Lean Translation — A Working Framework

This is a working document. It describes the principles behind how we build Codex and LangQuest at Frontier R&D. We're publishing it because we think the ideas are more useful shared than kept internal, and because we'd rather refine them in the open.

---

The Problem with Adding AI to a Broken Process

In The Goal, Eliyahu Goldratt tells the story of a factory that installs robots on the production line and celebrates a thirty-six percent efficiency gain. But throughput doesn't improve. Inventories pile up. Delivery times get worse. The robots made one station faster while the rest of the system stayed the same—or got worse, because now the bottleneck shifted downstream where nobody was watching.

Most of the translation industry is making the same mistake with AI.

The standard translation workflow was designed in the 1990s: a project manager receives content, assigns it to translators, collects output, sends it to reviewers, routes corrections back, manages a translation memory database, and eventually publishes. It's a batch production system. Content enters in batches. It moves through stations. Each station has specialists. The whole thing is coordinated by a project manager who keeps the line moving.

When AI enters this picture, it usually enters as a faster station—the same way Goldratt's robots entered the factory. Machine translation produces a draft. The draft gets post-edited. The post-edit gets reviewed. The review gets QA'd. The same stations, in the same order, with the same batch logic. One step got faster. The rest absorbed the pressure.

If your review team was already the bottleneck, producing drafts faster just makes the pile higher. If your translation memory was already full of inconsistent decisions from different translators across different years, AI that learns from it inherits every contradiction. If your quality process was already struggling to keep up, accelerating the front of the line doesn't help the back.

We wrote about this. But the diagnosis only gets you so far. The question is: what's the alternative?

Where the Principles Come From

The alternative comes from a pattern that's been proven in other industries, repeatedly, over decades.

Lean manufacturing didn't just make factories faster. It rethought what a factory is for. Toyota's insight wasn't "build cars quicker." It was: eliminate everything that doesn't contribute to the final product, shorten the distance between a decision and its feedback, and trust the person closest to the work to make the call.

Lean startup applied the same thinking to product development. Don't build the whole thing and then test. Build the smallest thing you can learn from, learn, and adjust. Shorten the loop. Reduce waste. Optimize for learning speed, not output volume.

Both of these disciplines share a set of core principles:

1. Shorten feedback loops. The longer it takes to discover a problem, the more expensive it is to fix.
2. Eliminate waste. Any step that doesn't create value for the end user is overhead.
3. Empower the person closest to the work. Decisions made by the person doing the work are faster and better informed than decisions made by someone coordinating from a distance.
4. Move from batch to continuous flow. Batch processes create queues, delays, and inventory. Continuous flow creates momentum.
5. Optimize for learning, not output. A system that learns faster wins, even if it produces less in any given sprint.

These principles have never been applied to translation as a coherent methodology. The phrase "lean translation" has appeared in a few blog posts and conference talks, but there is no defined discipline—no framework, no set of principles, no shared vocabulary. The conceptual territory is open.

We think it's time to fill it in.

What Lean Translation Looks Like

Here's how the old model maps to the new one.

Traditional Translation	Lean Translation
Batch production: content moves through stations	Continuous flow: translation happens in a single pass
Translation memory: static database of past output	Living memory: evolving knowledge that learns from every decision
Large specialized teams: translator, reviewer, QA, PM	Empowered expert: one person with deep knowledge, amplified by AI
Quality inspected after production	Quality created during translation
Feedback arrives days or weeks later	Feedback arrives in real time, with every keystroke
Learning means accumulating matches over years	Learning means the system improves with every validated decision

The shift isn't incremental. It's structural. Traditional translation optimizes process. Lean translation optimizes judgment.

The Five Principles

1. Continuous feedback instead of post-hoc review

In a traditional workflow, quality is something you check after the translation is done. A reviewer reads the output, flags problems, sends it back. The translator—if they're still available—makes corrections. The loop can take days.

In lean translation, feedback happens during the work. The AI responds to the translator's decisions in real time. If the translator changes a term, the system immediately reflects that change in future suggestions. If a passage deviates from the project's voice, the system can surface it before the translator moves on. The cost of catching a mistake drops to nearly zero because the distance between the decision and the feedback is nearly zero.

2. Expertise amplification instead of workflow specialization

The traditional model splits translation into roles: translator, reviewer, quality assurance, terminology manager, project manager. Each role handles a slice of the work. Coordination between them is expensive—meetings, handoffs, style guides that nobody reads twice.

Lean translation inverts this. Instead of distributing the work across specialized roles, you amplify and empower a specialized person, ideally the one with the most relevant project expertise. A subject-matter expert—someone who understands the source language, the target community, and the intent of the text—works directly with an AI system that has internalized the project's history, voice, and decisions. The expert steers. The AI handles the mechanical load. The roles and layers of tooling and procedure that existed to coordinate between specialists become unnecessary, because the work isn't split anymore.

3. Contextual memory instead of string reuse

Translation memory stores outputs: this source string was translated to this target string. It has no concept of why that decision was made, what the surrounding context was, or whether the decision still makes sense today. It matches strings. That's its entire function.

Contextual memory stores the conditions under which decisions were made. It knows the linguistic and extralinguistic context. It knows the project's stated philosophy and audience. It knows what the team has validated and what remains unconfirmed. When the AI draws on this memory, it doesn't find a match—it reconstructs the team's reasoning and applies it to new material. The memory evolves with the project instead of accumulating as a static archive.

4. Flow over batch production

Batch production creates inventory. In translation, inventory is untranslated content waiting in a queue, reviewed content waiting for QA, QA'd content waiting for publication. Each handoff creates a delay. Each delay creates an opportunity for context to be lost, priorities to shift, and motivation to erode.

Flow means translation moves from source to publication in a continuous stream. The expert translates, the system provides real-time feedback, validation happens as part of the work—not as a separate phase. There is no queue between translation and review, because there is no separate review step—it is all integrated. Quality is built in. The output is ready when the translator says it's ready.

5. Quality integrated into the translation process, not inspected afterward

This is the principle that ties the others together. In lean manufacturing, the insight is that you can't inspect quality into a product. If the process produces defects, catching them later is expensive and unreliable. You have to build quality into the process itself.

The same is true of translation. Post-hoc review catches some errors, but it can't fix a passage that lost its coherence three handoffs ago, or restore a voice that was diluted by five different translators over four years. By the time a reviewer sees the output, the decisions that shaped it are already made.

Lean translation moves quality upstream. The expert works with full context, real-time AI feedback, and a living memory of the project's intent. Decisions are informed the moment they're made. Quality isn't a gate at the end of the line. It's a property of how the work is done.

Why Now

These principles weren't possible five years ago. Translation memory was the best available technology for leveraging past work, and batch workflows were the rational response to the coordination costs of large teams.

Two things changed.

First, AI became good enough at language to be a genuine collaborator—not just a draft generator (though many implementations stop here), but a system that can learn from context, follow a voice, and improve with feedback. That makes expertise amplification possible. One person with the right knowledge, working with the right system, can do what used to require a team.

Second, the cost of content production dropped to near zero. As we've argued before, creating content was never the real bottleneck—trust, review, and coordination were. But now that production is genuinely cheap, those bottlenecks are no longer hidden. They're the whole problem. And they're the exact problems that batch workflows and translation memory were never designed to solve.

Lean translation is the response. Not a faster version of the old process, but a different process—one designed around the constraints that actually matter now.

What This Means in Practice

We're not claiming we've finished building this. We're saying it's the framework we're building toward—the set of principles that guides our decisions about what Codex and LangQuest should be.

Some of it is already working. Codex's living memory, real-time AI feedback, and translation health are direct implementations of these principles. LangQuest applies the same thinking to language documentation—empowering native speakers to document their own languages in a continuous flow, rather than waiting for linguists to run batch fieldwork sessions.

Some of it is still aspirational. We're publishing the framework now because we think the ideas are worth discussing, even before the implementation is complete.

The core claim is this: translation becomes continuous knowledge steering, not document production. The expert steers. The system learns. Quality is built in, not bolted on. And the teams that adopt this way of working will produce better translations, faster, with smaller teams and higher confidence—not because AI replaced anyone, but because AI finally made it possible to work the way translation always should have worked.

The old process wasn't anyone's fault. It was the best response to the constraints of the time. The constraints changed. The process should too.

Related: Ninety-Eight Percent Match Rate, Creating Content Was Never the Bottleneck, Translation Health in the Age of AI