Human-Centered AI Workflows in Localization with Shashi Bhushan

AI is transforming localization, but not in the way many fear. Despite the hype, the real opportunity isn’t replacing translators. It’s building smarter, cleaner workflows that eliminate repetitive work and free humans to focus on quality and creativity.

In a recent episode of The Agile Localization Podcast, host Stefan Huyghe sat down with Localization Workflow Strategist Shashi Bhushan to unpack how to design AI-enhanced workflows that remain deeply human-centered. Drawing from experience at Marvel, Google, Amazon, and now as an independent strategist, Shashi explains where AI genuinely fits, and where it doesn’t.

Listen to the new episode on:

Start With Your Workflow

Don’t start by choosing AI tools. Start by mapping your workflow.

Shashi compares localization to cooking, where you follow a specific sequence and use the right tools at the right moment. Before introducing AI, teams must understand:

How content enters the TMS
What types of content are they localizing
Who touches the content and when
Where the real bottlenecks and repetitive steps are

Only then can you decide where AI adds value. Sometimes the answer is: it doesn’t. Small, low-frequency content often doesn’t justify AI. However, large-scale, continuous product localization absolutely does.

The foundation is always the same: map, diagnose, and then introduce AI.

"

Back in the days when I was working for Google projects and for Amazon also, we were doing things with machine learning and all, but it was not basically not AI agents or full-fledged AI integration into workflows. But when I started using Crowdin three years back, then came the first full-fledged AI workflow thing. So what we did so I thank Crowdin for that because they have been doing a lot of work in the back end. They have been improving the workflow, cutting with agents.

Start With the AI Source Text Agent

Most teams’ first instinct is to use AI for translation. Shashi recommends the opposite: Deploy the AI source text agent first.

Why? Because any inconsistency or typo in the source language multiplies across all languages. A source-text AI agent can immediately:

Flag typos, grammar issues, and glossary mismatches
Enforce style guide rules
Alert project managers before problems cascade
Automatically comment on problematic segments

Fixing issues once at the source saves you from fixing them 20 times downstream. It’s the lowest-risk, highest-impact AI deployment you can make.

Build a Workflow Where AI Assists

Shashi’s approach keeps humans central, not optional. He explains that AI should never be used “just because it exists”. It should only appear where it meaningfully reduces repetitive work. A healthy human-in-the-loop workflow looks like this:

1. AI Pre-Translation (only where appropriate)

Use an AI translation agent for high-volume, lower-risk content. Skip it for legal, highly visible UI strings or content requiring nuance.

2. Human Review as the Core Quality Layer

Linguists fix nuance, ensure cultural fit, and validate terminology.

3. AI Proofreading After Human Review

An AI proofreader agent acts as a final safety check, spotting inconsistencies and leaving comments like a real reviewer.

4. AI Glossary Agent for the Biggest Pain Point

Glossary checks are one of the most frustrating tasks for linguists. A dedicated glossary agent will eliminate hours of manual checking.

5. Final Human Correction/Approval

The language lead or reviewer makes corrections based on the AI proofreader’s comments and gives the final approval before the content goes live.

"

And, again, I mean, the final check before it goes to live, the language lead or the translators or the reviewers, they will have this visibility to correct those things.

Involve Product and Engineering From Day One

Most quality issues have nothing to do with translators. They come from localization being looped in too late. Shashi urges teams to integrate localization at the beginning of the product lifecycle:

Join discussions when source content is created
Review early drafts
Understand context, intent, and UI constraints
Use tools like Figma integration to see how localized text fits the interface

When localization is an afterthought, you get last-minute requests, missing context, and UI overflows on launch day. When it’s built from the start, AI workflows become cleaner, faster, and far more predictable.

Start With a Pilot Project and Clear Metrics

Shashi strongly recommends piloting AI in small, controlled phases. A good pilot includes:

One content type
One or two language pairs
One or two AI agents
Clear metrics

After the pilot:

Fix workflow gaps
Expand slowly
Reevaluate model choices
Then scale to all languages or products

This de-risks AI adoption and gives management confidence through data, not hype.

Final Thoughts From Shashi

If there’s a single takeaway from Shashi’s approach, it’s this: AI should serve humans, not the other way around.

The winning strategy isn’t about automating more; it’s about automating smarter. Remember:

Start with the source
Eliminate repetitive tasks
Protect privacy
Collaborate early with product teams
Evolve through pilots and feedback loops

When you do that, you get better quality, happier translators, and a localization function that scales sustainably.

Shashi’s Background

Shashi Bhushan is a Localization Workflow Strategist and freelance consultant with vast experience designing AI-enhanced localization processes at enterprise-scale organizations. Originally trained as a journalist in India, Shashi transitioned into localization through work on Marvel Comics content and held key vendor roles at Google and Amazon, where he managed localization across multiple languages and products, including Google Maps, Search, and Payments. His expertise centers on workflow mapping and human-in-the-loop AI integration, specifically designing systems where AI agents handle repetitive quality tasks.

Listen to the new episode on: