a real home, a review workflow and an owner, and it is applied as a constraint on generation rather than as advice.

The last decision is structural rather than technical: the second language stops being a downstream step and becomes a parallel one, gated by the same publish action. That is what actually closes the three-week gap. The tooling only makes it affordable.

When a piece is marked ready in the CMS, a webhook fires. The pipeline drafts the second-language version under the glossary constraints, annotates it with confidence markers, and places it in a review queue with the reviewer notified. Nothing publishes until a person approves it.

Five decisions the pipeline is built around

1. The glossary is applied as a constraint, not a suggestion. Terms with an approved equivalent are locked, and where the model would have chosen differently the draft flags the disagreement rather than silently substituting. The reviewer sees the argument instead of having to go looking for it.
2. Confidence markers put the reviewer’s attention where it belongs. Idioms, proper names, legal phrasing, numbers with units and anything the model rated as uncertain are highlighted and sorted first, so a forty-minute review is spent on the fraction that needs a human rather than reading a clean paragraph for the third time.
3. Every correction the reviewer makes becomes a candidate glossary entry, queued for approval. The terminology improves as a byproduct of work that was already happening, which is the only form of continuous improvement that survives a busy quarter.
4. Publishing is atomic. Both versions go live together with correct hreflang tags pointing at each other, or neither goes live. A half-published bilingual site is a search engine problem on top of a credibility problem, and it is a surprisingly common one.
5. Every published piece records a named approver and a timestamp. For a regulated or member-accountable organization that record is not paperwork, it is the reason the process is allowed to exist at all.

None of this reduces the number of people involved. It changes what the person is doing from production to judgment, which is both a better use of them and, not incidentally, a much easier thing to recruit for.