Field Notes: Economies of Scale vs. Assurance in Localization

Join Stephanie Harris-Yee and Erik Vogt for a focused discussion on the shifting economics of the localization industry. As AI drives down the fundamental cost of language production, the financial models governing translation have split into two distinct categories: scale and assurance. Discover why a lower unit cost often results in higher overall spend, and learn how to properly evaluate content risk to avoid misallocating your localization budget.

Key topics covered:

• The divergence of localization economics into the volume-driven “scale curve” and the risk-driven “assurance curve”
• Why the total cost of ownership expands as automation drives translation unit prices closer to zero
• Treating high-risk translation as liability mitigation to prevent compounding errors, contaminated training data, and brand damage
• The risks of misallocating resources, such as applying human review to short-lifecycle content or full automation to critical materials
• Strategies for workflow orchestration based on content differentiation and the specific consequences of failure
• Moving away from outdated cost-per-word metrics to optimize for token consumption and cost-per-outcome
• The impact of Baumol’s cost disease on the rising expense and value of human expertise

More “Field Notes” Episodes

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

• The Governance Problem in AI Localization
• SME Bottleneck
• How to Price Quality
• Metadata
• Solutions Design
• Connectors
• Are Fuzzy Matches Dead?
• AI LQA
• From Idea to AI

Field Notes – Episode 10: Economies of Scale vs. Assurance in Localization

Stephanie Harris-Yee: [00:00:00] Hi, I’m Stephanie and I’m here again with Erik to talk about a new subject today. And that is something we like to break into two buckets. Scale versus assurance, in localization. Now, before we jump into that, let’s go ahead and set some groundwork.

So, AI has made reasonable quality translation dramatically cheaper, so shouldn’t that just mean that localization overall becomes cheaper now or should cost less now?

Erik Vogt: It should, but it doesn’t, because while the cost of producing language has dropped the cost of getting it right hasn’t. So instead of one cost curve, we have two. So we’re talking about economic cost curve here, where there’s different market forces affecting different parts of the industry. And this has been going on for a while.

So this shouldn’t confuse our audience but, I think, what we’re gonna be talking today is exactly what those cost curves are, and for better awareness of where people fit within those cost [00:01:00] curves will allow a better amplification of their core value proposition. So that’s the essence of what we’re gonna get to today.

Stephanie Harris-Yee: Okay, let’s jump off with that first cost curve. So let’s call this the “scale” curve. What does that look like?

Erik Vogt: So, the “scale” curve in microeconomics nomenclature, the lower you reduce the cost of a unit of an object, the more you increase demand. So if you can make your phones cheaper, more people will buy them. So we end up with this weird phenomenon where, and again, this is not news to anybody in our audience, but as the cost of translation approaches zero, then the amount that gets automatically translated explodes.

And that even small incremental costs ends up being a lot. And so we end up with some interesting phenomenon where, say, a team is translating a lot more, they’re getting a better unit price, but also, their total cost is actually going up. So you end up with a bigger [00:02:00] sort of overall spend. So the overall, the spend and the unit price are actually inversely related in some ways.

So the total output explodes and then you end up with an intense optimization pressure to lower the unit price even more, because that’s the one constraint left to continuing to grow and, that can show up by the way, not just in the technology that’s doing the translation itself, like the LLM or an MT system, but all the infrastructure that supports that delivery of that model.

So it’s shifts the economics tremendously. And so when we have a lower kind of unit price, we also can support more markets. We can support more content types, and then you end up with content that’s generated at scale, now you can translate it almost instantly into 60 languages or something, in ways that literally were nothing we even thought about being possible

10, 15 years ago, it just simply wasn’t even part of the equation, and yet now the total cost of ownership for this [00:03:00] particular submarket is still getting the attention of CFOs as they try to figure out how to manage this cost effectively.

If you thought about one way of summarizing, it’s like AI didn’t reduce localization demand, it just expanded the surface area.

So now there’s a lot more contact points, where translation demand can be created and where the supply of it can be increased.

Stephanie Harris-Yee: Okay. So that gives us a good idea then of that scale. What about the second one? So that assurance curve how is that different? What does that look like?

Erik Vogt: So I, I think this is the implicit space that most LSPs naturally live in, which is, what I’m calling the “assurance economics.” So I think this is more high risk content that doesn’t follow that pattern at all. It’s easy to go with the legal, medical, critical, brand sensitive type stuff, but it also is just as much true for the creating bad training data, t hat leads to underperforming models. And so the error cost, the cost of mistakes is high in this [00:04:00] subcategory, and then that is more important than the production cost itself. So, when you think about the consequences of a bad translation, it’s not just, and I’m gonna borrow another kind of business term here called “net present value.”

So, it’s something that’s you add up all the sum of the cost and the value of an object and multiply it by time, then you can end up with a number where that mistake is being replicated for a long time. You can imagine the very simple example of a translator having to fix the same mistake over and over again because, the terminology was wrong, and then that kind of multiplies out into a long tail of extra work, and extra damage, and extra QA steps, and all that stuff.

So that’s another example of where, it’s not just brand value or making a mistake that increases the liability. In this case, the value is not in volume, but in avoiding costly mistakes, like a contaminated TM or kind of a brand damage or liability.

So translation stops being a cost center, in this case, like, it tends to [00:05:00] behave more like insurance, so people will pay an LSP or a language expert to mitigate the liability. They’re outsourcing that liability and the consequences, they are owned by somebody you’re paying to be accountable, for that consequence.

Stephanie Harris-Yee: So we have these two concepts. When you explain it, it seems pretty clear they’re very different, maybe use cases, different curves going there. Where are companies messing up or where are they getting it wrong when it comes to addressing these two very distinct types of content?

Erik Vogt: There’s been a lot of talk about differentiating content types so that you’re appropriately applying the right cost risk opportunity kind of metrics to it. For example, one, one mistake would be to apply human review in absolutely everything. I think at this stage, if you are having a human translating something that’s a one-off short lifecycle, only a small number of potential users, less intimate context, [00:06:00] then that creates a lot of cost that isn’t converted into meaningful value.

But if you apply automation everywhere and you just translate everything automatically, then you’re increasing your risk exposure and can create liabilities, which can, by the way, not be apparent for a while. It can start adding up. And so, we’re very familiar with this, where somebody switches, switches a hundred percent to say automatic translation, and then suddenly they’ve start building this massive backlog of brand damage. They start hurting their international cells. There’s a lot of finger pointing and it just, anyway, so from an economic standpoint, you don’t want to overpay for low risk content, but you also don’t wanna underinvest in high risk content.

The biggest waste is not something you can think about as a cost. You think about as misallocated cost, where you’re thinking about how you’re spending money and you’re not thinking about how you’re managing risk in ways that really map to the real world.

Stephanie Harris-Yee: So what’s a smart way to manage [00:07:00] both to not make those mistakes?

Erik Vogt: The magic word that people like talking about is orchestration. But I think when you’re talking about orchestration, some people immediately go down to the simple, like workflow management. Oh, just automatically do this one thing. That can be measured in cost reduction in the simple way.

Yeah, you don’t have to manually copy and paste a bunch of things, that’s fine. But I think differentiating your content types is really important and you need to have a conversation about the risk, and it should be an honest conversation. And right now, I’m talking to the localization team managers within large organizations who have a bunch of just different internal stakeholders, but having an honest conversation about the consequences of failure in this case. Like what are the risks here?

What is the value that translating this content is creating? What is the expected impact? From those conversations, then you figure out where should we apply automation? Where should we apply expertise? Like where do we automate? Where do we slow down and make sure that a human is involved?

And then this is [00:08:00] not necessarily a black and white thing. It’s not either or. It’s just two fundamentally different worlds here. And sometimes, for example, it’s worth paying a really excellent quality assurance on an automated process to really deeply say, validate your QE model or make sure that you understand your risk exposure for certain types of things.

So the nuance is really important. This is not just about the systems themselves, but you’re looking specifically at the workflow friction and I’ve, we’ve talked before about this exploding wall of complexity that is addressing a lot of people when you have multiple different teams using different AI in different ways, and they’re just overlapping each other and creating a lot of turbulence. There’s also the human element is really important, the human usability. And I like to look at that both from the standpoint of the, our, the end customer, who’s consuming the translated content, the internal stakeholders within the kind of the architecture of our industry, and then [00:09:00] there’s the individual value creators, or the translators, or reviewers, the subject matter experts that we talked about too.

So I think this is, it keeps on going back to Baumol’s cost disease. That human labor isn’t getting cheaper, it’s the automation gets cheaper if you can automate it, but the humans are actually getting more expensive. So the orchestration is more important than ever of how, and you decide where to spend money and where you don’t.

Stephanie Harris-Yee: So if you were to take this concept and pull out, one takeaway for leaders what would that be? What would be the thing that you would say, remember this?

Erik Vogt: Yeah, the number one is stop optimizing for costs per word, blindly. ’cause if you just assume that it’s about cranking down the price list and getting the lowest unit rate without any real substantive understanding of the cost quality curve, and I think the industry have, and this goes back 25 years, like originally there was like standard translation [00:10:00] rates for UI.

It’s a little more, it costs more for legal. There’s just this multiplier, you just slap on 20% or something and you just get a different multiplier. Then there’s also this fuzzy match grid, which is also made up years and years ago, and it’s this shorthand for effort and risk and time management.

But I think we need to start thinking about another layer to this, which is: Are you optimizing for token consumption? That’s not something we typically think about, but for that high volume thing, token consumption can be a huge issue. Specifically, if you add up all the sum of all the incremental subsystems, if you’re trying to cost optimize for volume, now you need to scrutinize the incremental overhead of,

of data storage of, what is an update cost, like are you resubmitting things that don’t need to be resubmitted for, that are adding up. And I’m talking about literally 20 dollars per million tokens or something. But at the scale that we’re talking about that adds [00:11:00] up. The other thing that people often don’t really have a clear conversation about is the cost per outcome.

So again, if you start talking about unit price per word, and that’s where the conversation starts. If you’re not having a conversation about what happens if this fails, then you can’t apply the appropriate accountability model on top of the way that you’re selling these things to really build a model that delivers the accountability that you’re expecting it to get.

So, taking it up to the big picture again. We have two distinct cost curves. Very often people muddle them up and think about them as the unit price should go down because of automation. To some extent, there is some truth to that. Yes, we can.

Shave down MTPE by getting better quality content to the reviewer. But we don’t wanna over index on that because if you’re paying for accountability, you should mainly focus on making sure that accountability structure is really rock solid. So split it up into scale versus assurance and making sure that you know [00:12:00] the real challenges, knowing which one applies in which case.

Stephanie Harris-Yee: Okay. Words to live by. All right. Thank you, Erik. And yeah, we’ll see you again for the next one.

Erik Vogt: Thank you so much, Steph. I look forward to it.

Share this post