#287 How the Market for AI Data Has Become a Major Growth Opportunity

0:00Anna Wyndham:It's clear that it's not just a data industry. It's an industry that transforms human knowledge and judgment and expertise into the signals that AI models can learn from and need to learn from.

0:15Florian:Hey everyone and welcome to a very special episode of SlaterPod today. So today we are not talking about localization, translation, interpreting or anything in that area. Today, we are talking about the data that powers AI. We're talking about the companies that are buying it, the companies that are selling it, and what exactly this entire ecosystem is doing. So who better to have on today than our very own head of research, Anna Windham. Hi, Anna.

0:41Anna Wyndham:Hi, Florian. Hi, everyone.

0:43Florian:So you were in the lead for our new data for AI market report. So let me just give you a quick run through kind of the high level bullet points before I defer to you on the various aspects that we covered in this report. So we, in this report, we sized this market at approximately$9.3 billion. We looked at the data used to train, adapt, align, evaluate, and deploy AI systems. We mapped the buyer landscape, supplier ecosystem, the workflows, workforce, and the market dynamics. And finally, we also found that this industry really has evolved, of course, beyond traditional data labeling, you know, like the type of stuff that companies like Appen used to do, and into a much broader kind of ecosystem focused on making AI deployable in, you know, real business workflows.

1:39Florian:So before we start, I do need to give a quick plug for SlateCon San Francisco, because among the many amazing speakers, we also feature Emily Hsu, who's the head of enterprise AI at ScaleAI, arguably one of the leading providers in this data for AI ecosystem. them. So let's jump right in. And maybe Anna, before we start with kind of the core content of the report, tell us a bit more about your personal background and how this is relevant for this particular piece of research.

2:10Anna Wyndham:Yeah, so I started my career actually in the data for AI market working at Appen, mainly focusing on speech recognition, so speech to text. And that is really where the data annotation or data labeling market began. It's changed a lot since then. And as you said, we've tried to capture in the report what this market looks like now in mid-2026. It's a pretty broad and deep report, but the goal was really to understand what this market looks like now and the extent to which it's changed.

2:48Florian:Got it. So what made us, since later here decide to dedicate a report to data for AI?

2:55Anna Wyndham:Yeah, so I guess many of our listeners are language solutions integrators, and this is an important adjacent market for language solutions integrators. We did an analysis of how many LSIs actually operate in this market, and it's around 10%. But notably, a lot of the players at the top of the market, so the very large super agencies, RWS TransPerfect operate in this and they have significant data for AI businesses and notably fast-growing businesses. And there are obviously synergies between language solutions and data for AI, multilingual expertise and global operations. But as we dug into this, it became really clear that this market has much broader relevance.

3:38Anna Wyndham:So beyond being an adjacent market to language solutions integrators. So this is the market, as you said, that produces the data that is used to build, adapt, align, and evaluate AI models. So in that sense, it's a critical layer of AI infrastructure, and it underpins language AI, so all of the language solutions market, but also AI deployment across the broader economy.

4:04Florian:So we covered this, I remember, I think it was back in 2019, right? When Appen was like, I don't know, a$5 billion company. That's no longer the case because while they had to scale back on some fronts, others kind of exploded. But yeah, we did a report back in 2019. What surprised you the most when you came back to it now?

4:26Anna Wyndham:I guess what surprised me is how much the center of gravity has shifted and how little I recognized, how little was familiar in terms of the basic questions, who buys data for AI? What kinds of data matter most? Where is value being created? What kind of suppliers are in the market and how are they positioning themselves? All of these very basic questions had very different answers when I looked at them this year when we produced the report in 2019. Back in 2019, the main question was scale. So all of the focus was around these data annotation workforces producing data that was used to develop specialized task-specific models.

5:12Anna Wyndham:So we're all familiar with machine translation models, speech-to-text, and there were other areas that were generating a lot of demand, robotics, image recognition, and so on. So going into this research, I kind of expected to find an evolved version of that market. and because of the, I guess, the LLM, I don't know what we want to call it, the LLM explosion or the LLM moment in 2022, I was also expecting to see that there would be a whole lot of new demand for the unstructured large-scale data that we know is used to underpin large language models. But instead, what I found is that the conversation has moved much further downstream, so much further into the question of practical AI implementation into the question of, yes, we have these large language models, they have great broad general purpose capabilities, but how do we actually make them reliable, safe, controllable, usable in real world environments?

6:14Anna Wyndham:And that question changes almost everything about what is needed in terms of data. It changes the type of data being purchased. So we need data that will align, adapt and evaluate models. It changes who's doing the work. So instead of relying on those large pools of generalist annotators, there's now a new and increasing need for domain experts who can provide judgment that will shape model behavior. And we see, you mentioned Appen, we see leaders in this space repositioning around this need as well as new companies emerging that focus primarily on this. It's also changed who's buying. So before the demand was largely organized around applied AI, so machine translation, computer vision, defense applications.

7:04Anna Wyndham:Now, of course, we have demand coming from the frontier labs, so open AI, anthropic, coherent, so on. We also have demand coming from all of the enterprises across every industry who are deploying AI and aiming to extract value from AI. and as well we have demand coming from all the companies that are building their products and trying to make their products use case specific and trying to make their products differentiated. One implication of all of those changes is that the economics have changed significantly so in the past this kind of data work had a very low value per data point but today some of the most valuable data comes from experts whose knowledge has been built up over decades.

7:51Anna Wyndham:And that means that in those cases, the value per data point is dramatically higher. So in short, old world, we had a market focused on building model capability. New world is a market focused on deploying AI.

8:06Florian:It used to be like somebody annotating like bird, car, bicycle-ish, right? And now it's like, hey, does this physics equation look okay?

8:16Anna Wyndham:Exactly.

8:16Florian:So are there still like misconceptions out there about the data for AI market generally, maybe among some of the companies that have heard about it, you know, on our podcast or in other areas and, you know, kind of covered it superficially or there's some kind of broader misconceptions around there?

8:35Anna Wyndham:One of the big misconceptions or misperceptions is that the market is still just that kind of data labeling that you just mentioned, looking at an image, labeling it cat or looking at a social media post and labeling it as having a particular sentiment or so on. another aspect that gets a lot of attention in the media is that is the the need for this large-scale data to pre-train large language models and so there's also a perception that this industry simply provides that kind of unstructured web scraped data and to be clear both of those types of data demands still exist and are still important parts of this market but what's interesting and what's new is this newer layer of demand that doesn't get as much attention but is really the most strategically important type of data now which is the data that makes it possible to release and deploy models so as you mentioned an example would be say a medical physician looking at output to assess whether the response is not just medically accurate, but an appropriate response to a particular prompt.

9:52Florian:Let's talk about the market size. Obviously, people are very interested in that. We size it at 9.3 billion, which given the numbers that we're hearing now with like Anthropic and OpenAI, it seems kind of cute. You know, you have like these like trillion dollar numbers being thrown around for, you know, valuations when these companies go IPO. So what does the 9.3 billion include? What does it exclude? What do we capture with that?

10:16Anna Wyndham:We're capturing the external commercial spending on data for AI. So the Frontier Labs actually have a large component of internal data AI work that's being generated internally and managed internally without the need for external providers. So this is not included in the market estimate. what we're counting here is external spending from frontier labs, enterprises, AI product builders on data sets. So all of the spectrum of data used to build, to adapt AI models, to align AI models with specific behaviors and policies, and all of the expert work that is required to test the models to see where they fail and to evaluate the models to see if they actually do what they're meant to be doing.

11:02Anna Wyndham:So this ranges from, say, off-the-shelf data sets, customer-created data sets, managed data services and the provision of those experts, the AI data platform, so the technology layer. And it also includes licensed data sets. So there's an increasing need for data sets coming from, say, publishers or media archives that can be used either to pre-train large language models or to adapt them to specific domains.

11:37Florian:I guess one there would come to mind is we had Casper Grothwell on the podcast from Oxford. so yeah they would they will probably be in the licensing part of the market yeah and perfect

11:51Anna Wyndham:example of a provider that's really ridden this trajectory of change very fast so they they were previously providing those off-the-shelf data sets and now they're involved in discussions around do we want to license this material for repurpose this material as ai training data so So as well as not including what's being done internally at the Frontier Labs, there's obviously a lot of discussion around how enterprises use proprietary data to make their models deployable and useful and relevant. So this is a very important type of data as well, but it's internal. It's not a case of external spend, so it's not included in our estimate.

12:32Anna Wyndham:And even though that figure may look maybe dwarfed by some of the figures being thrown around by the Frontier Labs, it's much bigger than many traditional annotation market estimates. And that's because we're looking at something much broader than just traditional annotation.

12:48Florian:So take us into the production here of something like this 160-page piece of research. Like, how do you go about something like this? You know, who do you talk to? What do you analyze?

12:58Anna Wyndham:Well, in terms of the scope, the goal was to map the full market to understand the data categories, the demand associated with each data category, all of the different types of suppliers, buyers, and then how value is being created. So like the economics associated with different data types. So we interviewed stakeholders across the system. We spoke to Cohere. We spoke to many of the major data for AI providers. So scale, Appen, LXT, Telus. And we spoke to a lot of the language solutions integrators that have built AI data businesses. So RWS, TransPerfect and so on. So these conversations were so, so interesting.

13:39Anna Wyndham:I can't tell you how many times I've gone back to those conversations. So thank you to everyone who spoke to us. But it's really just such a dynamic and interesting space at the moment. And the other thing that I looked at in detail were the system cards and safety reports and model cards from the Frontier Labs, really super rich source of information. And they really validated what we were hearing from the providers, which is we need these human experts to take part in these alignment adversarial testing and evaluation processes. And that this is a really core need to building models and making sure that they're actually safe, they're actually useful.

14:23So, yeah.

14:24Florian:So one of the biggest themes in the report was the rise of something that we call deployment data. So what is deployment data and why does it matter?

14:32Anna Wyndham:We're using deployment data as an umbrella term to refer to everything that adapts a model after pre-training. So for pre-training a large language model, you need the massive unstructured data. And for under-deployment data, we're including the data used to adapt models to specific domains. So, for example, to make them specific to legal or medical or financial contexts, we're including the data that's needed to align their behavior. So this is data that gives models examples of how to give a good response or what does an enterprise actually need when someone asks, can you write me a marketing plan or analyze this Excel data?

15:25Anna Wyndham:and it also includes the adversarial testing. So this is where experts will go in and try to expose the failure modes and the weaknesses in the model and try and get it to go against policy or provide dangerous information or to provide inaccurate information. And finally, deployment data includes all of the data, so the gold data sets needed to evaluate the model. So in order to understand if the model's working, first you need to define what good looks like. To define what good looks like, you need experts. You need journalists. You need medical experts, financial experts to say this is what good looks like.

16:08Anna Wyndham:And then you need those experts again to evaluate the output of the model and say, yes, in this case, it aligns with what good looks like or it doesn't. And an interesting feature of all of these types of data or these processes is that they are very iterative. So the more you test and evaluate the model, the more you expose or reveal or unearth the ways you need to change it. And in order to change it and adapt it further, you need more data. So it's very much a cyclical process.

16:40Florian:Let's talk about, you mentioned that red teaming and also the alignment. So what does that work actually look like? What are people actually doing there?

16:48Anna Wyndham:Yeah, so I'll start with alignment. So this is how data is used to teach models how to behave. So it shapes things like healthiness, safety, the tone that is used, policy compliance, and when a model should answer versus when it should refuse. So alignment teaches a model how to act. So when you're interacting with a model and it refuses to answer a question, this is as a result of this alignment process. So this covers a wide range of tasks. You might have experts who write ideal responses. They might be ranking outputs to say this output is better than this output. the experts might be looking at whether or not outputs align with policy or providing examples of how an output should align with policy.

17:41Anna Wyndham:So it's about teaching models to follow specific instructions or formats. So imagine a prompt who asks, how do I treat chest pain? And you get two possible answers. One suggests a diagnosis. One is much more nuanced. explaining possible causes, highlighting uncertainty, recommending the user seek medical advice. So the human contributor working on this data task would be tasked with selecting the safer and most appropriate response. And the contributor's decision there becomes a training signal for the model. So this is through various training techniques used to update the model. So you can see why expertise is important here, because a generalist simply wouldn't be able to make a good judgment in that case.

18:32Anna Wyndham:The other one you mentioned is adversarial data or red teaming.

18:36Florian:Red teaming, yeah. Did you know the origin of that term, red teaming? Is that like a military term or something?

18:42Anna Wyndham:Yeah, I think it is a military term. So it refers to when, I guess, forces pretend to act like the enemy and try to find weaknesses or try to infiltrate. I think that this is where it comes from. But it's a really interesting term and it's all the way through the model cards and safety cards because this is how you find the weaknesses and failure modes. So red teamers are designing prompts. They're trying to bypass the safeguards. They're trying to exploit flaws in reasoning. They're trying to trigger policy violations. and they're trying to expose inconsistent behaviour across languages, domains and use cases.

19:26Anna Wyndham:And you can see again why it's so important to be a domain expert in these cases because as a generalist you may have some ideas about how you might be able to get around the model's safeguards but you really need to understand a domain to understand the safeguards firstly and then to detect when they've been bypassed. So just as an example of the breadth of human expertise involved with red teaming, OpenAI reported that the team for GPT-4, the red team included 100 external red teamers. They spanned domains from cybersecurity, healthcare, law, finance, psychology, misinformation, child safety, linguistics, across 45 languages across 29 countries.