Field Notes: Shadow Localization with Erik

Across organizations of all sizes, teams outside of localization departments are building their own translation solutions — a phenomenon known as “shadow localization.” Marketing teams need rapid content turnaround. Product teams are embedding chatbots with translation functions. Support teams are auto-translating knowledge bases and customer conversations. And with the rise of AI-powered automation and no-code tools, the technical barrier to spinning up a translation workflow has effectively disappeared.

This decentralization is not malicious or even intentional. It is driven by legitimate business demands: speed, agility, and the pressure from executives who expect velocity. Creating an arbitrary bottleneck without a clear business justification is no longer viable. However, distributed translation production introduces serious risks — fragmented terminology, duplicated effort, unclear quality ownership, and unmanaged compliance exposure. When no one explicitly owns quality, accountability eventually surfaces at the executive level in the form of uncomfortable surprises.

The solution is not a return to full centralization. Instead, organizations need “connective governance” — a framework that respects the distributed reality while anchoring accountability. Practical approaches include:

• Publishing standards and making content LLM-consumable — providing markdown files and terminology resources that distributed teams can feed into their own models for better output.
• Operating as an assessment layer rather than a gatekeeper — offering quality evaluation services (automated scoring tools, human reviewers) without blocking velocity.
• Playing an integrator and advisor role — actively supporting distributed teams with expertise, tooling guidance, and feedback loops.
• Maintaining a human-in-the-loop accountability layer — ensuring high-stakes content receives expert review and that quality output is clearly flagged as suitable training corpora for future AI models.

Organizations finding success are building content rubrics with multiple tiers — matching content types to appropriate quality workflows and accountability layers. This allows teams to manage risk proportionally rather than applying a single rigid process to all content.

The core shift: localization is not disappearing. It is becoming distributed infrastructure. The discipline must evolve to encompass both a speed track (facilitating the best available automated output) and an accountability track (human expertise ensuring quality where it matters most).

Key Insights

• Shadow localization is demand-driven, not defiant. Teams are building their own translation solutions because business velocity requires it and the technology now makes it trivially easy — not because they are trying to circumvent process.
• The primary risk is governance, not technology. The tools largely work. The danger lies in unclear ownership of quality, fragmented terminology, and unmanaged risk that compounds silently until it becomes a visible executive-level problem.
• Centralization is no longer a viable response. Organizations that attempt to re-consolidate all translation through a single team will simply drive more activity underground. Connective governance — standards, assessment layers, and advisory roles — replaces gatekeeping.
• A dual-track model balances speed and accountability. Successful organizations maintain both an automated speed layer for lower-risk content and a human-in-the-loop layer for high-stakes output, with clear rubrics determining which content flows through which track.

More “Field Notes” Episodes

Explore more topics in our Field Notes series, where we break down complex localization concepts, ideas, and experiments for industry professionals. Check out our other discussions here.

Field Notes – Episode 12: Shadow Localization with Erik

Stephanie Harris-Yee: ​Hello I’m Stephanie and here for another episode of Field Notes. And today it’s kind of exciting because as many of you may know, last week we just started our Companion Field Notes series with Giulia Greco, and today we’re going to do a companion episode with Erik talking about the similar topic.

So this topic is something that we’ve been hearing about a lot, and this is where teams are building their own translation solutions, sometimes without even involving localization at all. So Erik, from your perspective, what’s going on?

Erik Vogt: it was really amazing seeing Giulia’s input on this because she’s raising something that I’ve also been noticing as well, which is this decentralization pressure that we’re seeing. So because localization can be embedded this is not necessarily a brand-new problem, but organizations would either have different teams doing their own localization vendors and handling things their own way. This is a legacy of M&A activity or different sort of business purposes. You might have marketing and product working with different vendors, then they consolidate, and then they break it apart again. It’s a, it’s a not that uncommon of of a phenomenon. But because we’re actually seeing a new thing happening here where you can not only direct localization traffic via different channels, which we talked about last time, but also we’re seeing this pressure to add this capability in a much faster, more agile way in different places within an organization at the same time.

So it’s a very interesting phenomenon, and it’s hitting a much broader range of companies than I ever expected that would happen. It’s happening faster than I would have expected

Stephanie Harris-Yee: So we have now teams are somewhat, you know, localization is escaping the localization department. Why are teams doing this, do you think, instead of going through the typical localization route?

Erik Vogt: I think what we’re seeing is the demand is there, right? So, I think Giulia pointed this out as well, but th-there’s a demand, and the capability is there. So these teams are

Stephanie Harris-Yee: Yeah.

Erik Vogt: just reacting to these fundamental demands. So marketing team, for example, has a need to get things out very quickly. You know that. Sometimes it can be very powerful to use AI to create content or help shape the the understanding of what the market signal is and being able to bring that signal in and respond to it in a much more effective way. Meanwhile, the product teams are sometimes being empowered to or sometimes executives are demanding that they add a chatbot into there.

So now they have to add a, an AI capabilities like with some translation functions directly into the product. And then you might have support teams that are dealing with auto-translating content or which is sometimes end-user

Stephanie Harris-Yee: conversations with their users and their support

Erik Vogt: ecosystems or sometimes their knowledge base. And then of course, we have on top of that this crazy vibe coding reality where teams are needing to implement lightweight scripts or plug-and-play tools,

but we hear people joking about vibe coding a CMS over the weekend.

Stephanie Harris-Yee: little bit of a joke

Erik Vogt: Obviously it is. Not real not realistic. Please don’t try try this at home. But the mindset is real now that we have these powerful automation tools and we have individuals who are building hundred million dollar businesses with no code experience, they’re building an app themselves and then generating

like a million revenue, it’s unbelievable how fast this is all changing. So executives are definitely expecting fast and demanding velocity. The business is demanding that velocity. So I think it’s just realistically it’s not okay to create an arbitrary bottleneck if there isn’t a legitimate business reason to do so I think, yes, localization is escaping the localization department. But it’s actually it’s actually being driven by very meaningful business drivers underneath.

Stephanie Harris-Yee: So then where does this become a issue? So people are doing it. Is it an issue? When does it become an issue? What do you think?

Erik Vogt: I think it is when you have a distributed production system, then anybody who understands the consequences of poor terminology control will instantly see that as a major area of concern. You have fragmented systems, but I think these are technical things that we’ve been familiar with for decades. But I think one of the key problems is the ownership and the governance aspects of it, which it’s not clear who owns quality in this case. So if you’re thinking about this as a localization manager and somebody’s vibe coding some new tools and then packing in maybe your content without talking to you about it as part of some AI solution that they’re building, and then it spits out some less than ideal output. what happens

Stephanie Harris-Yee: Yeah. Yeah.

Erik Vogt: to the localization owners of this? How do they handle this ambiguity? Who’s accountable for fixing it? So without that clear ownership, then you’re gonna end up with inconsistent terminology, and there’s of course duplicated effort. But the big one that ties into what we’ve talked about before is the unmanaged risk. They’re not consciously managing that risk. So you have this demand force which is driving towards less regulated, less controlled output, and then you don’t really know or understand how you’re managing that risk. So what ends up happening is that, even though the technology could be owned locally or by individual teams, the accountability ultimately gurgles up to the CEO at the top, right.

We’re trying to align the ownership and the governance around each of these kind of emerging systems and you don’t really have the clear feedback loops, the consequences are gonna pop up later and you’re gonna end up with a lot of uncomfortable surprises. So I think this is really what we’re talking about, what Giulia was calling shadow localization. And I think it’s not necessarily that people are [00:06:00] trying to break the system on purpose. none of this is intentional. But because the market demands that we talked about earlier are forcing innovation, essentially trial and error sometimes very just run fast and break things mindset that you don’t necessarily see the damage you’re doing until later. But I think what we’re seeing is this kind of need to think about that responsive layer, to support business need that’s emerging.

Stephanie Harris-Yee: So then as an organization or a company, what do you do?

Erik Vogt: I don’t think that it’s gonna work to go back to a fully centralized system. It’s just not gonna happen.

Stephanie Harris-Yee: Yeah.

Erik Vogt: that creates an arbitrary bottleneck. But I think having a conversation about connective governance is like your starting point. So I think each of the… this is putting myself into the, to the shoes of individual owners of different systems. They often don’t feel empowered to really have that broad [00:07:00] level

of conversation. So I just wanna be realistic here. This isn’t, this is an easy thing to say, but publishing standards, for example centralizing knowledge one thought would be to make your content consumable by LLMs directly. So that’s bring your own LLM model.

So there’s a new emerging thing here where you’re basically trying to adapt or allow your content to be adapted to in this new AI ecosystem. In which case, you as the team are saying, “Hey, here’s the here’s the markdown files you should use when you’re running your model that we’re gonna shape for you so that you’re gonna have the best chance of getting better results from this.” there could be better, clearer rules of engagement. I think some companies have said, ” Go ahead and do whatever localization you want but we will be an assessment layer. We’re not gonna be a gatekeeper, but we’ll be maybe an assessment layer. So if you’re not sure what the quality is, we can do an assessment.

Maybe we’ll plug in TAUS or ModelFront or something like that as part of kind of this monitoring layer. Or we can have human reviewers look at the content and say, are you gonna hit any red flags here?'” You can also be an integrator and playing an integration role means actively looking for ways of supporting those folks out there.

Like how can you supply them with the best chance of being successful without creating this this stuff? So basically, you’re offering yourself to be a kind of an advisor and accountability partner, and then being an expert in the loop. And then, of course, that if you wanna get it right, and that’s again where the human arbitrage piece comes into it, that’s where we want the human in the loop to be and the team needs to make suret hat is getting the right answer.

That’s for all the gold standard output, that’s for all the training corpora. That’s another piece of metadata, by the way, that I think we’ve had or should be having in the loop here, which is: this content consumable by any earning efforts, or is this an output that we don’t wanna learn from?

So empty output is generally not something you wanna train on so the accountability layer is about delivering high-quality content you can stand behind as training corpora.

Stephanie Harris-Yee: Okay. So you’ve gone over like several points throughout this, but what would be your top takeaway? Remember this one thing.

Erik Vogt: Localization isn’t going anywhere, but it is becoming distributed infrastructure. And as we talked about last time, there’s these two tracks. One is the facilitating the automated translation to get the best available option a human-in-the-loop accountability layer, and then you’ve got just both of them. There’s the speed, and there’s accountability. So make sure you have both of these

Stephanie Harris-Yee: Yeah.

Erik Vogt: Think differently about this. And then some companies are making a rubric of their content with maybe even three or four layers deep of different content types and different sort of accountability layers, and they’ll create intentional workflows for each of work types [00:10:00] across the different teams.

So you’re not saying everything needs to go through this process and get to a hundred percent mitigating the risk with the right approach that balances cost and quality outcomes. So you’re managing risk,

You recognize that shadow localization’s gonna happen if you create unrealistic barriers to managing that content and then you end up with a governance

Stephanie Harris-Yee: Yeah.

Erik Vogt: So accept the reality that this is the direction things are going and think about ways of shifting towards connecting systems, deciding decision ownership, and anchoring accountability.

Stephanie Harris-Yee: Okay. All right, Erik. Well, thank you very much.

Erik Vogt: always a pleasure, Steph.

Share this post