If AI Makes Translation Faster, Why Isn’t It Cheaper? I Decided to Test That

My long-time client recently asked me a question that is becoming harder and harder to avoid:

“If you use AI for translation, can you lower your rate?”

It is not an unfair question. In fact, it is exactly the question anyone would ask. AI has become fast, fluent, and impressively confident. So the logic seems obvious: if the work is faster, the work should be cheaper.

That is the theory.

The reality is messier.

Rather than pushing back on instinct alone, I decided to test it. Not in the abstract, and not in the breathless style of people who seem convinced that every new tool is a revolution. I wanted to know something much simpler: how much does AI actually improve translation efficiency in real work?

The test

I looked at the kind of English-to-Japanese financial translation I deal with regularly and compared three workflows:

1. AI translates, then a human reviews

2. A human translates, then AI reviews

3. A human translates, then the same human reviews

The texts were different, but comparable in difficulty. The results were these:

• AI translation + human review: 375 words per hour

• Human translation + AI review: 366 words per hour

• Human translation + human review: 320 words per hour

So yes, AI improved productivity.

Just not by very much.

To be honest, I expected a more dramatic difference. What I found instead was something far less exciting and much more familiar: AI speeds up part of the process, then hands the hidden cost back to the human at the end.

The first problem: AI is convincing in exactly the wrong way

The biggest trap in AI translation plus human review is that the draft often looks good enough to trust.

That is not a feature. That is the problem.

AI is very good at producing language that sounds polished. The prose flows. The sentences are smooth. The tone is often plausible. At a glance, it can look impressively competent.

Then you read it against the source.

That is when things start to wobble.

Not always dramatically. In fact, dramatic errors are the easy part. What takes time is the sentence that is almost right, the nuance that has been quietly flattened, the meaning that has drifted just far enough to matter but not far enough to be obvious.

That sort of error is far more expensive than a bad translation. A bad translation announces itself. A plausible one asks to be trusted.

When I translate something myself, I remember where the difficult choices were. I know which phrase I wrestled with, which clause I deliberately kept tight, and which compromise I made for the sake of clarity. With AI output, that reasoning is invisible. There is no trail. So I have to reconstruct the entire thought process from scratch by checking the source line by line.

People imagine AI removes labor. Often, it simply relocates it.

It turns translation into a highly specialized game of Where’s Waldo?—except Waldo is hiding inside a quarterly earnings release, and missing him may create a very expensive problem.

Fluency is not the same as accuracy

This is where a lot of non-translators get misled.

AI writes smoothly. Very smoothly. Sometimes suspiciously smoothly.

That makes it impressive in general prose. It makes it dangerous in finance and IR.

Because in this field, the job is not to produce something that reads well in isolation. The job is to preserve meaning under pressure. A sentence can be elegant and still be wrong. It can sound natural and still fail the source text. It can be cleaner than the original and, for that exact reason, less accurate.

AI tends to optimize for readability. Unfortunately, readability is not always the highest priority. Sometimes the text needs to retain a legal hedge, a technical distinction, or an unpleasant ambiguity because that is what the original says.

Finance does not particularly care whether a sentence glides. It cares whether the sentence survives scrutiny.

AI still does not think in terms of structure

Translation is not just local sentence work. Or at least, good translation is not.

A strong translation often requires structural judgment: tightening a paragraph, repositioning emphasis, managing information flow, sometimes even reordering material so the argument actually lands in the target language. This is especially true in dense business writing, where what matters is not only what each sentence says, but how the whole document moves.

AI can handle individual sentences reasonably well. What it still struggles with is architecture.

It does not really see a document in the way an experienced translator does. It processes. It predicts. It smooths. What it does not do reliably is decide, with judgment, what the piece should sound like as a whole.

That part remains stubbornly human.

Human translation plus AI review is not the safe solution people imagine

The alternative workflow—human translation followed by AI review—sounds more sensible. And in some ways it is. But it comes with its own set of problems.

First, AI is better at spotting surface issues than substantive ones. Grammar, typos, awkward phrasing—fine. Subtle mistranslation, tonal misfire, logical slippage—less fine.

Second, AI has a habit of “improving” things that were not broken. It often cannot resist rewriting. A sentence may already be accurate, disciplined, and appropriate for the document, but AI will cheerfully replace it with something that sounds smoother, looser, and slightly less faithful. In other words, it confuses editing with interference.

Third, terminology remains unreliable. And this is where the cheerful optimism around AI tends to collide with professional reality. You can provide a glossary. You can define preferred terms. You can specify tone and style. You can explain the context in painful detail. And AI will still, sooner or later, decide to improvise.

That is irritating in casual content. In financial, legal, or IR translation, it is a quality problem.

Finally, AI often flattens tone. A human translator may intentionally aim for restrained, compact, slightly formal English—the kind that suits serious corporate communication. AI, in its endless enthusiasm to make everything “more natural,” may soften that tone into something friendlier, wordier, or more generic. Grammatically acceptable, perhaps. But wrong in the way that matters.

So, does AI make translation more efficient?

Yes.

Just not nearly enough to support the fantasy that translation has suddenly become cheap.

That is the part clients understandably want to believe. It is also the part that does not quite survive contact with actual work.

AI helps with drafting. It can reduce friction. It can save time at the front end, especially in non-native output, where getting to a grammatically workable draft is no small thing.

But the time saved upfront is often paid back during review.

And there is a reason for that: the cleaner AI output looks, the more dangerous it becomes to review casually. A messy draft forces vigilance. A polished draft invites trust. That misplaced trust is where time—and risk—creep back in.

So no, AI has not turned professional translation into a task that is two or three times more efficient. Not if the standard is still quality. Not if the translator is still expected to be accurate. Not if terminology still has to be controlled. Not if someone is still responsible when things go wrong.

Which, in the end, someone is.

The real issue is not speed. It is accountability.

This is the part that often gets left out of the pricing conversation.

Translation is not priced solely on the act of converting words from one language to another. If that were the whole job, then yes, perhaps AI would have already blown the pricing model apart.

But that is not the whole job.

The real work lies in preserving meaning, controlling tone, maintaining consistency, spotting distortion, and taking responsibility for the finished text. That responsibility has not disappeared. It has not even decreased all that much. It has simply moved into less visible parts of the workflow.

AI has made some tasks lighter. It has not made professional judgment optional.

And that is why “Are you using AI?” is the wrong question if the real question is “Why does this still cost what it costs?”

The better question is: who is taking responsibility for the final text?

Because that, more than speed, is what the client is actually paying for.

And yes, I will admit it: part of me is still quietly waiting for one of the many perfectly plausible AI mistakes hiding in plain sight to cause a large enough mess that people finally say, with feeling,

“Right. So humans are still useful after all.”

April 3, 2026