How Doist built an AI first localization engine with 46% production-ready translations

“Today, over 60% of our new acquisitions come from non-English markets, driving over 40% of our revenue. Transifex plays a central role in making that possible. — Daniel Garcia, Senior Growth Marketer at Doist.

Doist is a fully remote, bootstrapped company with around 100 employees across more than 25 countries. Its flagship product, Todoist, is used by more than 50 million people. But the company's relationship with localization wasn't always this strong.

"In 2017, localization execution was very weak," says Daniel Garcia, growth marketer at Doist and the company's localization manager.

"There were no real workflows and processes. There was totally a lack of quality assurance. There was no project management. There was no team management of the freelance translators."

Garcia joined the localization effort from the field. He'd been working as a Spanish translator for the company, bringing a background in journalism and linguistics to the role. That translator's mindset shaped everything that followed.

"My vision was simple. Users should never feel they are reading a translation. It should feel originally crafted in the language."

With a limited budget and no outside funding, the team tiered its markets by analyzing acquisition, activation, and revenue data. They prioritized based on funnel stage impact.

In three years, non-English-speaking acquisition grew from 30% to approximately 40%.

"Localization stopped being a cost center and it became a growth lever," Garcia says.

Building the foundation before AI

Before introducing any AI, Doist spent years building the operational discipline that would make automation possible.

The team developed:

• A robust localization quality assurance (LQA) framework

• Extreme context documentation

• Clear localization values aligned to brand values

• Glossary management

• And translation memory discipline

Garcia also restructured how work reached translators.

"I realized that we needed to structure and assign the work to the translators in advance," he explains.

"Minimizing last minute requests, protecting the translator focus, thinking that they have other clients. That ultimately gives quality to the localization."

"This foundation is critical because AI is starting to work today because all of these things existed first. Without this foundation, AI would have degraded the quality."

That discipline paid off.

Doist grew its international revenue share to between 40% and 45%, putting it in what Garcia describes as "high performer" territory for a company of its size.

The localization bottleneck that led to a strategic shift

By 2024, the cracks started to show. Doist's analysis revealed that:

• 51% of translation volume was support content

• 30% was marketing

• 19% was product UI

A significant share of translator effort was concentrated on structured, informational content.

At the same time, product shipping was accelerating. Localization, once a growth lever, was becoming a bottleneck.

"We stopped asking how do we translate everything well," Garcia says. "We started asking how does human craft actually drive growth."

"And that's how we created the 70/30 model."

The 70/30 model

Doist's solution splits content into two workflows based on funnel stage, not content type.

Around 70% of content types follow an AI first workflow with human LQA.

This includes:

• Product UI

• Help center content

• Educational content

• Video subtitles

• Lifecycle emails.

Transifex AI handles the initial translation, and human reviewers then evaluate in context.

The remaining 30% stays human-led.

This includes:

• Landing pages

• App store listings

• Campaign emails

• Paid campaigns

• Onboarding resources

"Every surface that defines our brand voice externally remains human-led," Garcia explains.

"Most of the acquisition and activation content is human-led plus human LQA. Most of the retention content is AI first plus human LQA. Same languages, same product. It's a different investment logic."

The AI first workflow isn't automatic without control. Transifex AI translates 100% of strings, then human reviewers evaluate in context.

Translators maintain glossaries and translation memory, follow brand and copy guidelines, and provide continuous feedback to the Transifex AI team to improve output quality.

What the numbers show

The results speak for themselves.

During the most recent month of tracking, approximately 3,800 AI translations containing more than 33,000 words were production-ready.

Meaning translators didn't edit them at all. That corresponds to 46% of all AI first content translated by Transifex in Doist's projects.

Among the remaining non-production-ready translations, the team found that the content was nearly perfect according to their quality standards, leading to simpler and faster edits.

"We track the performance on a weekly basis," Garcia says. "We can easily identify which languages and projects need more fine-tuning than others, and this way we can continue improving our style guides."

Fine-tuning style guides proved to be a turning point.

"Since we have improved the style guides, things have been improving very well," he says.

"Essentially we are fine-tuning the responses of AI to the tone, to the voice, to the language, and to the specifics of the different source content. This has led to a 20% increase of production-ready AI translations."

The human side of the transition

Garcia is candid about the challenges. Translators pushed back, and he understood why.

"Many experienced linguists are really concerned about being reduced to post-editors," he says.

"The key difference for us is that we didn't redefine them as cheaper translators. We refined their role as quality engineers, AI curators, cultural guardians."

He's also realistic about AI's limitations.

"AI sometimes flattens the language. It lacks nuance. It homogenizes the voice. It performs poorly in low-resource languages," Garcia acknowledges. "That's why we believe AI must be managed. It must not be adopted blindly."

For Doist, SEO content, emotional marketing copy, brand voice, and cultural adaptation all remain areas where AI alone is too risky.

But for product UI, instructional content, and educational copy, the performance has been strong enough to free up significant time for higher-value work.

What comes next

Garcia sees a path where product UI becomes fully automated with human oversight, and help center content follows close behind. The team continues to fine-tune style guides and track performance weekly across all languages and projects.

"The result is that this is freeing resources for us," he says. "We have more time for marketing. We have more time for brand. We have more time for SEO and we have more time for acquisition and activation, which for us are a key part of our product experience."

His advice for teams considering a similar approach? Build the foundations first.

"If you don't have context discipline, LQA structure, and clear values, I think AI will amplify your weaknesses," Garcia says.

"If you do have the foundations, it will amplify your growth. AI is a scale tool. It doesn't have to be a hype."