A batch of AI-translated plan materials goes out the door. The language reads cleanly. Nothing looks off. The project closes on schedule.
Eight months later, a compliance reviewer asks one simple question: who reviewed this, and against what standard? Nobody has a clean answer. The translation itself isn’t the problem. The problem is that nobody can prove it was handled the way it needed to be.
That gap, between content that reads fine and content you can actually stand behind, is where most AI translation programs quietly fall short.
Why “Good” Stopped Being the Right Question
Maciej Modrzejewski, VP of Artificial Intelligence at BIG, has spent much of this year comparing notes with machine translation researchers and platform builders, including presenting his own research at the Association for Machine Translation in the Americas (AMTA). He keeps landing on the same conclusion.
“Anyone can plug a commercial engine into a workflow and get something usable out the other end,” he says. “The real question is what happens around that output: who’s accountable for it, and whether you can prove it.”
A few years ago, translation quality was the whole pitch. Today it’s assumed. That’s not a knock on AI translation; it’s an acknowledgment that quality stopped being scarce. What’s still scarce is a record of why a particular piece of content got the treatment it did, and someone willing to stand behind that decision.
Two Documents, Two Very Different Paths
Picture two pieces of content landing in the same queue on the same morning: a routine FAQ update for a website, and a packet of plan materials a Medi-Cal member will use to understand their benefits. Both might start with the same engine.
From there, they diverge. The FAQ clears a quality threshold and ships. The plan materials get flagged for human review: a bilingual linguist checks it against the source, a second reviewer signs off, and the whole chain gets logged: which engine touched it, what confidence score it came back with, who reviewed it, when, and against what regulatory standard.
That’s the kind of record a health plan’s compliance team, a government agency’s language access office, or outside counsel reviewing a translated patent filing will eventually ask for. Not because they distrust the translation. Because “trust me” isn’t an answer that holds up in an audit, a public records request, or discovery.
Neither path is a shortcut. Both are deliberate. The skill isn’t applying one process to everything; it’s knowing which document earned which path, and being able to show the decision was made on purpose.
Where This Is Headed
Maciej’s own picture of where localization goes next looks less like a translation engine and more like an experienced project manager: pulling the right assets, coordinating multiple models, flagging exactly which content needs a linguist, and keeping that record automatically rather than as an afterthought.
Most of the pieces already exist. What’s still being built, across the industry, is the layer that makes all of it accountable, not just functional.
It’s also changing how this work gets priced. Per-word invoices are giving way to something closer to a managed program: a defined scope, a partner accountable for the whole workflow, rather than a bill that reflects word count and says nothing about what happened along the way.
Where the Gaps Actually Show Up
Skip that structure, and the gap doesn’t disappear. It waits. It shows up during an audit, during a client’s own compliance review, during litigation, or after content has already reached the people it was meant for.
At that point, your team is reconstructing a review history that was never written down. Your compliance or legal group is re-validating content that should already have been defensible. A government partner is fielding a records request with no clean answer. None of that is cheap, and none of it is fast, and it tends to surface at the exact moment you can least afford the delay.
Where BIG Comes In
The time to decide how content gets reviewed, and what gets logged, is before translation starts, not after someone asks. That’s part of how we plan AI-assisted work: matching the level of review to the risk, logging the chain of custody as we go, and keeping a record that holds up whether it’s a routine audit or outside counsel asking hard questions.
If you’re thinking through how AI fits into your language access program, especially content that has to hold up to regulatory or client scrutiny, we’d welcome the conversation. Talk to us about what a calibrated, accountable approach to AI-assisted language work looks like for your organization.
