--- id: "workshop:building-ai-ready-open-data" type: "InnovateUS workshop" title: "Building AI-Ready Open Data" url: "https://innovate-us.org/building-ai-ready-open-data" date: "2026-04-15" duration: "60 minutes" series: "Opening Data for AI Innovation" instructors: ["Luke Keller","Jose Plehn"] moderators: ["Stefaan Verhulst"] recording: "https://www.youtube.com/watch?v=ogpbSR8KkwM" transcript: "automatic, from the recording" source: "InnovateUS (innovate-us.org)" licence: "CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/), as stated on innovate-us.org" --- # Building AI-Ready Open Data InnovateUS workshop · 15 April 2026 · 60 minutes Examine what makes datasets usable for AI, including key design features, interoperable data ecosystems, and governance practices that support secure, responsible reuse. ## Learning goals - Identify what makes a dataset AI-ready — key design features that allow AI systems to ingest and interpret open data effectively. - Understand key concepts for structuring and maintaining data ecosystems which ensure that datasets can be securely combined, analyzed, and reused across agencies and tools. - Explore governance practices, processes, and infrastructure needed for effective and responsibly AI-ready data management. ## Speakers - **Luke Keller**, Vice President, AI Products @ Unanet, former U.S. Census Bureau CINO - **Jose Plehn**, Founder & CEO @ BrightQuery & OpenData.org ## Moderators - **Stefaan Verhulst**, Co-Founder of The GovLab, Co-Founder and Principal Scientific Advisor of The DataTank Part of the series [Opening Data for AI Innovation](https://innovate-us.org/workshop-series/opening-data-for-ai-innovation-series). Recording: https://www.youtube.com/watch?v=ogpbSR8KkwM ## Transcript Automatic transcript of the recording: names and terms can be misheard, and the times can be a few seconds off. Each time links to that moment of the video. The transcript alone: https://innovate-us.org/building-ai-ready-open-data.txt (plain text) and https://innovate-us.org/building-ai-ready-open-data.json (every segment with its start and end second). **Stefaan Verhulst** [0:00](https://youtu.be/ogpbSR8KkwM?t=0) Great, thanks so much Agueda and thanks everyone for joining us and as was mentioned, I'm Stefan Verhulst. I'm the co-founder of the GovLab here in New York where I'm calling from and also of the DataTank based in Brussels, but I'm also a research professor at the engineering school of New York University and I will be the moderator of today's session which is part of a series that is focusing on opening data for AI and we've started working on the concept of opening data for AI within the context of some of the work that we have been doing at the GovLab which is the work that we have labeled the fourth wave of open data because at GovLab we've done a lot of work on how do we change the way we make decisions leveraging new technologies and of course data is one of them and so over the years we've been focusing on how can you open data in ways that then can inform decisions and what we have seen is of course a evolution over times of the opening of data ecosystem where in the initial the first generation we were focused on freedom of information the second generation was really about providing data and especially government data via data portals which remains of course important the third wave was then really about how do we also have access to private sector data including for instance access non-traditional data for for instance official statistics but then obviously we are now in a new wave which we call the fourth wave of open data where of course open data and AI intersect and for those who joined the first session of this series I tried to explain the variety of ways how open data and [2:01](https://youtu.be/ogpbSR8KkwM?t=121) AI intersect where open data becomes a resource for AI in order to improve training in order to develop application or use data for applications such as rack applications but also we have seen that of course AI can be a tool for democratizing access to data including providing a conversational layer or leveraging AI to develop metadata standards describing data and so on and so the previous session we also had the pleasure of talking with the chief innovation officer of Boston where we looked into how Boston has used MCP protocols for providing a conversational layer with open data but today we are joined with a team I would say of experts that have plenty of experience to share with regard to how do we make open data AI ready and also how do we start for instance MCP applications for providing a more conversational layer with data and making of course also official statistics for instance AI ready so with that delighted to introduce two of our participants today and so we have Luke Keller who has recently left the US Census Bureau and is now a vice president of AI products at Unanet where he leads a portfolio of AI first SaaS products helping thousands of customers optimize and grow their businesses across AAC and government industries but before and prior to Unanet Luke was the chief innovation officers chief innovation officer at the US [4:08](https://youtu.be/ogpbSR8KkwM?t=248) Census Bureau which as we all know is the largest statistical agency within the US and at the Census Bureau he developed a whole range of AI products and I will ask Luke in a second to explain more what those AI products were and also some of his experience so far in addition to Luke we also have the pleasure to be joined today by Jose Plin and Jose is the CEO of Bright Query and also oversees Bright Query's open data initiatives for those who are interested in check it out at opendata.org which is Jose we need to have a conversation how you managed to obtain that URL which is of course I'm sure a very a very which is a URL that many probably wanted to obtain but he is also on the board of directors of the AI alliance and does a lot of work on actually developing standards including standards with regard to metadata and standards with regard to evaluation of AI as well and has an academic also background in a variety of contexts at UC Berkeley as well so with that let me first ask Luke to explain a little bit about your activities at the Census Bureau and also I think you were kind of one of the pioneers within the US government as it relates to leveraging protocols like MCP but also to start thinking about what does it mean to be AI ready especially at the level of **Luke Keller** [6:03](https://youtu.be/ogpbSR8KkwM?t=363) official statistics so over to you Luke. Thanks Stefan yeah so it's been a really interesting journey for us or I should say for the Census Bureau who I was employed with up until very very recently a few weeks ago we I actually started the Census Bureau in 2017 where I co-founded a team called XD which effectively served as the Census Bureau's AI center of excellence and over the years that team worked on a lot of really interesting you know internally focused projects thinking about how we can leverage AI machine learning especially non-LLM type projects non-large language model type projects but when large language models hit the market I think it fundamentally changed a lot of things you know particularly for agencies who see themselves as you know open data providers like the Census Bureau and what I mean by that is you know we realize that you know these large language models depend on such a tremendous amount of high quality open data in order to you know answer any and all questions the user may have either directly related to or tangential to the topic that they're talking about and your agency's data your data even just as a broader data provider whether you're a federal agency or not your data is part of that calculus for how those you know those models determine what to say and what not to say to a user and so at the Census Bureau what we were seeing was kind of this like age-old problem which was you know we were deeply concerned about how our official statistics get represented on the internet so like on Wikipedia and other websites like is it accurate is it faithful to the original statistic is it nuanced you know contextualized etc but now with large language models that problem just got you know exponentially more complicated you know the the context you know the lack of context even things like deficiencies in our API token limits all these other you know characters you know characteristics about large language models that make them both like wonderfully you know competent and and wonderful you know beautiful things to use we're also kind of working against us and our aims to have our official statistics be you know faithfully represented and so at the Census Bureau we undertook a project last year [8:04](https://youtu.be/ogpbSR8KkwM?t=484) to build a model context protocol server which was you know of course a standard that originally was released by Anthropic in November I think of 2025 if I'm not mistaken or 24 I think it was 24 and as now yeah November 24 thank you Jose I was like trying to get my timing right um but which is now of course run by I think it's the um agentic AI foundation if I'm not mistaken yeah and what we found most interesting about that you know standard was one it's an open source standard which is fantastic right so as an open data provider adopting an open source standard just makes a lot of sense and two I think the thing that that standard does for a lot of open data providers and we're happy to talk a lot more about this because I'm hearing a lot of like is MCP even relevant anymore given you know all the things that are happening in the space but one of the things I think it does really well that's really great for open data for providers is that it basically lowers the baseline required to interact and get meaningful insights out of your data so if you can improve the experience with your data across anyone who chooses to use you your MCP server and their large language model stack it means you are reducing the expertise and the complexity both technically and otherwise to use that data meaningfully in your pipeline it also means as a data provider that your data is being represented far more faithfully than otherwise would so it's something that we felt was really exciting in our testing we saw some really great things you know in terms of um you know a b testing it against you know with and without with certain models and we're really pleased with what we solved and we're really happy to support that project so we released it publicly last year in September um so yeah happy to talk more about that but just you know wanted to pause and give Jose a chance to also introduce **Stefaan Verhulst** [9:41](https://youtu.be/ogpbSR8KkwM?t=581) himself too great thanks so much Luke and so yeah Jose stage is set for you so uh tell us a little bit about uh bright query but also of course the work you've done uh I guess with uh the U.S. government on kind of understanding uh the state of uh U.S. government data as it relates to AI readiness and and also uh and I think where Luke I think also was referring to in the past we had this kind of whole discussion data is not AI ready so we cannot leverage it for AI but then MCP came along and so uh and suddenly uh uh of course what was the kind of the barrier for using AI of using data for AI uh suddenly uh uh was less uh less high let's put it that way so eager to hear from you Jose uh what you've been up to and how that has evolved yeah absolutely so thank you for uh the invitation and uh thank you everybody for for joining this uh this important uh discussion so uh to kind of help set the stage I I thought it might be fruitful to give a little bit more background on the kinds of things that that I do personally and the kinds of things that bright query does with respect to these issues so as mentioned in the introduction I am founder and CEO bright query or bq as we like to call it I'm also the founder of opendata.org um I will be talking more about that but long story short that's a project that we are setting up as part of a Linux foundation like the AIF agentic AI foundation which will be governed by a group of tech companies uh from around the world which I can talk more about but it's being seeded by data that we're providing uh which again I can get into further details later I'm also on the board of an AI alliance if you're not familiar the AI alliance is actually open to uh non-profits research labs companies and so on the AI alliance is both a non-profit research lab and a trade association I'm on the board of that and co-executive director of it and the idea is that we we try to have a common voice around AI issues and as part of that role being on the board where [11:45](https://youtu.be/ogpbSR8KkwM?t=705) I get to interact pretty much every day with the leading foundation model builders and companies that leverage AI a couple other things I'll just mention briefly about touch upon this as well the agentic AI foundation that Luke mentioned which is a project inside the Linux foundation we're also heavily involved there Jim Zemlin who runs it we work closely with him and his staff and we touch upon areas around identity accuracy and governance so we're very much in the weeds with Anthropic of course who donated MCP to the Linux foundation and all the other companies that are participating you know in this effort last thing I'll mention briefly which is perhaps the most relevant and what I'll perhaps my comments will be centered around is that my company bright query has five contracts or awards that would be granted with the national science foundation and specifically the NCSES the statistical division of the NSF although this is on behalf of U.S. government overall to contribute and build the national secure data service or NSDS if you're not familiar with this this is the U.S. government's first initiative or attempt to sort of have one centralized AI powered hub run by federal government that basically taps into federal state and local public data sets although the word secure is in the name of the NSDS national security data service in fact its primary function is around public data so statistical data that the agencies would normally release and these five projects a couple of which are very relevant one is we built a factual bot as we call it so whereby you ask it a question about some statistical fact and it will fish out the answer directly from the agency from the right table from the right columns right time period etc with provenance with links [13:50](https://youtu.be/ogpbSR8KkwM?t=830) academic style so it meets that academic you know threshold of diligence second we built an AI readiness tool which you know I'll be talking about further in subsequent questions the readiness tool has an evaluator and a transformer the evaluator kind of assesses the AI readiness of your website or data set and that transformer helps quote unquote transform it third and I saw here from the poll and from the participants a number of you are involved with government kind of state and local agencies you might find this interesting we built a platform what we call the state AI explorer which basically taps into open data sets of agencies and allows a and has a natural language interface on top of it with templatized answers and whenever the answer cannot be found from the state level agency it's a event queries a federal agency to see if a federal agency has a right and has has an answer to help answer question fourth project we're doing is around customer service, so that there will be one centralized site that everybody goes to, if you have a question about government statistics. So we're setting up like a ticketing system and a bot, kind of a customer bot. And the last thing, I know it's quite a bit, it's called the Data Usage Platform. And this, many of you might find interesting, it's how is your data being cited? How is it being utilized? How is it being quoted, referenced, etc. In many many outlets, news, government reports, academic articles, and online forums, or kind of the open web, so to speak. So we have licenses with over 700,000 sources in the world, and we're tracking every single day, monitoring citations and usage of data. Because as you can imagine, that's one of [15:52](https://youtu.be/ogpbSR8KkwM?t=952) the big challenges, right? It's like, you want to know, how's your data being used? So these are all still kind of at the POC stage, but part of the hope, the vision of the various folks at the federal level working on this, is to get approvals to put this into production. But having said that, we'd love to work with each and every one of you over here. Please, you know, reach out to us, that's partly why we're here. And you can feel free to even email me directly, simple direct email, jose at briquery.com. But happy to answer questions here and dive into further detail. Stefan, go ahead. No, no, thanks, Jose. And it's important to know for the audience here on what are some of the projects you're working on. And again, I think they are super relevant, and I'm sure for the audience, issues that they are struggling with themselves in terms of how do you make data AI ready? How do you evaluate whether your data is AI ready? And then also, of course, how do you understand what is actually being used? But I do want to come back to this notion of, and Luke, you perhaps can also reflect on that, because especially as Luke, as you mentioned, the real concern quite often for certain data holders, such as statistical offices, of course, is to make sure that the data is being used for factual answers, right? [17:21](https://youtu.be/ogpbSR8KkwM?t=1041) That anyway, you ask a question, and that you actually get a fact back based upon official data, as opposed to secondary data, that which is quite often used for training, generative AI models. And so, Luke, if you could tell us a little bit about what are the challenges here? Because when we started working on kind of the fourth wave of open data, we were approached by several statistical agencies who, anyway, came and said, we have a problem. Our data is not being used for actually answering factual questions. And one of the reasons was because their data was not so-called AI ready. And so, can you tell us a little bit about what are the challenges for having data sets being used for this kind of factual AI, and how would you go about addressing them? And of course, you used MCP, because he has been referring to kind of ways to transform data in AI ready format. So, what are we talking about here? **Luke Keller** [18:30](https://youtu.be/ogpbSR8KkwM?t=1110) Yeah, okay, this is such a great question, and I could probably speak for like two hours about this. **Stefaan Verhulst** [18:34](https://youtu.be/ogpbSR8KkwM?t=1114) I'm going to try not. You have an hour. We have an hour. **Luke Keller** [18:36](https://youtu.be/ogpbSR8KkwM?t=1116) Yeah, yeah, yeah. How many minutes do we have, right? So, okay. So, we had such extensive discussions internally about what we were trying to accomplish, and what were the limits of the project, and what were the limits of MCP specifically? So, for those not familiar with MCP broadly, I'll just start at like a super high level, and then kind of like dig into our thinking. So, at a high level, MCP is kind of like a specific type of API for LLMs to interact with data in any kind of system service, et cetera. It's kind of like, think about MCP as a way to plug your LLM into a data source, and then MCP has some opinions that you code into it through tools, resources, prompts that basically say, these are the ways I'm going to allow you to connect with that data and get it back. What MCP is not, though, is something that dictates what you do with that data, or how you use that data, which is, again, like a significant challenge like you suggested. So, there's a few ways to look at this. One way is to say, so I guess I'll like say like the way we looked at it, and then I'll like add like other ways you can kind of interpret this problem, and other things you can do, other solutions that are available to you to help solve this problem. So, at a high level, like the position we took was the way LLMs are using our data today is not great, right? We would ideally have a little bit more influence, not control, but influence on how that data, or at least how that data is being accessed and used, and helping LLMs make the best decision possible on behalf of their users. That's as respective as possible of our data, but understanding that we do not control how LLMs work, right? We don't own the stack. It's not a website that you're going to, like we're not offering a piece of software, like we're offering a data connector. And the theory behind that was like, if we don't do this, then, you know, third parties, individual developers, et cetera, are going to have to spend a lot of time and energy trying to understand what we understand, and don't [20:39](https://youtu.be/ogpbSR8KkwM?t=1239) have the mandate that we have to use that data in a way that we know we're supposed to use that data, but they will not adhere to that, right? So, the thinking was, if we build something like MCP, that helps to deeply influence how a large language model behaves, then at the very least, we can kind of like build a rising tide to lift all boats. It's not perfect, but it will hopefully help to like reduce the kind of development and engineering cycles that other teams go through, in terms of like the technical integration with our data, which then allows us to know that our data is being at least faithfully used, meaning like it's coming straight from us. We can't dictate how they use it. There are ways with an MCP where you can kind of try to, through things like prompts. The problem we had with prompts, and again, this is just now like my like individual opinion and slash like opinion of the team at the time. The problem we had with that was we didn't know what other MCPs you were trying to integrate with. So, if you have an environment where you're just using the Census Bureau MCP, and you're just trying to ask census-like questions, that's great. [21:41](https://youtu.be/ogpbSR8KkwM?t=1301) But what if you're using an MCP, you know, from, let's say like you're using MCPs with like overlapping data. So, there was like a, you know, I don't know if you saw this last year, but there was like a DoorDash pricing index to track inflation, right? That's one way to kind of like track inflation. BEA data is like another way to track inflation. What if you're using all these things at once, and they're all saying prompts, you know, introducing prompts in their MCP that say, no, use my data first when someone asks a question about this. Now you have a bunch of MCP servers in a cage match, and you don't know what's going to happen. So, we really wanted to shy away from trying to like deeply influence how those systems behave outside of just being faithful to the standard. But there was, you know, so that's kind of like the strategy of the MCP layer. And then I'll add that there were discussions, and I think there are ambitions to go further with that project where you use that MCP server as kind of like a base for how you like, again, like lift all boats through a rising tide. But then you also use that yourself as the agency to then provide enhanced search experiences, internal applications, et cetera, where you're now not just releasing something that's useful to everyone, but then you're using it yourself, right? And so that, if you stand up those kinds of environments where you have more control, now you have further ability to dictate how a model behaves. Again, it's not perfect because large language models are, you know, if you're a parent, right, it's like kind of having a toddler. [23:03](https://youtu.be/ogpbSR8KkwM?t=1383) It's hard, right? But at the same time, like you try to guide it and make great decisions. And that's the best that you can do. So if you're standing up your own environment, you know, you have system prompts that you can use. You have an agents.markdown file that you can use. There are lots of other ways. And I think, Jose, you probably have far more expertise than me in, you know, Bright Queries experience kind of standing up systems like this. But for us, we saw the MCP project as like the data connector and something to help influence good decision-making. And outside of that, you know, educating users of that project and helping them understand that context was one way we thought about it. The last thing I'll add is that there's been some really interesting work too, to think about like what the future capabilities of large language models are and like how they will ultimately, you know, like make decisions. There are ways to kind of like offer MCP functionality that aren't just like serving data, but serving, for example, common responses to prompts that, you know, questions a user may ask. There are like holistic ways that you can kind of approach this problem in a way that were frankly very different than the tech that we took, which was we need to connect you directly with the data source and the extent to which we can make that data browsable, searchable, token optimized. [24:13](https://youtu.be/ogpbSR8KkwM?t=1453) Great. But there's other things you can do with MCP too that I've started to see emerge in the space that have been very interesting. Interesting. Thanks, Luke. Jose, **Stefaan Verhulst** [24:22](https://youtu.be/ogpbSR8KkwM?t=1462) you want to add anything to what... Yeah, absolutely. So, Luke, I feel you. I could also speak a couple hours, a couple of days, a couple of years on this. Absolutely. So, I should point out, by the way, this suite of tools that we developed, Census is one of the agencies that was part of it, along with BLS, Bureau of Library Statistics, you know, there's a huge agency with a lot of data, like Census, BEA, Bureau of Economic Analysis, and various others. We literally dealt with millions of tables and billions of data points. So, it was a massive, massive amount of data sets. Three things I can recommend. Again, devil's in the details. It's easy to say, but three very specific things that I can point to that we found as pain points with regards to AI readiness. Number one is AI systems do not like tables. Okay? So, you know, Python, R, MATLAB, you know, the traditional statistical programs, they love tables, but guess what? AI hates them. It doesn't think that way. [25:30](https://youtu.be/ogpbSR8KkwM?t=1530) What's critical is for the information to be organized by topic, and you decide what the topics are. So, think of it as a tree of information. If some of you are familiar with JSON, J-S-O-N, kind of a way of structuring information, that's the way AI thinks. That's the way engineers often think. And so, the way to think about it is you don't want your information, your data as rows and columns. Instead, you want it as information trees. Think of it as a tree of knowledge, a tree of statistics, and with many, many branches, perhaps millions of branches. And that's the way to think about it, you know? And you decide the naming of the branches and the trees, et cetera, and how you classify them and how they relate to one another. So, it's a completely different way of looking at information. That's number one. Number two is aggregation. As you all know, I mean, when we're talking about statistical data, the data is by definition aggregated, right? It's typically anonymized in some fashion, aggregated at least over 80% of the time by geography or industry. Geography, when it comes to people, sometimes demographic characteristics as well, like census has a plethora of tables where things are, you know, demographic characteristics by geography, you know, cross tabulations. And then, when it comes to business data, often industry, you know, or things like size, employment range, revenue range, things like that. [26:58](https://youtu.be/ogpbSR8KkwM?t=1618) So, you know, the cross tabulations and so on. Well, we intuitively know, you know, say, you know, counties roll up to states, and states roll up to the country, and zip codes have a messy relationship, you know, with counties, as some of you may or may not know, or counties have a messy, you know, states have a messy relationship with zip codes, and so on. And, you know, you have ZCTAs, you know, it all gets very messy. But we intuitively understand that. And when we code in statistics, we kind of, you know, build that in. Well, guess what? LLMs, they don't understand things that way. So, the explaining it, and having a way to account for it, to tell it, hey, look, by the way, these tables are kind of like a subset of a separate one. And if you add things up over here, it adds up to this. So, you know, the aggregation or disaggregation of statistical information that has to be explained, it has to be laid out, you know, in very clear terms. The third thing I will say, again, devil's in the details, but the third thing, high-level advice, you know, here is, we loosely call it stat rank. [28:08](https://youtu.be/ogpbSR8KkwM?t=1688) So, think of how Google works, or ChatGPT, when you put in a search, and Luke alluded to this, if you put in a search about some statistic, your agency, or the agency that produces the underlying statistic, with high probability will not be cited, at least not initially. Why? Because it's a popularity contest. That's what it is. Google and LLMs, what they do is they surface to the top the sites that are the most popular. Or, of course, sometimes they're incentivized to bring to the top the ones that are paying them to rise to the top, right? I mean, after all, these are advertising, right? And now even OpenAI is going to be advertising. So, you're kind of competing with the advertisers, and you're competing with the popular websites, you know, so if you ask a question about some economic thing, chances are it might go to the World Bank, you know, or the United Nations, or Statista, which is a private company, or something like that, instead of going to Census, or Bureau of Economic Analysis, and forget about the states and the counties, often, you know, in fact, it's, you know, it won't even bother trying to look there. [29:14](https://youtu.be/ogpbSR8KkwM?t=1754) That happens a lot, right? So, Google, LLMs, and so on are popularity contests. Well, this is where the factual work, kind of, Breitbart does, you know, with federal government and a handful of state governments, and always comes in, what you have to educate the system about is provenance, where to look for what, okay? And that's this notion of stat rank. If somebody asks a question about employment, well, you know, where should they look, you know? And if you ask your colleagues, or ask among your friends, like, hey, you know, somebody wants to know what's the employment of, you know, we're a California-based company, so I'll give many California examples, you know, but, you know, if you want to know what's the employment in California, well, guess what? There are many answers to the question, right? You know, you could ask Census, you could ask... VLS. You could ask the State of California's Economic Development Office, and there's plenty of other surveys and data sets that are out there that have some answer on what is the importance in California, right? So it's a complicated answer. Well, guess what? LLMs are totally confused about this, as you would expect. Heck, even experts can be confused about this. So if even humans get confused, you can imagine LLMs are super confused. So you need to have a ranking system of who do you trust, under what conditions, for what kind of question. You have to do this. [30:32](https://youtu.be/ogpbSR8KkwM?t=1832) There's no other way. And you might say, well, you know, if a user is interested in this kind of question, on this level of aggregation, maybe look here first. Like, you know, for zip code level information, maybe, you know, this agency is hopeless, but this other agency does have zip code level. Or, you know, for if you're looking at county level, well, maybe there are three possible, you know, sources. And maybe you don't want to rank them, but then you mention all three. All of these rules of behavior would be, by the way, codified in your MCP. Okay. And MCP is the way to communicate now with, you know, between agents, and, you know, and between systems. But there are ways for you, even in your website, to make them AI ready, and in your data tools, and open data directories, and so on, that many of you have, and many agencies have, at that state and local level, for many tricks one can use to make your data, you know, more, you know, more discoverable. So, those are the three things. You know, so, grouping fields by topic, like a tree of knowledge. Number two, the levels of aggregation that has to be synthesized and explained and codified. And third, the statistical ranking, or who do you trust for what type of question. And if a question cannot be answered by this agency, who should you go to next? So, some high-level advice here. Yeah, no, this is great, Jose, and this is exactly, anyway, I'm sure the audience appreciates to get those kind of concrete suggestions on how to go forward. Now, building upon that as well, Jose and Luke, is that when we developed our report on kind of the fourth wave of open data, we had a big discussion about AI readiness. And so, some were indicating is that it depends on ready for what. And so, eager to hear from your perspective, if, anyway, if you're kind of a local data holder, [32:36](https://youtu.be/ogpbSR8KkwM?t=1956) should you aspire to be ready for kind of the ultimate use case, or should you say, well, anyway, my data can already be used for certain kinds of applications, and how would you know what the applications might be that your data is ready for? And so, basically, the question is, is there different levels of readiness for different AI applications? Because especially, Jose, if we go for factual AI, I think we have a higher standard, probably, actually making that data ready. And so, eager to hear, and again, the discussion that we had was two years ago, so things have evolved. And so, eager to hear whether that still is relevant to think about, anyway, how is my data AI ready, but then for what purpose as well, and is there kind of a taxonomy that you use in your own work? And I don't know, Luke, whether you want to respond to that, or Jose, I'm sure you have a response to that as well. Luke, you want to do the honors and get us started? **Luke Keller** [33:38](https://youtu.be/ogpbSR8KkwM?t=2018) Yeah, sure. I love the question of for what, right? I think there's different tiers of readiness, and I haven't had the time to codify it, but the way I would think about it is fundamentally, if you're interested and just genuinely concerned about your data being represented, even without MCP or direct integration, if you just want to make sure that an LLM is able to find your data, you might want to think about that problem the way that a lot of folks have thought about SEO for websites. I know there's an emerging practice of AEO, like AI engine optimization, where there are certain things you might be able to do to improve how your data shows up in a particular LLM. It's not going to be deterministic. Again, it's going to be deeply probabilistic behavior, and it might be a drop in the ocean, right? But when LLMs typically, with web search enabled, interact with the user, they're going to try to find the most appropriate website to answer that user's question. And if that data is on that website, you have a good shot of your data being represented there. So at bare minimum, making sure your data is following good open data form, making sure it's publicly accessible, ideally through an API that's documented publicly, something to that effect, where an LLM knows where that data sits, how to query that data. The more documentation you can provide there, the better for an LLM. So what's good for a human user is probably going to be good for them too. But if you're more deeply concerned about influencing decisions in a way that where AI engine optimization and discoverability isn't enough, where you don't want to be a passive participant in how your data is being used, but you want to be an active participant in how your data is being used. Maybe you're concerned about specific use cases that you have in mind of data customers dependent on your data, or maybe you just want to build something for yourself, right? That's where I think things like MCP come into play. And I think the thing that we found that was so interesting about this project was before we worked on it, that January of 2025, I want to make sure I get my dates right, Jose. In January, the Department of Commerce [35:39](https://youtu.be/ogpbSR8KkwM?t=2139) published a report on gender of AI best practices for its bureaus. And we didn't really cover MCP that much. It was so new. And we covered things like good schemas and ensuring your data is accessible and all the things I just described, right? But what that report didn't cover was, you know, when you actually kind of are like boots on ground, integrating your data with an LLM, when you've built an MCP, it's not enough to just hook up your existing data to an LLM through MCP. You have to ask like, is your data token optimized? Meaning like, if you, you know, take the Census Bureau's API, right? Jose mentioned that, you know, LLMs don't like tables. Well, we love tables at the Census Bureau. And so we would just, you know, like when you asked for something about the American Community Survey, we would spit thousands of pieces of information at you, right? Like, oh yeah, this is, you want to hear everything about this particular subject. We're just going to give it all to you. And LLM doesn't like that because that's, that's GPUs, you know, like burning credits, burning time, energy, that's costs, right? That is not efficient at all. Like the way a human user maybe would want to parse those things, maybe somebody who's like, you know, just comfortable with the intricacies of Census Bureau data, like a GPU and an LLM are not, right? Like they don't like that. So like, you have to think about like how you optimize your data to deliver a very targeted response based on what, you know, the LLM is trying to figure out. So a great example of the MCP project that we had was working with our geographic data. [37:09](https://youtu.be/ogpbSR8KkwM?t=2229) So in the Census Bureau API, you could pull all the geographic data, you know, that we have, meaning like all the kind of like, not like the, I'm talking about the Census Data API, so not the, you know, the shapefiles and all that good stuff, but rather like what are the FIPS codes associated with what different geographies, right? You can pull all that information today, but you can't search it. The way you get it is by summary level. So if you're familiar with our data, it means like for places, for states, for, you know, regions, divisions, you can pull all that stuff, but you have to do it by the summary level. So if you wanted to find the FIPS code, which is essential for, you know, kind of refining the type of data that's returned from API, if you wanted the FIPS code for your hometown, you would have to hit the places endpoint in the API for that particular vintage, say 2023, and you'd have to comb through thousands of places just to get that one FIPS code. That's not great for an LLM. So what we did in the MCPs server was actually to stand up a parallel local database that's installed alongside the MCP. And what it did was it indexed all those geographies and with some simple Postgres search functionality, right? [38:16](https://youtu.be/ogpbSR8KkwM?t=2296) Where you're essentially able to say like the LLM wants to search for this name of this geography. Let's enable it. Let's create a search function across our geographies. You know, give it a ranked response based on just likeness to the original query, which is again, old school, really simple kind of like database search, right? Real problem. Return it back to the LLM, but then lean on the LLM to use its probabilistic kind of competencies, right? To then look at that list and say, the user asked this, you've given me this list. I think the place that they want is this one in this list. So if there's a great example would be like the city of Philadelphia. There's several Philadelphias actually that are recognized by the Census Bureau. There's like different town names of Philadelphia in the country. There's different summary levels, right? There's like the borough, the county, there's all this stuff. So it's like, which one do you choose? But if you give that information to an LLM, they're gonna be able to say, I think the user wants the city and this is, you know, the place, right? And this is why. And so it's able to make that choice. So it's one of those things where like, when you think about your data and ready for what, right? Like, I think what you really need to be thinking about is like, is your data optimized to be delivering like as targeted and kind of like brief of a response as possible to an LLM asking a question. And if it's not, even if you build an MCP server, which frankly doesn't take a ton of effort, that's going to be deeply problematic. Where we spent, I would say like 85% of our time was on engineering our data to work with an MCP server in an efficient way. [39:44](https://youtu.be/ogpbSR8KkwM?t=2384) The MCP stuff is super easy. Like you could go in cloud code today and spin one up. Every single federal agency could have one today and just spun it up. But would the data be optimized? No, because not only do you have to pull data from the APIs, you also have to like round out that data, add contextual data, contextual descriptions that's buried deep in handbooks and PDFs and websites and all this other stuff that's frankly just like not available programmatically in an API. So I think that's where like subject matter experts come in and being able to lean on those folks and understand like not only how you engineer the data that you have today available through your APIs or other data sources for use by LLMs, but then how do you round that out? How do you contextualize that to give them as much, to give LLMs as much of an understanding as possible so that going back to that original question that was asked about kind of the faithful representation of data, it's like when you provide that context, you're inherently making an LLM far more faithful to the intent of how that data should be used than if you just give just a targeted response and say hands off. So we learned a lot. Yeah, yeah. No, this is great, Luke. And **Stefaan Verhulst** [40:46](https://youtu.be/ogpbSR8KkwM?t=2446) again, that's why it's so great to have you on the session here today is to really understand what does it take and especially also I like the, anyway, that you bring kind of the importance of context into equation because of course, anyway, a lot of discussion now about are we still talking about prompt engineering or is it context engineering? And I think what you clearly have indicated is actually it's, anyway, it's kind of both to a large extent, but at least without the context, the prompt will, anyway, will not, will be inefficient in actually how data is being used and exactly will not be appreciated for sure. Jose, any reactions from you? And perhaps because I do want to go to the Q&As as well, perhaps Jose, in your answer, you can also perhaps integrate a little bit more about what opendata.org tries to do, because I think it's somewhat relevant probably as well, right? Over to you, Jose. [41:49](https://youtu.be/ogpbSR8KkwM?t=2509) Yeah, absolutely. So I'm sure many of you, and I saw a few in the chat saying that you've experimented with building an MCP server. As Luke said, that's not the hard part. The hard part is the data engineering, the data prep that goes into it. That's for work. And sometimes it can take months. Part of what we've done in our work with census, BLS, BEA, commerce, and all the major statistical agencies is develop a rubric that the federal government approved for how to develop an MCP server that meets these very tough, rigorous standards of factuality and reliability. You also want it to behave the same way over time. That's quite critical. Don't forget the LLMs have a slight stochastic nature to them, right? And kind of by design. That can be a good thing if you're trying to be creative and write an article or write a book, but that's not so good when you're talking about statistics, right? Where it's all about factuality and reliability and consistency, right? So there are many things that have to be done on the data engineering side for the MCP server to work well. You know the old adage, garbage in, garbage out? Same thing applies to MCPs, right? So it has to be prepped the right way. Good news is, you know, we've done this already. We've developed a rubric with the Fed agencies and so on and a handful of state agencies. We'd love to talk to you to kind of help you all on your journey as well. So that's definitely one area where I would put in some thought and effort is on the MCP strategy. And again, you know, we're here to help. A second thing is to help sort of your agency statistics rise to the top of search results and things of that sort. And this is an area where we work very closely on is to work closely with the tech firms for them to be aware of you. You know, it's often about awareness, all right? And we have found that to be quite critical. [43:51](https://youtu.be/ogpbSR8KkwM?t=2631) You know, Google, Anthropic, OpenAI, Microsoft, you know, we work with them every day. And part of our job is to inform them about all the various agencies, what they have, and who to trust for what, and so on. Everything that we're building, by the way, integrates into the Google Knowledge Graph, which is what underpins Google Search. And we're doing now also the same with Microsoft, through which we're doing the same with OpenAI. And only recently we started also doing the same with Anthropic through the Linux Foundation, the Sage Antic AI Foundation that Luca referred to before. So you can build an amazing MCP server, but if they don't know about it, it won't be used, right? So challenge number one is build it well, but the data engineering that feeds into it, which touches on those things I mentioned earlier, how the data is prepped, et cetera, et cetera. Second, awareness. And then third, have what we call an AI explorer on your page. LLMs and search engines now, they love these, and they'll detect them and they'll use them. [44:52](https://youtu.be/ogpbSR8KkwM?t=2692) Again, we built one for the Fed agencies. You know, we can talk more about how that works. But long story short, an AI explorer is It can serve two functions. One is it can give templated answers to statistical questions, and or it can answer questions in conversational style. And we have found that building that can be really, really useful for the search engines and LLMs to at least become aware of, you know, what you have. And of course, for visitors to your site and such. Remember the following. The majority of visitors to your site are not humans today, okay? They are agents. They are AI agents acting on behalf of the scrapers, acting on behalf of the search engines, acting on behalf of news bots and various other, you know, things. We estimate that 60 to 80 percent of traffic is agentic today, and it will probably be 99 percent before the end of the year, which means your sites have to be, you know, AI-friendly, which means having, you know, these pieces of the puzzle. [45:55](https://youtu.be/ogpbSR8KkwM?t=2755) So Stefan asked about open data. So the agencies publish statistical information. Open data is about entity-level information. So down to the specific company, specific person, specific building, specific address. So, you know, the agencies produce through surveys and administrative records being collected and anonymized and so on, amazing data sets about the economy, about healthcare, about crime, about climates, et cetera, et cetera, et cetera. Those are wonderful statistics. But there's also some gaps of specific data about specific companies, specific people, et cetera, that the world doesn't need. So opendata.org, the initiative that I slash Brad Corey, you know, are launching, again, in partnership with the Linux Foundation and our partners at the AI Alliance and so on, is around building open-source data sets. It's like a global directory of every company, every working adult, every place of business on the planet, including the relationships between them. So it's like a monster entity graph. And then this graph, we are kind of donating to the world. You know, we've built it, millions and millions of dollars of investment. We've built this over the years, and we're now donating it to the world to help make these graphs more, and search engines and LLMs, you know, more factual. The agencies handle statistics, we handle entities. That's kind of the vision behind it. Yeah. [47:19](https://youtu.be/ogpbSR8KkwM?t=2839) Great. Thanks so much, Jose. And I do recommend everyone to check it out at opendata.org, indeed. And it goes back also, Jose, to what you mentioned as well, which is kind of, anyway, is that kind of graph the data is becoming more and more important. And I think what you've done there with the entity graphing is, I think, essential. So let's go to a few of the questions that have been posed in the Q&A. And so for those who still want to pose, feel free to do so as well. I think the first question was more about more examples of interoperable data ecosystems, and some of the lessons that we can learn from them to improve data set usability. And of course, you know, this is kind of goes back to kind of data engineering, perhaps, Jose, what you refer to kind of how do you make and how do you prepare the data? And of course, interoperability is always kind of the question that comes into play. And so that's one question. [48:28](https://youtu.be/ogpbSR8KkwM?t=2908) And Josua had another question with regard to what strategies would you suggest for organisations looking to streamline their data governance processes, while also prioritising innovation in AI? And so building upon here, and so the first question, Jose, and Luke, you can decide who you want to answer that one. But I think linked with that question, the second question is, is a question I have for both of you as well, is that do we need kind of new functions, when it comes to data governance functions, that really understand how do you make data AI ready, right? And so typically, we had kind of data stewardship, right, which was kind of focused on kind of the integrity of the data was really trying to understand, anyway, how do we provide access to the data? But do we need to upgrade or redefine data stewardship at the time of AI? And that's kind of a question to both of you is that do we have the functions in today's org chart of data governance, that actually enables us to leverage AI, both for data for AI or AI for data as well. So Luke and Jose, you can choose which of those, you can answer them both. [49:48](https://youtu.be/ogpbSR8KkwM?t=2988) Or you can choose which one you want to start with. Jose, I guess I can, yeah, I can kick us off and kind of, you know, give a brief answers on both of them. On the first one, you know, I touched upon this a little bit earlier. But I think it's perhaps worth reemphasizing businesses, if there's one thing you all should remember, is don't think in tables, think in terms of trees of information. That is absolutely critical. It's a new way of thinking. And, you know, I'm an economist by training, my PhD is in economics. I was a professor for many years, I advise many PhD students, I've won many prizes, published a bunch of papers. And it was always about tables, you know, and leveraging those tables and so on. This is a new way of thinking. We have to organize the information by topic, subtopic, sub-subtopic and how they relate to one another. It's a different way of structuring the data really radically differently from how we've done things, you know, in the past. The good news is there are tools for doing that. You know, we've developed some, there are some others out there. But if there's one thing I would think about is the tree of knowledge of your information. On the governance question, kind of a second piece. Yeah, absolutely. We've seen this come up in the AI industry as a whole. [51:00](https://youtu.be/ogpbSR8KkwM?t=3060) You know, I'm very involved in the AI industry and speaking with the leaders of the model builders and all this and such. And what we have found is that it's not enough that the data itself be reliable and accurate and so on. And of course, statisticians, economists, social scientists put a lot of effort to make sure the statistics that are produced meet those, you know, high quality standards that the agencies, fed, state, local are known for. The flip side is LLMs being stochastic by nature and often misunderstanding things will often say things or misrepresent things like these. We didn't understand or we got confused. So now it's like you have this interpreter. Don't forget, and maybe some of you are not aware of this, but LLMs were born out of technology of translation. Okay. You know, the transformer of GPT, general purpose transformer, is about, was all born out of research at Google, which was around translating text. In many ways, what LLMs are really, really good at is basically understanding and translating a text and we try to be creative when we do so. So what we have found to be effective from a governance standpoint is to set up a series of caveats and rules of behavior that the MCP and therefore the LLM has to follow when giving responses. For example, you know, it could be in the form of footnotes. It could be in the form of, you know, recommendations, things of that sort. You know, to always warn the user and always inform the user of, you know, hey, you know, there's a risk of, you know, things were misinterpreted. You know, if you're not certain, always check. You know, the LLM produced this response. There is some uncertainty. You might not get the same answer all the time, et cetera, et cetera. All these little caveats that for legal purposes should be mentioned, but also for practical data analytical purposes should be mentioned. And a lot of this can be done in the prompting and the design of the MCP, you know, server for sure. But then the last thing [53:04](https://youtu.be/ogpbSR8KkwM?t=3184) I mentioned to kind of help with governance is it's one thing to set up the standards, but then you need human testers, human evaluators. And that's one thing that I've seen agencies sometimes forget. You can't just put out there, you know, these tools. You need to have humans, experts on the staff who test them out occasionally on a random basis to see what kind of responses are being produced. And in that regard, one thing I would highly recommend to build into any AI system is a feedback loop, a feedback system. You need to allow the user to give a thumbs up or thumbs down and, or provide some kind of feedback of content of whether they liked the answer or didn't, what they liked, what they didn't like, and so on. This is what we call human reinforcement learning. You know, otherwise, LLM's making a haywire. And if a user cannot tell you, the agency that, you know, you want haywire somewhere, how are you going to know? You'll just have a disappointed user. So setting up rules and guidelines, that's critical. But then you do want this feedback loop and some human testing. We're still no substitute for the human mind. [54:05](https://youtu.be/ogpbSR8KkwM?t=3245) Great, great, Jose. And that generated a lot of feedback itself here on the session, Jose, by just mentioning that. So thanks for that. Luc, any response for you? And again, there were many other questions. I don't know whether you checked the questions as well. There were some of the other questions as well that were kind of specifically focused on US Census. And I don't know whether you want to address in the kind of the four minutes that we have left, address some of those. Like, for instance, there are some with regard to margins of error, which is kind of, anyway, if you have a large land area, small population, so you have a census block, census data, and how does that play out if you can provide them access to that through LLMs, right? **Luke Keller** [54:51](https://youtu.be/ogpbSR8KkwM?t=3291) Yeah, I'll start on the last question for like 10 seconds. I would say on interoperability, I would love to see local and state governments kind of combining forces and kind of unifying how they approach problems. There are common data platforms or the common MCPs, right? Again, think about like biggest swings you could take. That would be great. On data governance, I would just love to see folks, frankly, be more action-oriented. We had data sets sitting there where descriptions were, like the actual description of a data set was, this is a description. What is that, right? Like in 10 minutes, five, no, not in 10, in two minutes, you can run that through with the appropriate prompt, take Claude, say, you know, use this handbook, use this website, help me fill out my descriptions, get an entire kind of description for thousands of records in your data, right? Like for us, it was like descriptions of our surveys, right? Run those with your SMEs, spend a few weeks on it, and you'll never have to touch it again. There are such low hanging fruit in terms of like data quality that like everyone on this call today with a public LLM can use and not even hit your token limit. Then what comes after that is like the people in the governance. So like, how do you move that through and how do you get that integrated into the actual like data portals and your APIs? Like, I know that's a challenge, but like the tools exist today to improve data quality. So like, why aren't we doing that? That's one thing I'd push. For margin of errors with the census, you know, API, yeah, it's really tough. I think like one thing we didn't get to that we wanted to kind of like keep building on while at the Census Bureau was to like force some of those, you know, in terms of like how we're formatting responses. So like, you know, if we have margin of error data for certain, you know, for certain data, right? Like forcing that in the response or in the API, like, you know, there are ways to like ignore that information, not return that information. [56:32](https://youtu.be/ogpbSR8KkwM?t=3392) There are ways as a data provider to kind of dictate what comes with that data. And I think I would kind of like elevate that topic thematically to say like, if there are opportunities to further contextualize your data in ways that, you know, in your APIs today, like you can get around, like you might want to consider building MCP tools that actually force that functionality. And that way your data is being appropriately contextualized. So there are definitely things that you can do there. **Stefaan Verhulst** [56:56](https://youtu.be/ogpbSR8KkwM?t=3416) Great. Wonderful. Well, I'm sensitive. We have like two minutes left. So, and I will have to close this here because otherwise no one will ever show up if we go over time as well, because people anyway, our scarcity of time is so. But one quick question to you, Luke, having been a chief innovation officer, what was one thing you wished you could have done? And Jose, what is one thing what is one thing you wish you would be able to do in addition to all your great projects that you have or things that you're doing that says, look, this is kind of what I'm really excited **Luke Keller** [57:34](https://youtu.be/ogpbSR8KkwM?t=3454) about? Very short. Over to you, Luke. Yeah. So one thing I wish I could have stuck around for, but it's happening right now, thanks to the wonderful staff that are still at the Census Bureau, is they are now actively working with companies, data portal providers, et cetera, to pilot the use of the MCP and get feedback from engineering teams that are working with thousands of customers. That's the dream, right? We released something, we weren't sure how it was going to land. We're like, let's see what happens. Let's make some waves. We made some waves. And now we had a lot of folks reach out, local state, other federal agencies. I know the CDC was using it at one point in their own pipeline, seeing use and building that feedback loop, when you have a successful product launch, it's amazing. So these folks are now putting the MCP into the hands of folks, seeing how they use it, collecting their feedback, building out additional requirements and laying out a path for what the roadmap's going to look like in the future. So it's not something I got a chance to see, but it's something I'm so glad continues. **Stefaan Verhulst** [58:36](https://youtu.be/ogpbSR8KkwM?t=3516) Wonderful. Great. Jose, you have the last word. Sure. Well, thank you all again for joining. Please feel free to reach out to me if any of you want to learn more about all this work we're doing. I think I put my email in the chat. Jose at brightquery.com or JPLEHN at brightquery.com, they both go to me. No worries. We'd love for you all to get involved with this projects that we have going on with the federal agencies and various state agencies and trying to pull everything together into one kind of master AI system that understands all statistical information produced by the United States collectively. So I hope you can all join us on this journey. And that is kind of the answer to the question, what is it that you would like to see happening? So that's great. Thank you so much, Jose. Thanks so much, Luke. And thanks everyone for joining. I think we already added some information about the next sessions as well. And indeed, some already have asked, can we do this for four hours? I'm not sure that that's feasible for Jose and Luke, but I learned a lot in one hour. So thanks so much. And thanks everyone for joining. [59:40](https://youtu.be/ogpbSR8KkwM?t=3580) Take care. Bye-bye. Thank you. Thank you, everybody. * * * Original page: https://innovate-us.org/building-ai-ready-open-data · Licence: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/), as stated on innovate-us.org