On the Protect Europe Podcast, I opened up the thinking behind Fuzu Atlas to Colin Brown from Sparkmind Capital— and why the next generation of AI will need better human intelligence, stronger provenance and a more governed data supply chain.
At one point in my conversation with Colin, he reached for an analogy that stuck with me: Patagonia versus Shein.
It is an uncomfortable way to think about the AI data supply chain, and that is exactly why it works. We are becoming increasingly demanding about the technology behind AI — the models, infrastructure, security architecture and risks — but often know much less about the human intelligence that trains, evaluates and improves those systems.
Where did the data come from? Who produced it, under what conditions, and who decided whether an answer was correct, harmful or culturally appropriate? If something goes wrong later, can those decisions be traced back?
These questions go straight to why we built Fuzu Atlas: not as another annotation marketplace or anonymous crowd, but as a governed human intelligence layer for AI.
The model is only part of the system
Every advanced AI system still runs into questions that ultimately require human judgment. Is this response genuinely better? Is the answer correct in a specialist domain? Does the humour work in Finnish, Hausa or Arabic? Would a pedestrian interpret a situation the same way an autonomous vehicle does?
These are not simply labelling questions. They are judgment questions.
As more basic data work becomes automated, the quality of that judgment becomes more important. Atlas is built around making it structured, validated and repeatable.
1. Provenance: knowing where intelligence comes from
AI has an increasingly complex supply chain. Training and evaluation data may pass through platforms, subcontractors, annotators, reviewers and automated tooling before it influences a model.
That makes provenance a product capability, not just a compliance exercise. Who performed the work? How were they selected? What instructions did they follow? What happened when reviewers disagreed? Who approved the final result?
The same principle applies whether the work is RLHF, safety evaluation, video annotation, language data or expert review: good AI data should have a history.
More on Trust & Compliance →
2. Geography should be a capability, not a constraint
Global AI data operations are often organised around where a supplier happens to have its workforce. We think the right geography should instead depend on the workload.
A European organisation may require EU-resident teams. A foundation model lab may need native speakers across multiple markets. An autonomous vehicle company may need people who genuinely understand road environments outside California or Northern Europe.
The goal is not to force every workload into one location, but to determine where work should happen based on language, expertise, jurisdiction, economics and data requirements — while keeping one governance model around the operation.
European sovereignty is one important configuration. The broader capability is controlled placement of human intelligence.
3. Depth matters more than a giant anonymous crowd
Scale matters, but headcount becomes a poor measure once the work gets specific.
You may need a native speaker of a low-resource language, a medical professional able to spot a clinically plausible but incorrect answer, a programmer capable of evaluating generated code, or an experienced reviewer who can adjudicate disagreements.
This is where depth matters.
Fuzu Atlas grew from years of building talent infrastructure across multiple markets. We think in skills, languages, validation, calibration, quality tiers and continuous performance — not simply profiles available for a task.
For sophisticated AI systems, sometimes the only person capable of identifying an error is someone who genuinely understands the subject.
More on Talent Ecosystem →
The English problem is bigger than translation
AI has made extraordinary progress in English, but that can create the illusion that multilingual AI is mainly a translation problem.
It is not. Language carries culture, humour, social norms, indirectness and context. A model can be grammatically impressive and still misunderstand the person using it.
Europe alone makes the issue obvious: we operate across dozens of languages, while much model development and evaluation still begins with English as the default representation of human behaviour.
If a model is expected to work in Swahili, Arabic, Portuguese or Finnish, people who genuinely understand those languages and contexts need to be part of the evaluation loop. This is one of the areas where the breadth of the Atlas ecosystem becomes particularly valuable.
More on Multilingual Data Operations →
Benchmarks are a moment in time
Another point from the podcast deserves attention: benchmarks are a moment in time.
They are useful, but they are not the model. Performance can change after another training cycle, fine-tuning exercise or product update, while public benchmarks themselves gradually become part of the environment developers optimise against.
The better question is therefore not only, “Did our model pass the benchmark?” but, “How does it perform against the behaviours, risks, languages and edge cases that matter to us right now?”
That points toward private benchmarks, human-constructed evaluations, regression testing and continuous adversarial testing. The evaluation layer has to evolve alongside the model.
More on Model Benchmarking →
And then AI enters the physical world
The same questions become even more tangible in robotics, autonomous vehicles, drones and industrial systems.
Here, the long tail matters: unusual pedestrian behaviour, unfamiliar road layouts, strange objects, sensor occlusion and situations nobody included in the original taxonomy.
Commodity datasets are valuable for getting started. But as autonomous systems move closer to the real world, specific, traceable and purpose-built data becomes increasingly important.
Human intelligence remains part of that process — identifying edge cases, interpreting environments, creating new data and deciding what the machine has failed to understand.
More on Robotics & Autonomous Vehicles and Data Collection Programs →
So what is Fuzu Atlas?
The podcast gave us an opportunity to articulate something we have been building towards for some time.
Fuzu Atlas is the governed human intelligence layer for AI.
We bring together people, data operations, quality assurance and governance to support organisations training, evaluating and deploying increasingly capable AI systems.
Sometimes that means multilingual evaluators or expert reviewers. Sometimes RLHF, preference data or red-teaming. Sometimes thousands of pieces of image or video data from a specific real-world environment. And sometimes the defining question is where the work is allowed to happen and how its provenance can be demonstrated afterwards.
The common denominator is human intelligence made operational: structured, validated, traceable and capable of scaling without becoming anonymous.
The AI industry is moving incredibly fast. The systems around the models now need to keep up — with better evaluation, better representation, better data and stronger governance.
That is a much bigger challenge than annotation.
It is infrastructure.
Watch the full conversation with Colin Brown on the Protect Europe Podcast to hear us explore multilingual AI, changing benchmarks, data provenance, physical AI and why the human layer may become more important as the models themselves become more capable.
Link to the original YouTube video.
Or go deeper into the Fuzu Atlas Talent Ecosystem, Trust & Compliance architecture, Multilingual Data Operations and Model Benchmarking.
Full transcript
00:00:00 - 00:00:55
The first one there is we need to have a broad pool of native speakers in in in in you know, eventually in in thousands of languages, but but at the moment we're still looking at at a relatively sort of limited pool of languages. In our case, we we we are focused on 40 40 plus languages. We can expand that to 50 60 70 and basically we're walking this this journey with our clients. So whenever they tell us that hey, now they need something a bit more special, we can add those languages when
00:00:29 - 00:01:36
>> [music] >> when needed. >> Welcome to PEP, the Protect Europe Podcast, [music] where we interview founders and investors building the technology >> [music] >> that protects Europe's values, way of life, and infrastructure. So uh UC, thank you so much for coming on the podcast. This is really interesting. I love this particular space. As you know, we've talked about it many times before, but let's dive straight into what is Fuzu Atlas, right? And why should partners, particularly
00:01:01 - 00:01:57
BA, be paying attention to this space? >> Super. Thanks, Colin. Uh it's good to be good to be here and and really excited to share a a bit more about what we do at at Fuzu Atlas. As you know, Fuzu Atlas is is a next step in the evolution and the journey of what Fuzu has has been about for more than 10 years already. Something that we have been now building for the past past two years and and essentially what we're doing here, we're building on the legacy of of the Fuzu platform and the talent pool that we've
00:01:29 - 00:02:20
been building now for the past 10 years with millions of millions of qualified professionals and and already hundreds of thousands of people connected to labor. But what we're now adding on top of it is is a unique layer that is is specifically designed for supporting the growth of the AI industry going forward. And so that's what we're going to be covering now. And and you can see my deck there already, so let's let's dive in. >> Yep. >> And so essentially, you know,
00:01:54 - 00:02:46
what I'm going to be uh covering now is is taking a quick quick look at the market and what the market demand at the moment is. And then to look at uh what are the core solutions that form of Fuso Atlas, and what are then the use cases that uh that our clients are asking for from us. And then how does the back end itself work from the viewpoint of team activation and building up trust with the clients. And then what are then the kind of the engaging models and and you know, if you want to work with us, you
00:02:20 - 00:03:16
know, what should you do do next. But before we go anywhere, let's uh still sort of take a look at what this whole thing is about. And so when we think about where the world is heading at the moment, it's very clear that AI models are developing super rapidly and and they are adding incredible feature sets and and functionalities to them uh almost every single week. But still when we think about it now, we are getting to a moment where human judgment is more important than ever. And and that's
00:02:49 - 00:03:36
that's something that is needs to be kind of the foundational thing that ties these models to our reality, the way that we engage with each other, the way that cultures operate and so forth. So essentially the model safety is going to be a massively important topic for anyone who wishes to build any any any new innovations that are reaching uh significant scale. And and secondly then, what is important is that we're not just talking now about the English-speaking world. We're talking
00:03:12 - 00:04:05
about the entire planet. So we're talking about the smaller language groups. We're talking about the cultural nuance in local context and so forth. So So that's something that we we need to take seriously. We can't be just building uh you know, the the uni uni culture with with just English English English models. And then thirdly of course, we see that the regulators are getting more interested in this space and and they are waking up a little bit late in much of the world. But if we do
00:03:38 - 00:04:35
not build the the uh the safeguards, if we do not build the the processes behind these models for ensuring safety, we're not going to be ending up in a good place. >> Look, I think there's a really interesting point here that I've already seen insurance companies basically cancel a generative AI content from their policies. It's not insured, right? So, because effectively they're like, "What is that? This is a black box system where it produces a different response each time. Yes, it's trained on
00:04:07 - 00:05:07
your policies, but you don't know how it's done. How can I insure your business for the outcome of that product?" So, I've already seen insurance companies take very quick action in their policies, particularly in the US where people sue, and basically redlining generative AI out of policies. And in the same way, we're already seeing large enterprises struggle to adopt AI because effectively, not just cuz the insurance layer, but because that that trust layer, like, are they making a
00:04:36 - 00:05:28
commitment on behalf of the company, and is the company prepared to back up that commitment? >> Yeah, absolutely absolutely. The liabilities are of course massive here, and and and the funny thing here is that in most cases, the overall experience is is is still going to be better. But because you have those edge cases that may be super expensive or companies, they are they are more and more concerned about the repercussions of of the models, you know, getting out of hand or not doing what they're supposed
00:05:02 - 00:05:53
to be doing. So, so it's it's a big topic at the moment. But it's something that most companies are only thinking about after they are they are pushing something something already to the production readiness. so it's something what they should be building up front and embedding into the process from the start. >> Exactly. But but if we then >> sort of try to summarize, so what are we what are we then going what are we doing with Fuzzbuzz Atlas? We're trying to build a governance-first
00:05:28 - 00:06:21
workforce for artificial intelligence. And so, what we're delivering there is native language annotations, expert reviews, so basically people who understand the context or the domains that they are then reviewing, and and safety evaluation frameworks, which are all sort of aimed at building then those safe and models for for the human kind. And and and so, so when you think about it, what are we building this on? We're building it now on on a very deep a global talent ecosystem that we have
00:05:54 - 00:06:54
been building and are expanding as we speak to new regions and and to new competence areas. And secondly, we're building it on on a highly sort of structured activation process. How are we defining the scope of the proof of concept? What is then happening next as we are then pulling the the teams together and and rolling them out then to work with a client project somewhere. And then a layer that is foundationally important here is the QA layer. How do we ensure the quality? And then then how do we
00:06:24 - 00:07:25
embed the compliance into this process? So those are kind of the four building blocks on top of which Fuzu Atlas is being built at the moment. I already mentioned that we've been building Fuzu for several years. And and and as I mentioned already already before when we were chatting as an intro to this this discussion, the big differentiator between us and many of the other actors is that we are a technology company first. We started by building an AI-powered highly mature talent marketplace that has now been attracting
00:06:54 - 00:07:51
millions of people to the talent pool. Uh which means that that we are not a traditional AI consultancy or staffing company that then just ventured into building a platform. But we are first and foremost, we are technology company that is operating a large-scale talent marketplace. So therefore, we have millions of people in the talent pool. We have already supported thousands of companies over the past past years. And and and we have kind of the plurality of the of the competences out there. And
00:07:23 - 00:08:10
they're growing constantly primarily through organic outreach. So we are spending very limited amount of capital at the moment for talent attraction, but are still uh able to grow the talent pool by tens of thousands of people every single month. And so those those are some of the crucial assets what we are what we are leveraging at. >> Yeah, and and again, you see it was quite small on the slide, so just kind of just highlighting it for those who may not have been able to see the screen
00:07:47 - 00:08:35
or just listening to this podcast in audio version. You've got some big clients who've already used your platform so far. Do you want to just spell out some of the ones that were on the previous slide? >> Yeah, so so the talent pool side is is at the moment about 3 million. The number of companies we've served them with is about 2,000, actually closer to even even 2 and 1/2 thousand at the moment. And the languages we are supporting is more than 40. And and the amount of people that we are adding to
00:08:11 - 00:09:09
the talent pool is somewhere in the range of 20 to 30,000 people organically every single month. When it comes to clients, I can't unfortunately disclose who the clients themselves are, but what we have both local and global large-scale businesses all the way down to the foundation labs that we are working with. >> Okay, what about the capabilities that then you're looking to kind of really deliver deliver on this space? >> Exactly. So so there's basically five things or five kind of sub products what
00:08:40 - 00:09:39
we're building with Fuzu Atlas. And the first one there is I'm going to go in a bit more detail into this, but multi-multilingual sort of data operations. The second one is then what we call LLM evaluation and safety, really making sure that the large language models are operating optimally and safely. Multimodal annotation which then adds other forms of content, video and audio into the mix. And then a kind of the the governance layer of the of the quality assurance and then making sure that everything works as it should.
00:09:10 - 00:10:08
And then one thing that is quite important at the moment, data collection. So we are running out of data in some ways. And so so that's one of the one of the capabilities we are adding there. So going a little bit deeper into into each of these things. And so so the first one there is we need to have a broad pool of native speakers in in in in in our eventually in in thousands of languages, but but at the moment we're still looking at at a relatively to of limited pool of languages. In our case, we we we are
00:09:39 - 00:10:36
focused on 40 40 plus languages. We can expand that to 50, 60, 70. And basically we're walking this this journey with our clients. So, whenever they tell us that hey, now they need something a bit more special, we can add those languages when when needed. But essentially here, the whole point is that we are talking about non-English resources, making those models to function and operate in local languages so that we can start building innovative solutions that and whether they may talk about East African
00:10:07 - 00:11:09
language like Swahili or or we talk about Indian languages, we talk about Middle Eastern languages, or we talk about then then South Southeast Asian languages. So, so this is one important dimension and capability we're building is is the sort of multilingual assets for building then those those uniquely localized and culturally culturally sort of connected models. The second thing then is is really sort of adding this second capability capability is then this this LLM human evaluation, which
00:10:38 - 00:11:29
basically means then that how do we make sure that the LLMs are not doing something stupid? How do we sort of test them? You mentioned yourself the red teaming there as an example. So, how do we sort of prompt these models and see what the outcomes are? How do we make sure that you're not able to start programming your next next next venture by using a chatbot of a hamburger joint? And so, so there's different things here that that that we can be doing for making sure that the LLMs are doing what they're
00:11:04 - 00:12:04
supposed to be doing the way they are integrated then to other applications, they are behaving the way that they should, and adding them into this this this sort of supporting also then the fine-tuning of these models through reinforced learning through human feedback, and really sort of making sure that these models are aligned with how people behave, how people think, how people use language as well. And so, so the LLM, of course, this is a hot hot topic at at moment, and and and it's it's interesting because we see this
00:11:34 - 00:12:32
this world changing constant and so so so you what was needed three months ago is maybe is is is not what is needed at the moment and so therefore also then the talent pools that are involved in this type of work are changing as well. The third one then what we have here is is then the multimodal work world as we as as as we may discussed. And so basically this means then that it's not just about text, but it's then about image, it's about video, it's about other type of of content that these
00:12:02 - 00:13:04
models then then then then crave for. And annotating and and really making that content uh structured or labeled annotated in a way that then these models can be learning from them. And so this is also something that is of course massively important especially now that we are then venturing into the physical AI world where we are then trying to understand, you know, how does the the world around us operate and how do we then build models and also robots that are then that understand then their surrounding perfectly. Number four is
00:12:33 - 00:13:32
then adding this this domain expertise. And so so one thing that is also interesting here is that it's not just about sort of basic level annotation or basic level governance or red teaming and checking how these models operate, but it's also then about going deeper into thematic areas. So so so it might be about STEM, it might be about finance, it might be about legal, it might be about health care. And in this case you then need experts who are then specialized in those specific domains.
00:13:03 - 00:13:59
And and and here we see more and more specialized purpose-built models and solutions entering the market and so the need for this type of expertise is is growing. And then the last one what I mentioned was then the data. So now that we have all those other layers of model development in place, we then realize we still need more data. We need constant flow of fresh data so that everyone is is training their models based on same data sets and so so then creating fresh data, whether that's images, video, or
00:13:32 - 00:14:35
text, or audio, is something that is is is very important going forward. So, I think those those are kind of the four primary capabilities what we are building at the moment. >> Look, and I think it makes perfect sense why these are needed. Like, I think we already spelled out the accuracy level that you would get if you were basically unless you literally your entire customer base is US English-speaking people who never, you know, don't have a passport so they never leave the country, right? Like, unless that's your
00:14:03 - 00:15:01
demographic and you start touching anybody else in the world, the data sets just aren't there. Like, the accuracy levels with the models aren't there. The real challenge is it looks good on us good looks good in your implementation plan that you're going to, you know, service customer service via these kind of systems. But when the rubber hits the road, a lot of this will just fall down because the data is not there, the context is not there, data is not trained in the right way, and the models
00:14:32 - 00:15:30
cannot basically differentiate outside English white-speaking language set. >> Spot on. Spot on. And and even if you just look at Europe, we are in Europe, we're speaking dozens of languages. And even though, you know, people are highly educated, but still English is not a dominant language in in in almost any of these countries. And same, of course, applies in Southeast Asia. And so so this sort of singular world where we think that everything runs around it runs on English is is not the reality.
00:15:01 - 00:15:59
And the deeper we go into the economies, the more sort of also the cultural nuance is is required. And because not everyone is watching the same movies, not everyone's language is evolving, you know, based on the same experience around us. So so so there's much more nuance as we go forward. >> Okay. This makes sense. >> Super. But then let's move then to kind of converting those capabilities now then into actual problems that then the clients may be facing or what they need
00:15:30 - 00:16:24
to be then then then then then solving. And so, essentially there's sort of six problems that we are looking at the moment. So, first one is is then just centered around kind of the LLM evaluations and regtech in it. So, really supporting the companies in in figuring out how do we prevent these models from doing something stupid. Now, the second one then links back then to the capabilities. But is there's is there making those how to make those models then multilingual? Just as we discussed, cultural nuance and and
00:15:57 - 00:16:56
linguistic nuances massively important. Multimodality. Then again, how do we then add those those additional layers into this rich world of ours, which is not just text, but it's also understanding the physical physical world around us. And then, companies who need that that quality assurance layer. So, they may be able then to do some of the foundational work. And and as we know, many of the AI houses are just focused on the core AI innovation, but they do not have the capability then for ensuring quality or
00:16:26 - 00:17:24
then then fine-tune and and and improve the performance of those models. And the model benchmarking is also something that is is emerging more and more as as the fifth fifth need that we see emerging. And then, lastly, but not least, of course, then the data collection problems. And so, so as you can see that capabilities we're building are very much aligned with then then the problems that our clients are then then seeing out there. I'm going to be running these through very quickly because we already covered them in quite
00:16:55 - 00:17:49
a bit of detail. And so, obviously, when when the client needs evidence, it's not about volume. It's really about kind of the trustworthiness of those models. And and when we when we then think about kind of the adaptability of those models, that language the linguistic nuance and and polarity is is is of course massively important. And so, so that's that's that's clear. Then, the multimodality, whether then we talk about kind of the physical AI and robots that are trying to perceive what is
00:17:22 - 00:18:21
happening in front of me. Maybe it's a sewing machine. Maybe then it's it's a welder. Whatever. Or then it's it's really sort of adding different sort of layers on top of the image data what we are doing. Lidar, radar images, 3D images, annotations, and so forth. So, that's how machines of course learn about the physical reality around us. And then when we think about the expert in the loop layer, obviously, it is really about adding temporary teams because you don't want to necessarily
00:17:51 - 00:18:51
have an army of 50 people sitting on your on your payroll all the time, but rather than you would have experts that are abroad on board whenever you have the next release coming. You can you can increase the size of the team flexibility and then bring it down when when you then have done done with it. And so so things like that. >> Yeah, and just on this one, I've got a bunch of friends who have either running companies that basically do this AI kind of onboarding for large-scale enterprises. And every single I'm not
00:18:21 - 00:19:18
going to name the companies, every single one of them will talk about how not data ready any of the enterprises are, right? So, the the enterprise data set is siloed is a sweeping generalization, but it's sweeping generalization cuz it's completely true. The data is it just in thousands of different places. >> Absolutely. >> None of it can be gathered. None of it has any degree of labeling. So, it's kind of garbage in, garbage out. And it's Sorry, it's even worse. It's
00:18:49 - 00:19:46
incomplete, incoherent garbage in, and whatever, you know, wording you want to use on the kind of result coming out. >> Absolutely. And you know, of course I started my career with with product data management, knowledge management in the late '90s sort of late 1990s. And we were very sort of cutting edge in Finland in that domain because that was the only way for how we we able to you know, build innovations that that where we are able then to serve massive amount of different client needs with with
00:19:18 - 00:20:09
highly sort of modular product back end. But but one thing that the companies were always feeling that why would I need to be cleaning this data? Why would I need to be doing investing, you know, millions of dollars in in cleaning those data sets and building those metadata models and and and labels. But then whoever did that now is sitting on a treasure trove of of data that then helps them to become much more agile. >> But of course, it's becoming a little bit easier with all the annotations and
00:19:43 - 00:20:43
and and and whatnot. But but still, if you don't have your your kind of core data set solid, it's going to be super difficult for you to go forward. And so so absolutely. And then, you know, very quickly sort of covering these remaining two things here. So so kind of the model benchmarking, of course, is also increasingly important as we're looking at especially in this sort of multi multi-modal and multilingual world where we're trying to understand kind of the depth and the expertise of these models
00:20:13 - 00:21:17
and in different type of contexts. And so this benchmarking exercise is is super important especially as the kind of the the demand for more complex solutions is is constantly constantly growing. And then >> Yeah. I think we've beaten that Just on the Just on the benchmarking, like it just annoys most people. The benchmarking just annoys most people because fundamentally, it's like humanity's last exam or these bunch of maths problems or this e-value that I But it's not related to you and it's not
00:20:45 - 00:21:48
related to your customers and not related to your use case. So I wouldn't say people are prepping for the test, but they clearly are, right? It's just I want to make sure that my chat GPT, my Gemini, my z.ai is basically, you know, further to the left than the other guys, right? And and the benchmarks kind of don't mean anything anymore. It really needs to be your benchmark to your system and your customers and how close to level of accuracy that you're currently doing and can we exceed it?
00:21:17 - 00:22:15
That's clearly where we should be going for this. >> Yeah, absolutely. And it's to be honest, it's it's a difficult world. I'll give an example. We were just on our our side as we're building our own tech. And we were just now sort of switching some of the embedding models that help us then to build these, you know, massive vectors that we then use for and for doing rapid database driven calculations. And and we had to shift from one of the models then to another because we were not able to ensure GDPR
00:21:46 - 00:22:35
compliance of the of the first model. But then you know, there comes the problem of how do we compare the performance of these two models now? And how do we sort of really see the real-life difference? And and it took quite a bit of time uh to figure out that what actually is that is the right right solution for us. And and we still, you know, we're just now pushing it now into production. So, finally when the rubber hits the road, we'll see that how well it actually performs. But but yeah,
00:22:10 - 00:23:12
this benchmarking and the testing and and evolution is tricky. >> And also just benchmarking is just a moment in time. One of my friends did a partnership with one of the frontier labs. They rolled out a very highly research-based piece of benchmarking analysis. Two months later, the model had changed and the analysis was actually completely different. Right? So, it wasn't like it went up and to the right. In fact, the model performance on a bunch of tests had got worse. Right? And on certain tests it got
00:22:41 - 00:23:37
better. It was just it was basically like, you know, meet somebody on the street, assess their performance, then randomly two months later meet their friend on the street and assess their performance. It was totally different totally different experience. >> And you think about kind of the past past world where the product life cycles were measured in years and now we are measuring it in in weeks. And so so yeah, it's a constantly moving target that we are we are playing with here. I'm not going to go deeper now into this
00:23:09 - 00:24:10
data side of things, but but of course we we already discussed it quite a bit. But but what we see now is more and more demand for different type of of data, whether then it's image data, video data, conversational data, documentation about specific things, which then sort of somehow represents the real world. Not an easy topic, because the demands are always highly bespoke and and specialized. And so sometimes getting that data is is is difficult. But but this is something that I'm sure that there's a lot of
00:23:40 - 00:24:34
people around the world who are playing with this as well, then figuring out what are the more scalable ways of producing that data. But still raw real data is is needed going forward as well. And so if we sort of bring that back those capabilities and then those needs, and then we think about who who are then those actors behind those needs, I think that it's it's the usual suspect. So so you have the foundational model labs, you have the robotics companies, you have the enterprise AI businesses who
00:24:07 - 00:24:58
are who are then building a bespoke, let's say, business solution or legal tech, whatever that it is. And your trust and safety teams, which could be anywhere. It could be your your localization and language teams that are thinking about how do I make this solution now work in a new market that we are just now entering. And then of course then the regulated industries. It might be your defense, it might be your health care actors and so forth. So those are kind of the six different buckets we'll be looking at
00:24:32 - 00:25:29
when we are sort of reaching out then to the to the clients. If we then take a step back and we go back to those assets we've been building, the millions of people in the database and the thousands of companies that we have served, then when we think about the footprint, as you know, Fuzu originated from Finland. So we are Nordic EU-based company, but big part of our initial journey was around Africa. I've been myself working in the Africa and East the region for the past now 20 years already, and know the region
00:25:00 - 00:25:53
inside out. And and I wanted to come up with a solution that solves the job market challenges of that region. But now that we have proven that the asset that we building, the platform itself, uh the marketplace, is something that we can use more broadly. We have now started that building those linguistic assets and the competence assets in other parts of the world as well, and expanded then to LATAM, expanded to Europe, as well as then to to to some parts of Southeast Southeast Asia and South Asia. And so so the footprint is
00:25:27 - 00:26:17
growing, and as I mentioned, we are growing the footprint together with our clients. So so whenever there are needs, we then then explore new markets and then figure out how can we start building those interesting talent pools for our clients' needs. And so the footprint is is constantly evolving evolving there. A practical thing, of course, everyone is asking us always, "So how do you then get things going?" And so so so how can we start testing out and figuring out what we could be
00:25:52 - 00:26:41
doing together? The big insight here is it doesn't make sense to run into a massive project straight away. But it's really about understanding the context, understanding the data, understanding the needs of the client initially, which is kind of step number one. It doesn't take that much time, but but typically it's good to start from a ring-fenced problem. And then secondly, then it's about setting up then then a team. So we identify the the professionals that could be good in in in resolving this
00:26:17 - 00:27:08
particular particular challenge. And then we roll out then this this sort of government POC, where we try to have a relatively small team initially work working on the ring-fenced problem, and proving that we can actually deliver and we can solve this issue. If there are issues along the journey, then we iterate. We go back to the starting point, and we refine, or then we may even change the team members if that's something that's required. If there's some special competency that we didn't
00:26:43 - 00:27:42
actually see at the beginning that we need to be then adding then to the pool. That's the reason why we have those millions of guys in the in the database. And then before we then start scaling, it's about reviewing and and figuring out what did we learn from this exercise? What did we do well? And then, you know, move from there. And that's kind of the process that we like to follow, starting from ring-fenced POC, and then scaling from there. Super. And so, if we then just take a still still another dimension in in this
00:27:13 - 00:28:06
So, now that we looked at also the kind of the operational capabilities, what are we What are we doing? What are the steps to getting getting a project moving? The underlying element behind all of this is is still kind of the governance. And and so, so, it's not something that is that should be just bolt on, but it's something that needs to be a foundational part of what we do. And that is also where this this European heritage and the European and Nordic heritage also comes to play. And
00:27:39 - 00:28:38
being plugged into the European ecosystem, allowing us to be then that bridge also then between different parts of the world and and EU. And so, whether then you are a Chinese company that wants to expand to Europe, or then you're a European company that needs to have a super solid foundation and GDPR compliance here as as you are then expanding to new markets. Everything goes there. But but essentially, the the the four big things what we are looking at is is the quality assurance authority, which means that
00:28:08 - 00:29:08
how are we then setting up those QA processes and steps in in place. How are we then building the audit trail as we sort of understanding kind of what have we been doing? How are these changing? How the model itself is is is performing. Data handling, which also matters a lot sometimes, but you know, when there is no PII data, when the data itself is not sensitive that sensitive, then we can have a bit more relaxed ways of handling the data. But then, if it's super secure, super sensitive, then of course
00:28:38 - 00:29:37
the processes are look then quite different different. And then lastly, has to be based on of course ethical ethical labor. As you remember, where we started from is about providing people with opportunities to grow and and and earn a decent income. And so, it's important to us as well as a as an organization that we are providing people with meaningful job opportunities that uh that then then help them then to grow and and accelerate their their journey to the next opportunity. And and the idea as well is that we're providing
00:29:07 - 00:30:00
people with a trajectory forward. So, it's not just doing that entry-level task for the next 5 years, but it's about learning and acquiring new skills and and mindsets and then growing from that. But so, I think those are the kind of the the the the the key things that are under underlying underlining everything we'll be doing. So, yeah, uh very very actually excited about and proud about the work we've been doing now in putting the building blocks in place, really sort of articulating, you
00:29:33 - 00:30:30
know, what those specific clients client industry needs are in this in this in in this time. And then, of course, keeping this agile. And and whenever new needs and and new requirements are emerging, then of course, uh we need to be ready to then modify the setup as well. But yeah, that's in a in a nutshell what we are what we are what we think. Yeah. >> Yeah, and look, I love the fact that your defaults are set this way, right? You make it very easy and very clear that what you are and what you're not.
00:30:01 - 00:30:59
It's very much a European {slash} African approach initially. So, effectively, it's it's almost like what you know, this is Patagonia building a garment, right? So, the supply chain's going to look like this, the ethical standards are going to look like this, the end consumer like this. This is not Shein building a garment, right? So, it's very clear to an outside uh someone looking at Fuzzable Atlas is going, "Okay, so this is kind of on that scale." Right? This is what This is what it
00:30:31 - 00:31:32
should do. The model is this, the defaults are this, this is where it goes. And you're making very therefore a very clear counter position between certain companies and yourself. So, it's making it easy to differentiate and force that choice. >> Absolutely. I I I'd add still one more layer there. Because of the of the origin of Fuzu, where we started this whole thing, Fuzu has always been super lean. And so, the cost efficiency is massively important to us. And And cost efficiency doesn't mean that you are
00:31:01 - 00:32:02
unethical. It means that you're acquiring the talent where the talent is. And you're then plugging that talent, high-caliber talent, into your your ecosystem. So, so that's one thing that that we are super efficient at because we have been operating in this space for quite some time. And so, combining kind of the best of both both worlds. So, super. Any final questions? >> No, no, no, no. So, again, I think it's more a question of So, you've laid out the model, you've laid out how quick But
00:31:31 - 00:32:35
now we haven't given anybody some timescales, right? So, how long does it take to get a a new client up to speed? Right? So, it's that kind of like, you know, inbound lounge in your in your J bar, how long will it take to do X path? So, people have an understanding. Right? Now, I appreciate, you know, if they take seven 70 days studying the POC document, then it takes 70 days plus, right? But I mean, all things being equal, let's look at the timelines for onboarding. Let's look at again the kind
00:32:03 - 00:32:59
of the sort of client use cases. So, people are really clear, where is Fuzu Atlas playing and where would some other players like Macquarie and others be kind of in the value chain? So, first timelines and then kind of like, you know, ideal customer profile. >> Super. >> Yeah. Uh w- when you as I mentioned, the key thing for us is is to understand exactly kind of what would be that first initial ring-fenced problem that we want to be solving. And so, so that's something that typically ta- takes a few
00:32:31 - 00:33:27
few meetings and and getting to know know each other. You have to build a relationship. You have to have to build that trust between the between the actors. And so so typically the POC, the way we are framing it is that it should be should be a doable within 30 days from start to finish, okay? And and and so that's kind of the goal what we have. Of course, sometimes when you're talking to talking about larger organizations, you might then just to spend 6 months with the contractual framework, but
00:32:59 - 00:34:00
that's partly why we like the POC concept to start with because that that yes, you need to have of course all that data protection all that all that in place, but you can you can typically start with a lighter agreement framework and to do that sort of 30-day POC to prove your value without then, you know, having these massive MSAs and and so forth in place. And so that's that's what we what we aim for. 30-day POC and proving our value and and then moving from there then to a higher scale and longer-term longer-term
00:33:30 - 00:34:31
engagement. But to be clear, also then we have nothing against working with even sort of short-term engagements. And so so this is something that that we can do. You know, it can be a few week engagement as well, but but of course ultimately, we're building long-term relationships with our clients and then going from there. If we then maybe take one step back there, so so when we go then to that scaling phase, the scaling phase, the way that we are kind of dividing the competence in the world world in our own internal jargon
00:34:01 - 00:34:57
is that we're looking at talent pool that we can scale within a week. We are looking at talent pools that we can we can engage and roll out within two to three weeks. And then we have teams that we can then, you know, kick off in about a month's time. And so so what we are systematically doing, we're building those competencies and talent pools, dividing then those those competencies into these sort of three three sort of categories depending on the location, depending on the competence. And so so
00:34:29 - 00:35:24
typically when the clients come to us, they have a specific need in mind and then we go over it transparently with them. Hey, these guys we have available, you know, within let's say two, three days. But then these this this competence pool will take some time for us to build. Are you fine with us spending, let's say, the next 20 business days in in in finding, validating, testing, and and sort of setting up the contractual framework for for the engagement? And very often that's completely fine. Of course, there
00:34:57 - 00:35:55
are there cases when the client needs a team tomorrow. But but yeah, the the those are some of the some of the kind of the variations what we see. >> Okay. And then in terms of when should they be calling you see from Helsinki to work on Fuzu Atlas? When they should be calling them a core guys or Turing or Handshake or any of the other players in the market? I'm not I'm not asking you to kind of badmouth your competitors. I'm saying is different folks for different strokes or different horses
00:35:25 - 00:36:20
for different races. Everybody has their market position. So, where do [clears throat] we Where does Fuzu Atlas sit in comparison to the other guys? >> Yeah. I think the big thing with us is is the scalability and the agility. And so so being able then to enter flexibly new markets and to start building those talent pools is is a key thing. And as I mentioned, a big part of this is is the is the platform that we're building now for for several years. So, I'd say that that's that's one thing. The kind of
00:35:52 - 00:36:57
agility and ability to scale. The second thing then that that I would still point out is is because of that sort of frugality in our in our setup. And and being very lean and ethical at the same time, I still see that we are highly competitive when it comes also then to costs. Okay. So, so that's I'd say that those are the two things, scale and cost. >> Okay. Yeah, yeah. And look, ultimately there's scale, there's cost. But ultimately, if you are turning energy into intelligence, which is really what
00:36:25 - 00:37:16
we're doing with these models, right? You know, that's what you know, we put energy into a data center, we we get training, and we basically get intelligence out. A lot of it's got to come down to value, right? You know, it's not it's not it's not about how many tokens I can spend in my enterprise. We kind of that was just to get people using the stuff. >> Yeah, yeah. >> Now I've got to make sure that I actually get the value out of it. So, the question being is then, how does
00:36:51 - 00:37:48
then the food And this is less related to price or kind of cost. This is more related to the quality of what you deliver, and how close that not just matches what they have today, but actually exceeds what they have today. >> No, no. Absolutely. And I think that kind of the word what I left out from there is when we talk about scale, I'm talking about scalable high-quality talent pools. And and so so that so of course the quality of talent is is a key thing here. And so the combination of
00:37:19 - 00:38:15
that agility to scale, the quality of the workforce, and then your cost level, those are kind of foundational things. But of course then if you don't have those other pieces of the puzzle in place, which was related then to how do you manage this this project itself? How do you then then ensure the security of course of your client's data? All of those things are are kind of foundational things that have to be in place. And things that that are of course extremely important to us. And and and
00:37:48 - 00:38:49
and of course need to be extremely important to everyone in this industry. But but that's kind of what what we see is that when we think about what our clients tell us when they engage with us, I'd say that the biggest thing there has been that our teams have been very nice to work with. They like the return on investment. They like the quality of the of the of the output. And so so because in this world where you're very often dealing with relatively large, you know, large teams, which may go from few
00:38:18 - 00:39:14
people to thousands of people. And so then it's also very important that that entire relationship is frictionless. When you have a problem, the problem is solved and you're not just left waiting. And so, so that engagement itself is and how you manage it is is is massively important as well. >> Mhm. Yeah, no, that that makes more sense. In terms of it's it's kind of time to value. If we're talking about value, it's also the question around, well, that's great, but I've got my
00:38:46 - 00:39:52
board {slash} executive breathing down my neck. Actually, when can we get the value? And a lot of that is not just my ability to as an enterprise or or a customer to work with you, but it's then how quickly we can scale this thing up into something that actually can start looking at a quarterly scorecard. >> Yeah. >> Right? Cuz I think that that makes a difference, right? If I can start to show, okay, first of all, I was just on, do I have an AI plan, right? Have a plan. Then I'm judged on, okay, now when
00:39:19 - 00:40:15
is this plan going to be operationalized? Now I'm judged on, okay, and what's the performance against the plan? And then there's the what's the budget against the plan and and not just getting that return, but how quickly are we getting that return? And also, I think there's a degree of, you know, this isn't, you know, dogs are for life, not just for Christmas. You can't just do this thing once and presume that your data set covers absolutely everything and all customer queries for
00:39:47 - 00:40:44
between now and the dawn of time have been solved. We only had to do this once. That's not realistic, right? As you launch new products as a company, as you basically build new services, as you encounter new situations, this is an ongoing provision that's required. >> Yeah. No, that's spot on again. When I'm looking at the engagements we've had with clients, it is very very often that the clients do not initially exactly know what they do or what they need. And and so, it's it's it's also something
00:40:15 - 00:41:11
that that that we need to need to inbuild build into the process of of this sort of iterations and those rapid rapid sort of sprints where we are able then to provide feedback to each other and then to sort of course correct as quickly as as possible. Because the last thing what you want is that you have spent millions of dollars in building your models and then your model refinement or or or safeguarding whatever and then it's done poorly and so so you need to add need to ensure that that there's clear alignment there
00:40:43 - 00:41:40
and and of course to be fair as as we discussed already many of the many of the companies who are building building this solutions are hyper fixated on on the tech itself. Is it doing what what it's supposed to be doing? But then this sort of governance layer and and really making it sure that it is real it is you know behaving the way it's supposed to is then something or those edge cases then are typically you know little bit out of the focus. But but now that we are seeing this of course world changing
00:41:12 - 00:42:09
rapidly companies are burning their fingers and and we see more and more companies then maturing and realizing hey this is not just about compliance it's really about the core experience we're building and making sure that we actually delivering what we promised to our clients and so so this is a foundational component of that of that delivery process. >> Yeah and look again any frontier technology as it's being rolled out against the public there's going to be some rough edges. But again you couldn't
00:41:40 - 00:42:44
have a I've just rolled out a new drug to the pharmacy most of them don't work. Some of them some of them will kill you and a few and a few will work right? That wouldn't be a good corporate position to take right? And really again I think we can all see the potential in AI and we've all had the nice little honeymoon period and some excitement around it. What we now need to do is actually get back to the real work of ensuring that this technology is actually rolled out properly with the
00:42:13 - 00:43:14
appropriate safeguards and the appropriate quality so that it's a value add for humanity and we're not just a worse version of customer service, a worse version of interacting with a company than you had previously and really just a downgrade. Like that makes no sense. That makes no sense. Okay. >> Yeah. This shouldn't be a cost-cutting cost-cutting stream. This is because we we saw as you know, we saw it in the previous iterations of customer service improvements that we were trying to use
00:42:43 - 00:43:40
chatbots and so forth to save costs. Now it's actually it's about it's about making that service actually work and to improve the quality of that experience. And so also now we are seeing interesting leaps in in in that space and and yeah, happy to be part of that that that revolution. >> Brilliant. Well, you see, thank you so much for coming on the podcast. You are a star as ever. I'm always super impressed with how they how how you can get across really quite complicated so
00:43:11 - 00:43:47
so simply. So thanks very much and we will love to follow up in the next six months find out how you guys are getting on. >> Super. Thanks Colin. A good connecting again. >> Thanks for listening to this week's episode. To make sure you catch us again, please follow us on your favorite podcasting platform.