A rig maintenance manual landed on my desk last month tagged "Tier 2 — human edit of MT draft." It was a wellhead pressure-control procedure. People die when that's wrong. Somebody in a workflow tool had decided it was brand-adjacent content that just needed a polish. That decision was made by someone who couldn't read a word of it, in either language.
That's the thing nobody says out loud about how specialized work gets routed now. Before you see a single sentence, the job's already been sorted into a bucket, and the bucket decides your rate, your deadline, and how much of your brain you're allowed to use.
The tier system is real, and it's blunt
The model everyone's converging on is risk-tiered. Tier 1, low-risk: machine translation plus a light check. Tier 2, customer-facing: MT draft plus a real human edit. Tier 3, legal, medical, regulatory: human-only with CAT and QA tooling. On paper it's sensible. Match the effort to the consequences. I've got no quarrel with the logic.
My quarrel is with who assigns the tier and when.
In oil & gas the categories collapse into each other. A single PDF might open with marketing boilerplate about the operator's safety culture — pure Tier 1 fluff — and then thirty pages later give you the shutdown sequence for a gas compression train, where a mistranslated "close" versus "isolate" is a fatality. Same document. Same word count. Same tier tag on the folder. The person who tagged it read the title and moved on.
Here's what the tiering conversation misses: oil & gas isn't a soft technical domain that happens to have some risky bits. It's high-consequence technical work, sitting right next to legal, and it should default to Tier 3. HSE documents, permit-to-work systems, equipment safety cases — these have legal consequences and physical ones. But nobody's built that recognition into the routing. The industry talks endlessly about legal, medical, regulatory as the human-only trio. Energy safety documentation gets lumped into "technical" and pushed down a tier to save budget.
Automation bias, but for project managers
The legal side has finally named the problem. Automation bias — overconfidence in the machine without oversight — is now an explicit worry in legal translation, and rightly so, because the consequences show up in court. Slator's 2025 legal report leaned hard on human-in-the-loop for anything carrying legal weight. Good. That message is landing.
But automation bias doesn't only live in the translator trusting the MT output. It lives one step upstream, in the PM trusting the tier label. Once a job is tagged Tier 2, everything downstream inherits that confidence. The rate's set for editing, so nobody budgets time for terminology research. The deadline assumes a draft exists that's basically fine. If I flag that the machine's Turkmen is fluent and completely wrong on a torque specification, I'm now the difficult vendor who's blowing the schedule. The tier fought me before I opened the file.
And Turkmen makes this worse, because the MT draft you're "editing" was likely built on a thin model that doesn't know a drilling term from a kitchen appliance. Fluent and wrong is the default failure mode. So the Tier 2 promise — that a solid draft exists and you're just cleaning it — is a fiction for my pair before the work even starts.
What actually works
Stop tiering the document. Tier the segment, or at least give the translator the authority to re-tier on sight.
The agencies I like working with do a version of this already. They send the file with a note: "we've marked this Tier 2, but tell us if any section needs to bump up." That single sentence transfers judgment to the only person qualified to make it — the one reading both languages. Costs them nothing. Saves everyone the incident report.
The money supports this, if you look. The market's bifurcating hard. Commodity MTPE margins are getting crushed by legaltech startups doing AI-first document translation, while premium human work — certified court interpretation, patent translation — keeps its pricing power. Spanish is still over 30% of EU patent requests because Spain never joined the Unified Patent Court, and that resilient demand tells you something: the high-consequence, human-required work isn't shrinking. It's the middle that's hollowing out. Oil & gas safety documentation belongs in the premium half. It gets priced in the middle because of a folder label.
OpenAI shipped ChatGPT Translate in January, tone controls and fifty-odd languages, and it's genuinely useful for the Tier 1 stuff. It also, by its makers' own admission, falls apart in dense regulatory and legal terminology where errors carry legal or clinical weight. Nobody's arguing that. What's unresolved is who decides, per job, which side of that line a document sits on. Right now it's a dropdown menu selected by someone who can't read the content.
So when you send me a job, don't just tell me the tier. Tell me you'll believe me when I say it's wrong. That's the part of the workflow that no tool has automated, and the part that keeps somebody from tightening a valve the wrong direction because a machine was confident and a human was busy.