How to Enable Agentic Commerce with Nell Thomas, VP of Data at Shopify

Nell Thomas: Yeah. That's just more of a personal experience I live through every day. It's a good question.

I think ideally what Shopify is trying to do is be this sort of... shopify's always been an operating system for commerce, and what we're doing right now with AI and agentic commerce is just extending that into the newest frontier of technology. And in terms of types of merchants, we wanna be able to provide that for all different variants of entrepreneurs, whether that's like very small stage, building, a jam business or, the giant entrepreneurs that we know of BarkBox or Skims or Aldo, well-established merchants.

So ideally, we're building tools that give all of those different type of businesses what they need to be able to connect with buyers, and similarly for buyers to connect with all types of brands. So that's a long way of saying I, I think ideally the goal is for this to be agnostic to the type of individual product being purchased.

But right now, I'll say we d- we certainly see it working very well for places that have, very spec-driven types of tools, technologies that people might buy that need a lot of very detailed search-related queries, and that's where you can see agents do a really great job of parsing through options.

Richie Cotton: Okay. So I'm thinking then something like electronics, where you've always got like 20 different drop-downs you have to go through selecting oh I'm gonna filter on, I want this particular spec and that particular spec, and so this just makes it a bit more natural. 

Nell Thomas: Absolutely. I think that's a really great example.

Richie Cotton: Okay. All right. And you mentioned that, Shopify is the operating system for commerce then. So I guess that's lots of different stages in the buying process. Do you wanna talk me through, like, how AI is used in all these different steps? What's the flow and where do you use AI?

Nell Thomas: Absolutely. And this is where I will talk both about what it's like for merchants who are building up their businesses on Shopify, and also for buyers, 'cause obviously AI plays a role in both sides. On the merchant side, we're thinking a lot about what we call agentic storefronts, so which is the ability for a business on Shopify to be able to very easily connect to these new discovery sur- discovery services like ChatGPT or Microsoft Copilot or Perplexity.

Places where users are increasingly gravitating towards discovering the internet, , and we're making sure that it's very e- easy for merchants to plug into that, and making that under the merchant's control part of their existing Shopify admin experience where they're used to managing their business.

And again, this is about lowering friction, lowering toil for a business. 'Cause if you think about a business they're just trying to make sure they're, building the things people wanna buy, and they're connecting with their buyers not necessarily having to be experts on every emerging technology.

From that perspective, agentic commerce is really about bringing it into merchants' lives in a really easy way. On the other side, we have offerings like the Shopify Catalog, and this is something that allows allows us to ha- really help merchants Take all of that data we have about the products and catalogs each of them offer.

So the maybe tens or hundreds or thousands of SKUs a merchant offers with all sorts of attributes like dimensions or sizes or colors, and helping structure that, helping make that AI-friendly, and making it easy for that source of truth product data set to be available for platforms to build on top of.

And I'm offering it as an API that, obviously we can, as Shopify, make sure we're building experiences on top of, but also that other parts of the ecosystem can plug into as well. And we're, again, m- making it so that all of those billion products across millions of merchants are accessible via the emerging ways AI agents are looking for products.

Richie Cotton: Okay. Data featured quite a few times in that. And yeah I'm thinking if agents are looking at your data to try and figure out what should be bought, there's an inherent sort of assumption that the data's right there. So talk me through, what do you have to do to make sure that you're providing good data for the bots?

Nell Thomas: I just want to say making sure we have good data is my lifelong pursuit, regardless of bots or no bots. This is just, to detour slightly, as someone who cares deeply about data foundations and data quality, I think we're always in a pursuit of making sure that the foundational layers we have of trusted data are as good as possible.

And what that looks like in practice is measurement. So making sure we are measuring quality, measuring how well we are able to map merchants' products back to the correct attributes, like I mentioned before. Or even we have s- this idea of being able to identify the same product that appears in multiple merchants' catalogs, right?

So there's repeated SKUs, so how often can we actually identify those? Making sure even things like how quickly our APIs are responding, so latency, reliability, those are important things we measure to ensure that- Those foundational sort of sets of data we're building are as trustworthy as possible.

So trust looks like the data's right, the data's available, the data's fast, and it can be improved upon, right? So also cap- capturing feedback and making that feedback part of the loop of self-improving data sets. 

Richie Cotton: Okay. That's cool that you are just basically monitoring the data quality and hopefully having some kind of feedback loop.

I'm curious for individual merchants, like what do they need to do? 'Cause y- you mentioned that you have a lot, s- very small merchants. Y- you mentioned like jam sellers, and I guess like someone who's selling jam isn't necessarily going to be like super enthusiastic about data management as a general rule.

It'd be cool if they were, but I guess we can't assume that. So talk me through what do you need to do as an individual merchant to make sure, okay, I'm providing the right data? 

Nell Thomas: Hopefully what we're doing on the Shopify side is making it as easy as possible for merchants to, when they're managing their stores or they're uploading or creating new products that they might be offering, we're making that enrichment of the product offering, so adding in all of that additional metadata that we need, as easy as possible.

And increasingly, we can do that- Even by doing something like when a merchant uploads a photo, that we can identify using AI magic what is in that photo, and what that infers for us about the type of product it is, and the additional attributes that need to be suggested. I think in this way, what's happening right now with AI is not that different from what Shopify has always tried to do, which makes it easy for merchants to spend their time where they want to, which is like the creativity of running a business, and not on the sort of paperwork of correctly updating a thousand fields inside their admin.

But making it easier and easier to correctly infer the information we need to have to make sure we have all that rich data for the product catalog without it being a huge burden on an individual seller. 

Richie Cotton: Okay. Yeah. I think nobody wants to be, like, editing JSON files or something just to be like, "Oh, yeah."

Nell Thomas: Through a thousand drop-downs saying, red, yellow, orange. Yeah. So- 

Richie Cotton: yeah. So I love that these sort of things are being automated. And I'm curious how do you think this is gonna then change the experience for merchants? What do you have to do differently suppose you've got all these AI tools?

Nell Thomas: Again, I think my impression of this, and certainly this is not an area where I'm a deep expert in the being a merchant and what that day-to-day looks like, but I think ideally the tooling that Shopify is providing is creating the confidence for merchants that they are able to reach buyers, which is ultimately what they care about, on these agentic channels without having to radically pivot how they're spending their time and energy.

Because again, most of these merchants and these entrepreneurs wanna be spending their time on developing new product offerings, connecting to buyers, collecting consumer feedback, building and growing their brand and their business, and that what Shopify can do is- Make that as a natural extension of their existing workflow as possible.

So that's where automation comes in, that's where, really building this into the existing merchant experience comes in. I think ideally, as merchants are managing their stores we're using more and more of what we call our Sidekick, which is our merchant assistant, to make that also very much like that...

And what we're now getting used to as a sort of natural AI chat experience, where s- people can just ask for what they need out of the platform, and it responds to those asks very intuitively, as opposed to having to search through a million settings, fields to find what you need. You can really...

Merchants can now interact more natively with that Sidekick assistant. So I think that's another example of where we're trying to make it a natural extension of how merchants are spending their time as opposed to an additional set of work. 

Richie Cotton: Okay. So the idea is that you essentially get all the agentic stuff for free, and you can, carry on making jam or whatever you were doing before and focus on that instead.

Nell Thomas: And I would say many entrepreneurs are repeat entrepreneurs too. So it's like many of them wanna be thinking about what is that next business that they wanna be building, how do they expand and grow their business? And Shopify is right a- right there with them to help them through that expansion journey.

Richie Cotton: So I'd love to nerd out a little bit about some of the underlying technology 'cause I know there've been a few of these different protocols that cropped up for for commerce, and Shopify's been quite involved in the Universal Commerce Protocol. Do you wanna tell me what that is and yeah, how the, all these things fit together?

Nell Thomas: Yeah, absolutely. Yeah. Universe- Universal Commerce Protocol, UCP, it's a mouthful. What I'll say very simply it is a shared set of rules that lets any AI agent shop with any merchant store. And again the internet's full of protocols. Many of them are really critical actually though to how the internet works.

And with AI entering the picture, it's really a time where we are now the- forming the new set of rules by which the internet will work, and specifically work for agents. And that's what U- UCP steps in to do, is to help provide a leading example of how AI platforms interface for things like product discovery, the buying experience, checkout to make it a shared foundation so that instead of each individual AI platform, so ChatGPT or Microsoft Copilot or Google AI mode, each of them individually having to negotiate on how to talk to each merchant or each type of online seller, instead the UCP Lets a merchant's checkout rules, discounts, terms work consistence- consistently no matter which of the agent initiates that interaction.

Richie Cotton: Okay. It's really just a way of making sure that there's a standard, for how agents interact with your shopping site, and that consistency to make it easier to, have agents that can shop on multiple sites. Is that about right? 

Nell Thomas: That is perfectly said. And again, this is all about making it easier and abstracting some of that complexity away from the merchants so that they can just make sure it works.

And we're, leading that with Google and 20-plus retailers. We're very heavily involved to make sure we have that consistent standard going forward. 

Richie Cotton: Okay. And I guess from the merchant point of view, do they need to do things to, to implement this, or is this just stuff that gets set up by you?

Nell Thomas: Exactly. It should be things that are happening behind the scenes. And again, as this protocol is adopted as the consistent set of rules, it should work for all those AI platforms. 

Richie Cotton: Now, you mentioned checkouts, and this is interesting 'cause I saw a talk by Daniel Danka from Walmart recently, and he was saying how agent commerce is great, but when it gets to checkouts, bots aren't really as good as the existing technology, like checkouts for websites have been around for decades and they're mature.

I'm curious at your take. Are there any kind of limits to what you think bots should be doing at the moment? 

Nell Thomas: Yeah. I think this kind of verges into the question of like, where do humans vend to the occasion versus bots? And I... I don't think it's as simple as, AI buy stuff for you and people are abstracted.

I think, we talked before about discovery, right? There's the discovery part of commerce, there's the brand interaction part of commerce, and then there's the transaction, and I think agents will be involved in each of those on kind of a different timeline. I think today, right now, what we're seeing a lot is the focus on discovery and the pre-purchase with humans still at that loop around interaction and transaction And again, that's where I think our data infrastructure is really focused on powering that discovery in particular, because I think we-- many consumers will say discovery is a pain point.

And I think many merchants would say, "We want to be able to reach more buyers." And so it's a really a good example where I think the technology can make things better for everyone. But I don't think right now it needs to be focused on the buying piece. I think that's actually, in some ways the buying piece is the straightforward piece.

The harder part is finding the item, considering it, evaluating it, and getting to the place where you have trust in the merchant to buy it, and I think that's where we still have a lot of work to do. 

Richie Cotton: Absolutely. And certainly there are an awful lot of websites where you think it's just really difficult to find the right products that, that you want on the site.

So I think my, my test for this, I always want like new, like amazing shirts to show off on camera. And yeah if a bot can match my taste in clothing, I think that's the point where we know the technology's matured. 

Nell Thomas: Y- And just to add onto that, I I don't consider myself someone who has amazing design aesthetic taste, and yet I have strong opinions, which is a bad combination.

But what I've loved doing recently is uploading a photo of a piece of furniture I currently have and saying, "I'm trying to buy a dresser that coordinates with this existing chair. What are the styles?" And now having that visual and textual support to say, "Oh yeah, like this is a, mid-century modern piece that has some boho elements, and the type of item you might be looking for is that."

That is a whole expansion of my abilities as a consumer that I didn't have before, and it makes me really excited to actually buy and discover more, because I also feel like more empowered as a buyer to be like, "Oh, I maybe, I couldn't articulate what I was looking for before." Like you said, Richie, it's hard sometimes to know exactly what I'm looking for.

But this, now I have this tool that can help me articulate myself in a way I didn't before. All right. Extend, expanding upon the from the idea of agentic commerce, we were talking before about, a lot about data quality and how you need to have like good data in order to work in an agentic world.

Richie Cotton: I feel like for data teams in particular, you've got to be able to deal with agents everywhere. And so what do data practitioners need to worry about in terms of preparing for this agentic world? 

Nell Thomas: I think this is a thrilling time to be a data person in this agentic world personally.

But I do think it is changing the game a little bit. And- one of the fundamentals of great data work and great data science work is building trust in your objectivity and building trust in the underlying data quality and data assets that you're using, right? So people need to know when you're doing a, deep analysis of something that you're using the correct data, that you understand the problem space, and that you are speaking objectively as you tell the story of what you found.

And in some ways, that is the same issue we're having right now with AI, where AI is making it easier to do a lot of the analysis faster, right? So I can now throw a model at a problem of "Hey," "help me understand the drivers of change in this complicated space," or, "Help me improve the forecasting model I have for X."

That's, it gives me a faster and maybe a deeper ability to drive at the problem space, but it still relies on having really high-quality data that is well documented in a way that I am confident that the agent is correctly utilizing the dataset, and that as I get the sort of findings back from it, that I've been able to validate them and build confidence in a way that I can maintain the objectivity and trust that I have as the person speaking for it.

And so in that sense, the old adage in data of garbage in, garbage out is the same. You need to make sure you are avoiding that situation, you have high-quality data going in, that your underlying data engineering is top-notch, that you have rich metadata and documentation around your datasets that when you use a, an agent to help with your analysis, that it is correctly interpreting your dataset, right?

And you're giving it the right product and business context to be able to do and then on the flip side, that you're really, you are stress testing it, and like you, you are taking accountability as a data person for the work that is happening by that agent in that data, and that you can then interpret and speak for those findings in depth.

So it's changing the workflow a little bit, but I think the fundamentals of the work are the same around high-quality data and objective truth-telling. 

Richie Cotton: Okay. So that's interesting, the idea that the data team is now almost responsible for making sure that the bots do the analytics right.

So yeah, that's interesting. So I, I like, I did like your analogy before about an operating system for shopping. Is it like a new operating system for data then? 

Nell Thomas: Yes, and let me expand upon the storytelling a little bit about what happened at Shopify. So a- actually, every place I've ever really been, you're usually in the process of Migrating to a new data infrastructure, 'cause it's an age-old problem.

Ugh, our data infrastructure needs an overhaul. We're going to upgrade it. We're always modernizing and innovating. And along with that often is ensuring that your central data assets and your, maybe your data warehouse has the most highly trusted data in it. Shopify did that process back in 2023, where we overhauled a bunch of our data infrastructure, we rebuilt our data warehouse, we rebuilt central data assets.

The reason I'm going on this tangent is because a lot of that was about solving this problem for having a source of truth layer. And we got a little lucky in that we did that right at the time AI was also emerging, so that we now have this amazing source of truth of highly governed data that had been the focus of a lot of scrutiny to make sure that it was trustworthy and trusted.

That gives us more confidence, again, when we put, when we throw an agent at a problem, and we give it the dataset, and we give it the corresponding documentation around that dataset, which is a lot of labor to get that in good shape, it can correctly do its job. And then we can create scrutiny on top of that, and there's lots of ways we can also create measurement on top of the agentic kind of analysis to ensure that it is, we're calling the correct tables, that it is interpreting the columns correctly, that it is, writing its code correctly, and that it's ultimately returning correct results.

So we can also measure the accuracy of the agent. But again, I think a lot of the responsibility here is still on the human to oversee that sort of workflow and ensure that at the end of the day, they are taking accountability for the outputs, and that they are confident it will lead to the right insights for the business and for the products, and ultimately, of course, for the users.

Richie Cotton: Okay, yeah. Lots of different layers there. You talk about having things like documentation just to make sure that the agents understand what's going on, and then you talk about having measurement to make sure that it's getting the right answer, things like that. So it feels like there's a lot of setup and I guess administrative work there.

So do data teams have to spend their whole days writing documentation to, say, describe datasets? 

Nell Thomas: In some ways, hopefully they were doing that before Warren, because that's the thing. Before we had agents to do this work, we had humans to do this work, and the humans also needed documentation.

I actually, I remember in one of my first jobs working in data, there was one person on the team who knew in detail the meaning of every single column and every single dataset, and they're basically were a living librarian for the data, right? They were like, "I can recall exactly what, orders_V2_trusted means versus orders_V2_, like canonical."

And but that was written down nowhere, but it was in this person's head, and that's an anti-pattern, right? You-- it was an amazing job security for this person, but it was it was not scalable, and when you, the, when you tried to grow the team, you get bottlenecked on the one person who can understand it and explain it.

So I think having, anytime to have a high-functioning, scalable data team, you need to have great documentation, and it is work, and it's real work that often data engineers, data scientists take on that often was not really celebrated or praised. I think it's so important. I think right now we're seeing the fruits of that labor because now we can see how it's not just other humans that are being unlocked, but it's also the vast capacity of our AI agents as well.

So yes, to answer your question, I think documentation, like metadata around a table and saying, "Here's all of the meanings in human language of those columns." Comparing that also- With business and product context. So saying, in addition to having an... Let's say we have a dataset that explains our Sidekick agent I mentioned before, Sidekick.

It ex- we have a dataset that it talks about or doesn't talk about it. It stores information about interactions with Sidekick. We also probably want some information about what is Sidekick. Sidekick's a, it's a name that people might not be able to interpret. Who is eligible to use Sidekick?

What countries has it been launched in, when? What is all the information we give a merchant about Sidekick? So all the sort of lay knowledge that a PM might have or a business owner might have, or even a merchant might have about this thing. If we give that to the agent as well along with the dataset, it makes the results that much more accurate and more usable out, out of the get-go, because we're just giving it upfront a better dictionary to use.

Richie Cotton: Okay. Yeah, one of the big problems is historically, humans never read the manual, but AI agents at least will. If you give them a product manual, then they- 

Nell Thomas: Exactly ... 

Richie Cotton: hopefully if it's in context they'll remember what's going on. Okay. There are quite a few other things beyond documentation that have been popular recently.

I think about a data catalogs to make data more discoverable and semantically so the agent knows what to use. Do you wanna talk me through what other bits in your stack you need then just to make sure that all this data foundation works? 

Nell Thomas: We have the layer that is the raw data itself, so to speak, or I should say the structured data, the model data itself.

We have metadata associated with that table I mentioned already, which is like what do the columns mean, what does the table mean? You might have a semantic layer. A semantic layer will map the business definition of a metric to that table. So let's say we have a metric that's like number of Sidekick messages per day.

How do you actually calculate that off of that table? That will be part of the semantic layer that kinda maps that metric to that actual definition. So that's another definitely an important part of the set of context we're giving to the agent. I mentioned before product or business domain context, so that might be something like the help center documentation that we give we have about Sidekick, or the internal documentation we have on our kind of internal wiki about the product, so that we're further enhancing what can be understood around it.

And then- obviously we wanna make sure that we're using measurement around that to understand whether or not the agent is correctly passing these different steps of recalling the correct dataset, interpreting the dataset correctly, writing code against it correctly, understanding the problem space when prompted correctly as well.

And the other thing I'll really mention here in terms of all those steps is making sure all that work is happening out in the open, because we wanna make sure that the same way we would expect humans to check in their code and be transparent about the work and their workflow, we want the agents to do the same, so we can also scrutinize the work and understand the q- the decisions being made at various steps, understand the code that generated the work product.

And so making sure that we have workflows where that code is written in a way that can be checked in and can be later interrogated and interpreted that also makes it really great from the perspective of having a internal repository of those deliverables that can be enhanced over time.

Richie Cotton: Okay, yeah. That level of transparency does seem important 'cause sometimes what the data team does can be a bit of a mystery, and then the people who you're working with like other teams are like, "Oh, I don't know what this data is." So I suppose if you are doing th- all this, has it changed how your data team works with the rest of your organization?

Nell Thomas: Yes, and I think this is, by the way, I should mention, this is all, an evolving and rapidly changing environment, so I think that's part of what makes it really exciting right now is to see how we're still figuring all this out together, and that's part of the fun part being about Shopify. It's a super high agency, intense, like very curiosity-driven culture.

And so- The data team is figuring out how to evolve and keep pace with the emerging tools, as is the rest of the company, and we're able to do this in concert in a way that is very it's very thrilling. But to get back to your question directly, we, there's a lot more self-service tools in it.

Or self-service tooling is always a goal of data teams to make sure that other fellow colleagues can get at questions as easily as possible so that we're not a blocker. It's why I think great data teams invest in foundations, because foundations make self-service more trustworthy and better, and that makes it easier for a more democratized data access.

And I think, again, AI is enabling and furthering that because, again, it's so easy. We have many ways now of your, someone who works in our customer service team or someone who works in an engineering team or a PM to be able to ask a qual- question that's relevant to their job and get back an answer easily.

So we have a MCP that connects to our data warehouse. That's embedded in our internal chat as well as accessible via, cloud code or a lot of people here work in Pi using whatever model they want to. And so we've tried to, again, really bring that data accessibility along with really straight, strict, security, privacy and compliance, but bring that together so that people can get the context they need to do their jobs.

And we can also get feedback ab- about the data systems themselves. So in that sense, again, it's a continuation. It's probably an enhancement of that self-serve flow. But fundamentally, I think many people still turn to the data teams to be their trusted partner in their sort of that end product accountability.

And in, and at the end of the day "Hey, did I get it right? Did my agent get it right? Is this the right way to be thinking about it objectively?" And I think that's where data folks continue to play a really important role as partners to the organization. 

Richie Cotton: Okay that is interesting. I think self-service analytics has been this kind of dream for at least a decade, and we're getting closer to that because if anyone can just chat to an AI and they got all the data analyzed for them, then we get a lot closer.

But I do like the idea that you wa- often people want that, reassurance from someone in in the data team just to be like, "Is this right or not?" 

Nell Thomas: And I think, yeah. And I think that there's also a wide range of data questions that are asked, right? And it, there are some questions that are like, "Hey, this is just a kinda like small, I just I need to know like how many people in Poland yesterday did a thing," and it's, I just, a rough estimate's fine.

And then there are questions where it's like, "I'm deeply trying to understand- The financial breakdown of X, Y, and Z, I need a high degree of accuracy and confidence 'cause I'm gonna make a major partnership deal based on it. And those are two different types of questions that require two different levels of confidence.

And I think now we're seeing that data folks can spend more and more of their time on those really thorny, complex strategic questions, as some of those more like micro questions that may require also a little bit less, a little bit less confidence, a little bit more just like I need a rough range here.

Those can be handled in a self-service way, and that shift I think is real and it's happening. And I think we're seeing something like 40% of our Shopify emplo- employees are using data tools daily, like those data access tools, so a really high degree of that self-serve ability. But redirecting the data team to those hardest problems where we wanna make sure we can go very deep on rigor.

Richie Cotton: Absolutely. Yeah, definitely agree there's a very wide range of questions you can answer with data, and some of them, yeah, you do have to be a lot more, paranoid in terms of is this the right answer or not? And this seems to be a bigger question around AI because generative AI is notorious for hallucinating sometimes.

It doesn't always get the right answer. And so once you start introducing AI into a lot of your data workflows, how do you make sure that you're not gonna suffer the consequences of like wrong decisions or hallucinations or mistakes from the bots? 

Nell Thomas: Yeah, I think this is where like ultimately humans are still accountable, and I think it's... I hold my team to a really high bar of rigor, objectivity, and accountability. And and I think that's, that's not just true of data folks, that's true of everyone at Shopify. I do think that there are no excuses for passing off shoddy AI work as shoddy human work, like shoddy work is shoddy work, and people should be ex- you know, expecting their, the work that they're doing with AI might be getting faster and deeper, but it should also be getting better and not like just easier. 

Richie Cotton: Yeah. So I quite like the idea that if you get a wrong answer, you can just be like, "Oh yeah.

AI, am I right? It's the bot did it." But I think you can maybe get away with that once and then after that your boss is not gonna be very happy about it. So you mentioned that the the job that a lot of your data team are doing has changed a little bit. Has it changed the profile you're hiring for?

Do people need different skills now? 

Nell Thomas: Yeah. A little bit. I, in the sense that I think, I think curiosity and the ability to really re- excel and thrive in change. Those were v- those were values that ShopBuy's had for as long as I've been here, and I'm sure longer than that. But I see how much, h- how amplified that is because things are moving so quickly right now.

Or how, how we think about our jobs is changing as we just described, and it's again, I don't have all the answers here how that will continue to unfold and change. What I do know is it will continue to change, and so I need to hire people who are excited about that possibility and are not afraid of the disruption that is coming and happening because, that is we know that's true. We don't know the extent of it or what it'll exactly look like or how it will change roles. We know things are changing, and so people who Are excited to keep pushing and learning. I always say find the edge of where you're comfortable technically and push past it. And I'm seeing a lot of that happening right now in a really organic way internally, and I think hiring in new folks who have that same attitude has been fantastic.

Richie Cotton: Yeah, certainly curiosity and the ability to deal with uncertainty seems pretty much essential for getting by in almost every role at the moment. Now, I know a lot of people who work in data science I found have a weird career history, and I think it's true of myself and true of yourself. Do you think that helps in data science, or is it just a coincidence?

Nell Thomas: It... it is very true that there are a lot of, And I say this with love and I consider myself one, but there are a lot of weirdos in data science. We tend to have unpredictable career paths. And I haven't thought deeply about whether that's a coincidence, but there probably, as a data person, I should probably say there is a pattern here.

I do think there's something to be said for folks who have the a lot of data folks have the technical skills without having gone full bore engineering. They have a lot of natural analytical and knowledge-seeking, truth-seeking abilities. I think they tend to also, a lot of data folks tend towards the slightly pessimistic.

They tend to be like, again, that sort of "I'm gonna find the truth. I'm gonna find the counterargument a little bit." I... So I think all of that is a little bit agnostic to individual career path, and so you find people who navigate their way over time. But yeah. I'm... i'm interest- I...

Obviously that's changed a little bit with the last, I think 10 years where there's a lot more data science programs and degrees. So you do see a lot more people who are just coming straight into the field. But historically, a lot of folks coming out of social sciences physics very, obviously out of consulting or finance even.

So I love that part of it. I think you get a lot of different perspectives and methodologies too that combine to make really rich and dynamic teams. 

Richie Cotton: Absolutely. It's very fun coming from a different background and ending up somewhere completely different. So do you have any advice for people who want to get into data or AI coming from a totally different career path?

Nell Thomas: Oh I would say the war- the water's warm. You're not gonna be alone. There are lots of people who end up pivoting through different career trajectories myself included. I would... In terms of advice, I think the most important thing is the obvious thing, which is do the work.

D- spend your time Getting deep on a problem. And that can be through a publicly available data set. It doesn't need to be one that you only have proprietary access to via job. There is so much data in this world that is available to discover and understand better. So I would say find a pro- fall in love with a problem and find ways to explore it using your skills.

Find the edges of your technical skills and keep pushing. Use AI. Hold accountability to yourself. Hold your agents accountable too. Hold a high bar for your own work, whatever tools you're using. And AI is just another tool, right? It's an amazing tool, but it's just another tool, and you need to figure out how to use it and hold it the right way for yourself and for your workflows.

But ultimately, at the end of the day, it's like the work is the most important thing. The ability to do the work, to solve the hard problems, to use your skill sets to the best of their abilities and find answers that illuminate something people didn't know before. That's amazing. And just do it.

Richie Cotton: Yeah, th- that should be a slogan. No I do agree with you there that you just need to try some projects, and once you come up with some interesting answers and interesting insights, then that's gonna stand you in good stead and inspire you. 

Nell Thomas: Yeah, I'm g- I think sometimes people focus a lot on some of the oh, I need to network or meet the right people or, make s- sure I have certain things on my resume.

And I think, I've, we've, I've certainly hired people in my career off of the side projects that they post about on their, on their GitHubs, on their personal blogs or whatever, personal websites, because that's always the most telling, is that someone is has that natural curiosity and that love of learning, and they wanna apply those skills, and that's what they wanna spend their time on.

It's just, it's so telling. And I, and then I know they'll bring that into their work too, and th- people who fall in love with their work, they make the best work, and they do the best. They tend to thrive. And that's ultimately what I'm looking for, is someone who's gonna come and thrive and want to do amazing work.

Richie Cotton: That's very cool stuff. I love that you've hired people based on side projects. And, we talk a bit about having a data portfolio and things like that, of examples of stuff you've done. Do you wanna talk me through some examples of these projects then that people have done that you've gone, "Wow, that's very cool, and I wanna hire you now"?

Nell Thomas: Sure. I will give one example. This is, this one's actually from a decade or more ago, so it's a little bit dated. But I hired a woman based on she did an analysis of her first names, popularity over time as tied to certain historical figures. She had a name that was very well known.

I won't use it 'cause I don't wanna reveal too much about her. But and she did this like great analysis about name popularity, but like tying the rises and falls of her particular name to world events that were happening. That was just an example of where wasn't a, wasn't like a crazy involved project, but it was like a really tight, nicely done piece of work that showed a level of like fun and yeah, like creative juices that got me excited and differentiated her out of a pack of people applying.

Richie Cotton: Nice. That's just a really fun project I think as well, and it's something that everyone can understand. It's not super technical that only three people in the world would care about. So yeah a fun idea for a project for anyone who's going to analyze some names. Okay, cool.

Just before we wrap up, what are you most excited about in the world of data and AI at the moment? 

Nell Thomas: I'm really excited to learn from my team right now. I feel like we're running a million miles a minute, and I'm just doing my best to keep up with them, and that's an amazing feeling because like the level of, I've used the word creativity a couple times, the level of creativity in how people are tackling the problems that we can solve.

I a member of my team, he got access to some, to Google Deep Think and threw a problem that he grapples with at it and did some amazing work on it and ended up in a presentation for Google I/O because they were like, "Oh my God, this is an amazing example of how to use this model." And that was just like, the type of thing that like, that's just one example.

But there is so much happening right now where people are applying their brains and these new tools to these hard problems and uncovering like new layers of how to tackle the problems. And not gonna... By the way, that also is managers too. Like managers I'm finding all sorts of fun and exciting new ways to extend my skills as a manager using AI.

And So yeah, I'm just, I'm very excited to learn from other people right now and that's people inside my team my peers at other companies in similar roles. I am-- I have some amazing colleagues at other companies that I just love sharing and hearing how we're, this field is evolving and how we're all trying to navigate this moment of uncertainty, but also immense possibility.

Richie Cotton: Absolutely. I'm-- This, yeah, you're right. There's just so much going on all the time. It's yeah, impossible to be bored, and whatever stage of your career you're at, you can find some uses of AI, and it's gonna change how you're working, and there, there's always something exciting going on. Nice. All right.

So just finally, I always want more people to learn from, so whose work are you most excited about at the moment? This is a great question. I'm gonna, I'm just gonna to give, I'm gonna give a shout-out to my team member, Andrew Plourde, whose video was featured by Google, like I mentioned last week, about using Deep Think for some of our growth forecasting challenges.

Nell Thomas: So I, I need to plug that because I think it's a great example. And there's a nice little video about it online. Yeah. I think I'll share that one. 

Richie Cotton: Okay. Deep Think for improving your forecasting. That sounds like a very cool use case. All so yeah I'll look out for Andrew's work.

All right. Thank you so much for your time, Nell. 

Nell Thomas: Absolutely. It was my pleasure to be here.