Out Through the Front Door

· @Steven

I’m leaving the translation profession. I’m winding down CotranslatorAI, the tool and business I’ve spent the last few years building. And I’m turning my time and effort toward something that isn’t translation.

I’m writing this during a week of hiking alone in the Korean mountains. I usually hike with my wife, but this time I had enough to think about that I wanted to walk by myself. Four hours a day on the trail, blisters on both feet, and the rest of the day at a table with my laptop, thinking it through.

This isn’t an announcement. I’m putting it somewhere quiet on my website, mostly for myself. I want to set down, in order, how I got here, so I can close this door properly. If you’ve found it anyway, you’re welcome to read it. Maybe it will help you think through your own situation.

One thing up front. I didn’t reach this conclusion because I gave up early or didn’t try hard enough. I reached it because I tried everything I could think of, honestly and in the open, and kept running into the same wall. Most of what follows I’ve said in pieces to friends and colleagues over the past year. Here it is in one place.

The bet I made

For the past few years, everything I did rested on one idea. If human translators were going to survive alongside AI, we had exactly one lever left. We can’t be faster than the machine, and we can’t be cheaper. The only thing left is better.

CotranslatorAI was built on that bet. I designed it as a precision AI instrument for a human who runs the process with full control. It isn’t wired into agents and doesn’t automate anything on its own. It’s a blank slate: you write whatever prompts you want, using whatever technique you want, and it moves text in and out of whatever you’re working in through ordinary Windows cut-and-paste.

That meant it worked anywhere. Trados, memoQ, a client’s quirky online portal: if the interface had quirks, you could prompt around them. I priced it at $87 a year.

Behind the tool was a workflow, and the workflow mattered more than the tool. Working segment by segment inside a CAT tool, you lose context. So I stopped staying inside the CAT tool. I export a bilingual file and translate 50 to 500 segments in a single pass, with a robust anchor prompt, the relevant TM matches and the client’s terminology. Then it goes back in with tags and formatting intact, end to end.

In memoQ especially, I made the software do things it was never designed to do. I used to joke that it was black magic. The result is a cohesive translation, consistent from top to bottom, that I would put next to anyone’s.

I didn’t just build it. I taught it. I demoed the tool to thousands of translators and taught the workflow in depth to dozens. I recorded hours of training videos and ran weekly coaching for the better part of 2025. With a colleague’s help, I ran a five-day challenge, one hour a day at midnight my time in Korea. For the first time, students actually came out the other end able to do it. I hosted a couple monthly drop-in clinics on Zoom for anyone who wanted help.

Then I took the workflow to agencies as a service. I’ve pitched it to hundreds of LSPs, many of them more than once; by now they’re sick of hearing from me. Every time someone pointed out a weakness, I went back and fixed it. Today the workflow simply works.

I’m proud of all of that. None of it changed the outcome.

Fourteen clues

At the end of October I started keeping a list of signs that the translation industry is dying. I reached fourteen without trying very hard. They fall into three groups: what happened in my own business, what I watched the AI do with my own eyes, and what’s happening across the industry.

In my own business

  • CotranslatorAI won’t sell. Not through my funnel, and not in any packaging I could come up with. Translators will happily attend a webinar about an AI approach. Ask them to actually learn it or buy it, and nothing happens.
  • The people who do buy don’t stay. Users are losing faith in it. I set up my system to email everyone who cancels and ask why, and I’m about to turn that off because the replies have become too depressing to read. One came in this week: he’d moved to Claude, because he could pile everything in and still get a good translation. I’ve always taught that you have to cultivate your context, not throw in the kitchen sink. The honest truth is that the AI has gotten good enough that he’s probably right. Even one of the most committed students from my five-month course found that, as the course went on, there were fewer and fewer projects to use it on. Clients rarely send a whole manual to translate from start to finish anymore. They ask you to train their machine, build their terminology, or check their output.
  • Agencies listen, but they don’t buy. This was the one I never expected. I assumed LSPs would jump at a workflow that lets them deliver better work at lower cost, especially if I do all the work for them at a penny a word. Everyone is interested; nobody buys. I read that as a verdict. They don’t see a future worth investing in either, so everyone is hedging.

In the AI itself

  • The AI catches what I miss. A couple of months ago I took a post-editing job at half price. I machine-translated it, post-edited every word, and delivered work I was satisfied with. Afterward, I ran my QA prompt over it. It found mistake after mistake, things I had no real chance of catching on my own. I would have delivered better work by skipping my own word-by-word pass, running the QA prompt and fixing what it found. I don’t think I’m an idiot. It was a humbling day.
  • It does document work better than I can. I put a PDF into Claude and got a far better Word reconstruction than anything I could have produced myself: faster, cheaper and better. I haven’t yet had the courage to ask it to translate that PDF into Word in the same pass, but I’m fairly sure it could. Over the long term, why would anyone pay a translator for that?
  • One setup can serve a whole company. Nothing stops a company today from building a single company-wide translation skill or project in Claude, loaded with its terminology, style guide and reference material, and handing it to every employee. Add one localization person for the hard questions, and I’d estimate that 90, 95, even 98 percent of the work gets done with no outside help. Consistency comes built in. Is there an occasional mistake? Sure, but there were always mistakes when humans did it. For most content out there, raw AI with good QA checks is already better than the best humans were a few years ago. And the vast volumes that were never translated at all can now be translated, and translated well.

Across the industry

  • The flagship conference is shrinking. The ATA conference had close to a thousand attendees not long ago. Then 800, then about 400 last year, and from what I hear about 300 this year. Membership has reportedly fallen by more than half.
  • Other gatherings are disappearing. memoQ didn’t hold memoQfest this year. TAUS has ended its own conferences and handed them over to LocWorld. The BP conferences have come to an end too. One after another, the places where this profession met are closing.
  • The AI summit became a direct-client summit. A well-known online AI in Translation summit didn’t run this year. In its place came a summit on finding direct clients. Translators have quietly given up on AI as their way forward and are pinning their hopes on new clients instead.
  • “The future of translators” now means leaving translation. I attended an industry webinar billed as being about the future of translators. It wasn’t about the future of translating at all. It was advice on changing to a different job.
  • Post-editing volumes are falling. On that same webinar, one of the industry’s longest-standing observers remarked that even machine translation post-editing volumes are now dropping. That’s exactly what I would have predicted.
  • The experts are guessing. I read a major research firm’s book on AI in translation. When it gets to what translators will do, it reaches for cultural insight, judgment and process design. To me it reads like guesswork dressed up as analysis. I don’t believe the people writing it actually know.
  • Even a big following can’t fill a low-cost community. A well-known translator trainer with a large audience launched a membership for under ten dollars a month and could get only a few dozen people to join.
  • The market has noticed. Stock prices of translation companies are falling hard.

Why “better” stopped being enough

Put those clues together and the logic is hard to escape. The one lever I bet everything on, better, is losing its value from every direction at once.

Clients can’t see the difference, and don’t need to

Most clients can’t evaluate translation quality. A few trust their translator and believe the work is good, but they don’t actually know; they just believe it. Most of us don’t have that kind of client.

When we try to explain why AI output falls short (this term is inconsistent, that word is slightly off), clients hear noise. Their answer is: just pick one and move on. Maybe they’ll budget fifty dollars for someone to settle a glossary. Once that glossary and a style guide are in the AI, good enough is good enough.

And good enough keeps getting better. Why work five times harder to get from 98 percent to 98.5 percent when you can get 98 percent for almost no effort? Four years ago, the extra effort of a careful human workflow bought a lot. Today it buys a little, the margin is shrinking fast, and nothing guarantees it won’t reach zero soon.

Human quality was never as good as we claimed

We like to say that if you want real quality, you need a human. My QA experience says otherwise. Humans don’t deliver such great quality; we overestimate ourselves. The AI is finding far more mistakes than we are.

“A human has to take responsibility” is mostly psychological. The client wants a person to say they stand behind the work. But if we’re talking about the actual probability of producing a flawless document, a rigorous AI-based workflow with a small amount of human intervention beats a traditional all-human workflow. It beats machine translation post-editing too.

That includes the regulated content people point to as the safe harbor, the material that can’t afford to be wrong. Human translations of that material contain more errors than AI translations do now, and AI can help squeeze those errors out. Put a translator at the end for a few checks, maybe. But the translator’s central role there is basically gone.

Post-editing was a bridge, not a destination

For years we were told that full human translation was declining but post-editing was growing. So I positioned CotranslatorAI as a post-editing tool, and that’s where my best users found the most value.

But post-editing is a transitional workflow. In the end, you either go all human or all automation. All human means a shrinking niche of bespoke work for a handful of translators. Everything else goes to full automation, to agents and pipelines, and none of it will run through a tool like mine. It won’t be as good as what a careful human workflow produces. It will be good enough, and good enough keeps rising.

Nobody invests in a shrinking business

This surprised me most, and it’s the part I finally understand. Translators aren’t turning down CotranslatorAI because of the $87. They’re turning down a new effort.

Their volumes are dropping, and I’m offering a way to do more work, faster and better. Those two things don’t match. You need more work before it’s worth learning to do it faster. I’ve tried every angle I can think of to square that circle, and I can’t.

The agencies are in the same position. Why buy a new workflow to produce something you’ll have to sell more cheaply than you sell it now, to clients who are already leaving? Some of it is inertia. Some of it is a rational plan: milk the current work for as long as it lasts, then get out.

When I talk about an approach that could help them in the long run, they listen politely. They just don’t believe there is a long run. Even the agencies that keep their business alive with AI translation don’t need human translators to do it, and there’s nothing they can do with the translators they have.

A couple of years ago, everybody wanted to learn AI. This year, almost nobody does. It’s interesting, even fun, but there’s no point learning to work faster when the work is disappearing.

Confidentiality won’t save us

We keep telling ourselves clients will pay for confidentiality. I’ve thought for years that this argument is a farce. Clients say they care, and they use it as a stick against translators. But the moment someone shows up a penny cheaper, they take it.

Five or six years ago, before ChatGPT, I tested this. I had a spare computer in my office, so I offered my clients a deal. For one penny more on my rate, I would do all their work on a machine with no internet connection, nothing sent by email, nothing that could be hacked. Not one client took me up on it. One agency contact wrote back, half laughing, asking whether anyone would really pay an extra penny for that.

Today there are breaches every day, and agents roam everywhere. Our data gets stolen whether we like it or not. When an AI provider says paying customers’ data will be kept secure, that’s all most clients need to hear.

The tools are being swallowed too

CotranslatorAI isn’t the only tool going under. memoQ, Trados and the other CAT tools that sponsor our conferences are being swallowed the same way. You can already do better work in Claude. You build a project, load your terminology and the relevant TM matches, write an anchor prompt, and go. An agent can even paste the translations into the client’s interface for you, which was half the reason my tool existed in the first place.

It’s also what people want. They don’t want a simple, limited interface. They want something that feels like it opens onto the whole world. For years I’ve been saying, in effect, come back to Kansas City, it works better here, while everyone else heads to Las Vegas for the lights. Nobody believes me, and lately I’m not sure they’re wrong.

We still have cabinetmakers. A few people with money pay for something built exactly to spec. Everyone else goes to IKEA. That’s where translation is headed: a small bespoke niche, and everything else assembled at scale.

Every way forward is a career change

Look closely at what’s being offered to translators now. Nearly all of it is a career change with a translator’s name attached.

Become an intercultural communications consultant. Become a multilingual marketing consultant, or a localization consultant for small companies. Go to trade shows and talk to exporters, but don’t use the word “translate,” because nobody there is buying translation. Learn sales and marketing. Move up the chain with your clients and become a business person.

These are real paths, and some of the people teaching them are doing serious, thoughtful work. Some translators will make it. But none of this is translation. These are different jobs that borrow a few translation skills.

I’ve looked at the in-house side too. I do think a localization role in every company, even a part-time one, is viable, and I’ve explored it. But that’s a career change as well.

And we’re asking a lot. Most translators aren’t young. We’re telling people who spent decades building a deep, specialized skill set to carry a small piece of it into a new field, with a new mindset and new skills learned from scratch. Some will manage it. Most won’t. That isn’t a judgment of anyone; it’s what happens when a whole profession is asked to reinvent itself at once.

Many translators have already left. Some are holding on to the hope that direct clients will save them. For a few, they will. For most, they won’t.

Once this direct-client wave passes, people will discover that working directly with clients means becoming a business consultant, and most will realize they can’t or don’t want to. I think that’s when we’ll see the final poof. My guess is that much of what’s left of the profession disappears sometime next year. I’d be surprised if the ATA conference happens more than once more.

I hope I’m wrong about that.

What I’m doing now

Here’s what this means in practice.

CotranslatorAI is winding down. The monthly drop-in clinics are over. When I find the time, I’ll change the website so the only thing you can buy is a three-month subscription. That way my support commitment to anyone who signs up never runs longer than three months, and I can wind things up responsibly instead of just switching off the lights. I’m not planning a big announcement.

I’ll still take translation work. If a job comes in, I’ll do it. I need the money; things have been tight enough that putting food on the table has been hard, so this isn’t an abstract decision. But I’m no longer investing my time and energy in growing a translation business.

I’m looking for work. I’m applying for localization roles, and if someone wants to hire me as a localization person, I’ll take it. In the short term I’ll also do data annotation and similar work.

I’m building on Korea. The longer-term direction draws on my Korean business background. I hold a PhD in business from a Korean university. Korea is popular right now, and people want to do business with Koreans. I don’t plan to lead with “I’ll translate your website.” Exactly what I’ll offer, I’m still working out.

One loose end remains. Next month I’m scheduled to speak at a conference on generative AI in localization. I wrote that talk as an encouraging message for boutique LSPs, and I don’t believe it anymore. I’ve gone back and forth about whether to give it as planned, rewrite it, or withdraw. I’ll probably give it. Maybe some LSPs in that room will take the workflow, run with it, and prove me wrong. I’d be glad if they did.

Out through the front door

Through all of this, I tried to do things the honest way. No shortcuts, no gaming the system, no pretending. I came in through the front door. I built the tool, taught it, sold it, listened to every objection, fixed every weakness I was shown, and kept going back.

That has made my view clearer. I’m not guessing from the sidelines. I’m reporting what happened when I gave it everything I had.

There’s an odd relief in that. For a long time I was working to believe something, hoping against hope that what I was seeing wasn’t so. That’s exhausting. Now I have a conclusion, and a conclusion is something you can act on. The pressure to keep believing is gone.

A friend who walked much of this road with me put it better than I can: we didn’t fail; history went faster than we did. I think that’s right. The tool was good, and the workflow is good. The ground moved.

I’m grateful to the people who kept me company this year: the colleagues who listened, argued, encouraged me and told me the truth. I couldn’t have gotten even this far without them.

Maybe I’m wrong. I hope I am. If I am, I’ll be glad to say so here, in this same quiet corner of my website. And I’m not disappearing. I’m still reachable, still in touch with this industry, and still curious where it all goes. It’s just not where I can invest my time and effort anymore.

For now, I’m walking out the same door I came in.