When translation costs hit zero

I recently gave a keynote at the GenAI in Localization conference, where I shared a provocation I’ve been testing at industry events over the past few weeks. I told the audience that our industry needs to stop defining itself through the language of translation and localization. When we lead with those terms in conversations with business stakeholders, we have already lost their attention.

When I made the same argument at LocWorld, half the room nodded and the other half looked genuinely shocked. At this session, the reaction came through the chat, and it was just as divisive. And what followed, in the Q&A and in the conversations, was exactly the kind of exchange I was hoping to provoke.

So here’s the argument, and what came back.

The commodity trap

There is now research confirming what many of us have sensed but haven’t always been comfortable acknowledging. The cost of translation is heading toward zero. The models keep improving, the unit economics keep falling, and the trajectory is clear.

If we as an industry continue to define our value by the cost of moving content from one language to another, we reduce ourselves to a line item. And business leaders have a predictable response to line items. They try to drive them as close to zero as possible.

This happens in customer conversations all the time. Engineers build a translation feature over a weekend, the results look passable, and someone in the room says, “It looks like Spanish.” And for anyone who doesn’t speak Spanish, that can pass as good enough.

The same dynamic applies to quality. When we lead with BLEU scores and COMET and edit distance, it’s the equivalent of an engineer walking into a CEO’s office and talking about beautiful code. The CEO doesn’t care about beautiful code. They want the software to ship. They want to know if it sold anything.

If the game is selling hotel rooms, the question is whether our content is selling hotel rooms. If the aim is to warn users about a zero-day vulnerability, the question is whether people understood the warning. These are business outcomes, and they have very little to do with the metrics our industry has historically used to describe its own value.

The model isn’t where the value lives

I spent a good portion of the keynote on this point because I think it’s one of the most misunderstood dynamics in our space right now. For the past two years, the conversation has been dominated by models. Which model is best? Will models render everything else obsolete? Do you need to build your own?

Even the model companies are now reaching for context. Claude recently launched a Slack integration. Why? Because Slack is where company context lives. The signal is becoming very clear.

Research from Stanford that I’ve mentioned before, (Meta-Harness, Lee et al., 2026) published in March 2026, demonstrated that the same model can produce performance differences of up to six times depending on harness design. The model stays constant. Only the scaffolding changes.

This matters for every brand that’s using AI to communicate. If Nike, Adidas, Puma, and Reebok all use the same model with no differentiation layer, they all sound the same. That’s a disaster. The moat isn’t the model. It’s the harness around it.

And here’s what’s encouraging. As the underlying models get better, the harness performs better too. It doesn’t get swallowed up. The investment in context, in brand-specific intelligence, appreciates over time rather than depreciating.

“How can we sell this internally?”

This was the first question from the audience of localization professionals and the most honest one. The host put it directly. We’ve been selling localization internally for a long time. The vision you’re describing is way beyond localization. It puts us up against engineering teams who think they can build this in a weekend and AI platforms adding language as an afterthought. How do we sell it? And to whom?

It’s worth acknowledging how hard it is to break habits. I catch myself saying “localization” constantly, sometimes deliberately when I’m talking about specific roles and departments, and sometimes just out of muscle memory. So I’m not pretending this is easy.

But I think our industry faces some very strong headwinds. Language teams across the industry are under pressure, and it’s almost always because the business underestimates what getting language right actually involves. The path out of this is to sell a different version back to the business. Language is the programming language of people. If you want an emotional connection with consumers or business users, you cannot get there with raw machine translation and zero thought about nuance, tone, or cultural context.

The headless distribution model helps to change this conversation meaningfully. Instead of asking the business to follow a process and comply with standards, a localization team can now say to any function, “Here’s an MCP connection. Hook it up to whatever system, and you have context, governance, and effective language capability at your fingertips.” The business doesn’t have to change its workflow. The language intelligence just flows through in the background. You go from playing compliance officer to adding value invisibly, to everyone, all the time.

The simplicity question

One of the most revealing questions came from the live stream. Some language technology products have scaled through simplicity. Agent builders, orchestrators, and development environments are inherently more technical and complex. Can they achieve the same level of growth?

I think there are two things going on here. The first is that companies betting their entire strategy on a single model face a structural challenge. The world is not unimodel. We aggregate models at Phrase and can see from customer usage data that model consumption is not unipolar. No single model wins every use case every day. And as the cost of frontier models rises ahead of potential IPOs, there’s a growing commercial incentive to evaluate open-source and on-premise alternatives for specific workloads.

The second point is that simplicity and complexity aren’t opposites in the way the question implies. Context and harness engineering are genuinely complex. But the user experience on top of that complexity can be, and should be, simple. You can have a team that reaches into the mechanics of these systems and configures extraordinary capability, and then users who have no idea what’s happening in the background and never need to. The user experience should feel conversational. People should be able to ask for what they need in simple language, while the platform handles the complexity behind the scenes. That’s a sensible architecture, and it’s how most great software already works.

What happens when you try to scale a second brain

A question from the audience about personal knowledge management led somewhere I hadn’t planned to go, but it turned out to be one of the more interesting threads of the session.

I’ve been building a personal knowledge system, an LLM-connected wiki that now has around 18,000 pages, linked to my agents through a memory layer. When it works well, and it does take continuous effort to maintain, the experience is transformative. The agents I interact with daily have massive context on my business, my conversations, and my decision history.

But here’s the unsolved problem. Individual knowledge tools are getting quite good. What nobody has cracked is how you layer those together across a company. If I have a brain and my colleague has a brain, how do you link them? How do you respect privacy? How do you stack individual, team, department, and company-level memory with proper role-based access control, auditability, rollback, and injection hardening?

Every company I’ve seen claiming to do this at scale is actually running one shared brain for a team of 10 to 30 people. That’s not the same thing. The architecture for structured, multi-layered organizational memory with appropriate governance doesn’t exist yet. I’ve spoken with AI researchers who are working on it, and I think breakthroughs are coming. But as of today, this is one of the most significant open problems in enterprise AI.

Humans aren’t going anywhere

Someone in the chat offered the view that there would be no need for human post-editing going forward. I wanted to address this directly because I think the nuance matters.

Post-editing as a percentage of total content will decline. The machines keep getting better, and as they do, businesses will gain confidence to push more content through automated pipelines. That is simply the trajectory.

But a declining percentage of a rapidly growing total may still mean growing absolute volumes of human post-editing. We haven’t yet seen the content explosion I predicted in 2023, but I still believe it’s coming, driven by Jevons Paradox, the principle that when you reduce the unit cost of something, people use more of it and total demand grows. As unit costs fall, demand expands, and the total volume of content that needs to exist in multiple languages will grow substantially.

More importantly, the people who are skilled at post-editing today have an adjacent capability that’s becoming increasingly valuable. They understand quality at a granular level. They know how to evaluate whether something works. Those skills translate directly into running and tuning the harness itself, doing evaluation, setting guardrails, managing drift when underlying models change. Humans need to be involved at both ends of the pipeline, and the role is shifting upstream.

The vocabulary you use decides your future

Finally, one of our customers presented recently about how they now generate a third of their global revenue from international markets they weren’t in before. Another has massively reduced customer support volume through better multilingual content. These results aren’t measured in cost per word. They’re measured in revenue, market access, and operational efficiency.

I want to be precise about one thing. Driving down cost per word still matters, because of Jevons Paradox. Cheaper unit economics unlock more volume, more use cases, and a larger total addressable market. But it’s not the whole game. It’s an enabler. The game is the intelligence, the context, and the compounding value that sit on top.

The commodity trap is real. The question isn’t whether it exists, it’s whether we walk into it or around it. I think the answer is to stop describing ourselves in the vocabulary of the past, start talking about business outcomes in the language of the boardroom, and build systems that make language intelligence an appreciating asset rather than a declining cost.

That’s the conversation I wanted to start. Based on the response in the room and the questions that followed, I think it’s one the industry is ready to have.