---
id: "workshop:ai-101-for-human-services-leaders"
type: "InnovateUS workshop"
title: "AI 101 for Human Services Leaders"
url: "https://innovate-us.org/ai-101-for-human-services-leaders"
date: "2026-03-23"
series: "AI and Human Services: Innovating to Serve People Better"
instructors: ["Dave Cole"]
moderators: ["Rob Asaro Angelo"]
recording: "https://www.youtube.com/watch?v=MEZLb5EOZG4"
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"
---

# AI 101 for Human Services Leaders

InnovateUS workshop · 23 March 2026

Human services leaders are being asked to make decisions about AI, often without clear, shared definitions of what these systems are, how they work, or where they are already shaping service delivery. This session provides a practical, plain-language introduction to AI in human services, grounded in real operational contexts rather than hype.
Participants will explore how AI is currently used in areas such as eligibility determination, case management, fraud detection, and client communications. The session focuses on building informed leadership judgment: understanding what AI can and cannot do, where risks tend to emerge, and how leaders can ask better questions when evaluating tools, proposals, or pilots.
This workshop is designed for human services executives, program leaders, and managers who need a clear foundation to engage responsibly with AI initiatives.

## Learning goals

By the end of this workshop, participants will be able to:
•    Distinguish between common types of AI systems and how they show up in human services operations
•    Identify current and emerging AI use cases across eligibility, case management, and client communication
•    Recognize common misconceptions and risks associated with AI adoption in high-stakes public programs
•    Ask informed, practical questions when evaluating AI proposals or vendor claims

## Speakers

- **Dave Cole**, Chief Innovation Officer, State of New Jersey.

## Moderators

- **Rob Asaro Angelo**, Senior Fellow, Burnes Center for Social Change & Professor, Northeastern University

Part of the series [AI and Human Services: Innovating to Serve People Better](https://innovate-us.org/workshop-series/ai-and-human-services).

Recording: https://www.youtube.com/watch?v=MEZLb5EOZG4

## 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.

**Cassandra Madison**

[0:00](https://youtu.be/MEZLb5EOZG4?t=0) It's such a pleasure to be here with you all today. Super excited to kick off this series with our great partners at Innovate US. This is the start of a whole bunch of conversations we're going to have about AI's applications in human services. So as Agata mentioned, my name is Cass Madison. I'm Executive Director of the Center for Civic Futures. We are a nonpartisan nonprofit that's focused on helping state, tribal, and territorial government leaders make good decisions on emerging tech. We do a whole bunch of things, but I think of them typically in three pillars. What pillar one is convening and community, and this is the idea that for all of you out there who are leading in your states trying to make decisions on AI's applications, it's really important that you have a safe space to talk to each other about what's working and what's not working. We run a couple big communities of practice to create safe spaces for this dialogue. One is for Chief AI Officers, the other is for Health and Human Services Secretaries, and are talking to a lot of the senior leaders on the teams of folks who are in the room today.

[1:08](https://youtu.be/MEZLb5EOZG4?t=68) So that's thing one. Thing two is all about readiness and capacity building, and this series that we're rolling out today is a big piece of this. But this is the idea that you all in your day-to-day work are needing to make decisions, and it's tough for states to have all the capacity that they need to make good decisions on emerging tech at this moment because things are moving so fast and the pressures inside are so strong. So we spend a lot of time on things like education and framework development and on-the-ground technical assistance with that North Star being that you could walk into a room with a vendor who is pitching you something and know what questions to ask, what good looks like, what to be concerned about, and make good decisions on behalf of your beneficiaries. And then the third pillar that we focus on is our living lab, and this is a strongly held belief that particularly with AI, the ability to govern AI, make good decisions on it, build it well, requires really understanding how it works. And so this idea that states, territories, and tribes need safe spaces for experimentation where the goal of that experimentation is primarily learning. And so we work hard to create those spaces because sometimes they're difficult to create in government. And so for us, some of that looks like direct funding of pilots, of which we've done some in New Jersey, and Dave Cole and friends are a great partner. Sometimes we're on the ground actually bringing states together and experimenting with use cases, but it's all applied work. And then our goal is to create connectivity between those three pillars.

[2:38](https://youtu.be/MEZLb5EOZG4?t=158) My background is state government service. So I've been in the seat as a deputy commissioner, as a director of operations at our state Medicaid agency. And early in my career, actually started as a policy person and had a front row seat in Vermont to a couple big, really troubled tech rollouts. And that has really influenced the second half of my career, which is, hey, let's try to provide government with the support to roll these technologies out. Better, faster, stronger in service of the beneficiaries that we support. So I think this series that we're rolling out today is really this recognition that you all are at this really pivotal moment where not unlike we went through with our healthcare.gov disaster in Vermont a few years, 10 years ago now, but policies are changing. There's really tight implementation timelines and layered onto all of that is AI is moving really fast. And that provides these technologies offer a tremendous opportunity and a bunch of risk and states, this recognition that states, especially in health and human services, where you're serving some of the most vulnerable populations in the country really need support in navigating that world effectively. And so that's what this series is really designed to do. So in partnership with Innovate US over the next coming months here, we're going to be exploring things like what AI is and what it isn't, which will be the focus of today. We're going to explore what responsible AI looks like in practice, how data underlies so many AI use cases, including identity management.

[4:15](https://youtu.be/MEZLb5EOZG4?t=255) We'll focus on how to design AI systems to improve program delivery. And finally, how to engage frontline staff and community being served in those programs before these technologies get rolled out. And so our goal through all of this is that you become better equipped to ask the right questions and be more confident in the decisions you're making. And we're also hoping to learn from all of you because as we continue to work with Innovate US and other partners in this space, we want to make sure that we're developing the program supports and content that are best going to support you as you embark on this journey. Really appreciate all that you're doing to serve beneficiaries out there and support each other through these changing times and excited to dive in today. So I'm going to hand it over now to our moderator and our guest speaker, Rob Asaro-Angelo, who's a senior fellow and professor of practice at the Byrne Center for Social Change and former commissioner of labor at the Department of New Jersey, sorry, State of New Jersey. And over to Dave Cole, who's New Jersey State Chief Innovation Officer. Thanks, Cass, for all that. Appreciate

**Rob Asaro Angelo**

[5:22](https://youtu.be/MEZLb5EOZG4?t=322) that. I'm so excited to see so many fellow public servants from around the country here today, from let's see, from LA to New Hampshire, got the whole country covered, Virginia, and all different levels of government, which really makes me thrilled. Nothing makes me happier than to be around my fellow public servants around the country. When I was asked to moderate this series, I was told emphatically that the first, second, and third rules of Innovate US series was that they have to end on time. So I will spare you all my hysterical witty banter that I had trying to get right into the program. You already know about today's topic, which is why you're here, but I wanted to go over some of the goals for this session for us to learn together, to distinguish between common types of AI systems, how they show up in human services operations, identify current and emerging AI use cases across eligibility, case management, and client communication, recognize common misconceptions and risks associated with AI adoption in high stakes public programs, and ask informed practical questions when evaluating AI proposals or vendor claims. So I mentioned the Q&A before. Please get in there and ask questions now or as you go along. We have a huge gamut today of folks as far as their experience with technology at all, let alone AI, or where they are in government. So there's no bad questions. The more we ask, the better we learn all together. But right now I'm happy to introduce my friend and former colleague, Dave Cole, who is the chief innovation officer for the state of New Jersey. And if you read his bio online, it talks about coming to the state in 2020 to help with COVID response.

[6:50](https://youtu.be/MEZLb5EOZG4?t=410) And I am so lucky that a big part of that was his work with the Department of Labor. I don't think I need to tell anybody here about what departments of labor who administer unemployment benefits went through during COVID. We didn't have a lot of friends out there, but Dave Cole was a fantastic friend to our department, to our 3,000 employees, and to the 1.6 million workers in New Jersey who received unemployment benefits during COVID and helped us navigate through that time and try new things in the most stressful time in our government's history with great success. So I'll turn it over now to Dave Cole for today's teacher. Thank you.

**Dave Cole**

[7:26](https://youtu.be/MEZLb5EOZG4?t=446) Go, Dave. All right. Thank you, Rob. Appreciate it, everyone. Thanks for joining. And thank you, Rob, also for your leadership over the last eight years to create the space for us to do things differently to try to deliver better services. It's been wonderful to get to work with you and your team over that time. For today's conversation, I think our opportunity here is really just to ground truth the conversation about AI in some terms, some use cases that we think are accessible but practical for public servants, and then to open up a discussion about how you all in your various work are seeing the use of generative AI, any questions you have around it, and how we can make sure that we're deploying this new emerging technology in an effective but responsible way as well. So to start with a couple terms here, I want to make sure that we're talking about AI in the same way. When we on our team at the New Jersey Innovation Authority talk about artificial intelligence, as I'm sure many of you as well, we're principally talking about generative artificial intelligence at this point. So these are large language models, LLMs, software that is designed to provide an interface for people, to provide context, to provide prompts, and what we think of as asking questions to the model. And then based off of all the knowledge that that model has been trained on, try to get distilled answers. The commercial examples, ChatGPT, Claude, Gemini, Copilot, are some of the most common use cases out there. But of course, we can take that approach and build custom applications that solve more discrete businesses as well.

[9:00](https://youtu.be/MEZLb5EOZG4?t=540) We also work with predictive modeling or predictive artificial intelligence. For us, this means using patterns and data to try to find predictions of what might happen next, or classification of information. And we can use this for trying to find, if we look across the data set, what the probability is of a match of a name, for instance, that are in two different data sets. Machine learning, I think of as a bit more of a general field under artificial intelligence beyond generative artificial intelligence, really how we're training, building algorithms to train computers to approximate how humans learn, how humans think. But it's a bit more of a broad topic for the use cases that we're going to discuss today. And then of course, more recent evolution of generative AI, of agentic AI. Some examples of this would be scraping the web or using other tools to generate codes, to interface with databases. But largely, this is having prompts or a series of prompts that you might use for a generative AI system be chained together and then have agents have access to tools to be able to interact with other services, whether that's through the terminal on a computer to access the computer's information through a web browser, but to be able to extend beyond just the AI context and take action on an additional set of tools. For our purposes, of course, we're going to talk about how this all relates to government. So there are two key topics that we think about on our team when we're working in the state of New Jersey. The first is using AI to make governing easier, more efficient, more effective.

[10:34](https://youtu.be/MEZLb5EOZG4?t=634) These are applications that we might design and build to help staff manage the workload that they have to simplify repetitive work, to automate things that can be automated, while making sure that we have a careful line about the things that really require human judgment, human decisions, and that that is always preserved. And we also think about how AI can help residents get access to better services and the information they need. So these might be resident-facing services. Even before an agent gets an application for a program, how do residents get access to the information they need so they know that they're providing the right information in their applications? There are a bunch more use cases and things the government needs to think about from the economic development perspective, how states can attract artificial intelligence companies and businesses to locate within the state, how they can train their workforce to make sure that across from early education through higher education, folks have access to understand the technology to learn about how to use it to better themselves, their work, and the outcomes that they generate, as well as then how do we create the right regulatory framework for ensuring public safety, consumer safety, and health, as well as addressing some of the potential externalities, particularly around environmental consumption, that can be caused by the training and the use of artificial intelligence as well. Our approach is to start with our values. So we have values on our team at the Innovation Authority around all of our work being human-centered, making sure that the people that we serve are involved in the design of the systems that we're building that get the opportunity to provide feedback and help us improve those systems and to be iterative in the work that we do as well to make sure that we're delivering value very early to residents and to businesses and the folks that we work with and using that feedback that they provide to then make improvements. This makes sure that we stay very rooted in that first principle about being human-centered. So we want to bring this to our approach to using artificial intelligence, generative artificial intelligence tools as well. Across the state, we have a series of guidelines that are available for how anybody in the state can responsibly begin to leverage artificial

[12:39](https://youtu.be/MEZLb5EOZG4?t=759) intelligence. Those are rooted in the values of empowerment, inclusion, respect, transparency, and accountability, responsible innovation. To say a bit more about that, we want to make sure that AI is supporting our workforce to deliver effectively, safely, and equitably, that we're using AI to uplift our communities. We are public stewards, so we want to make sure that these tools are used to reflect the values that we have in delivering public benefits, and that transparency and accountability means disclosing the use of AI so people understand how we're using AI, where we're using AI, have the right to review, and have consent over their data in the use of some of these tools as well. And finally, that we are embracing experimentation with the right guardrails, particularly around privacy and understanding where the potential for bias or what ethical constraints we need to be thinking of as we're using artificial intelligence tools as well. The way that we bring these values to life has to do with a series of responsible guidance that we provide. In New Jersey, we have what we call an acceptable use policy. It's a document that provides guidance to the state workforce with specific values, with specific use cases, and a place to get help to because we recognize that with an emerging technology, this is an interactive conversation. The first pillar of this is to start with training.

[13:56](https://youtu.be/MEZLb5EOZG4?t=836) So we've made available through a partnership with Innovate US video-based training for all of our state's workforce so that folks can get hands-on with generative artificial intelligence tools and see public service use cases. How might these tools help improve our delivery of services and our work within a public service context? We also then make sure that we have the right privacy and security constraints in place. So when we're rolling out artificial intelligence tools, we want to do so in a way that fits within the state's IT infrastructure to make sure that things like data and privacy contracts and regulations that need to be met are already in place before the staff has access to these tools. By providing tools proactively, we hope to limit the cases where people might use tools on their personal time or their personal accounts and channel these through state resources that we have a higher degree of trust in and can make sure that we have done the reviews to ensure that they're secure. We also want to make sure that human reviews are explicitly defined in the use of generative artificial intelligence tools. It can seem very easy to set up a tool and let it start making decisions, but there are a variety.

[15:00](https://youtu.be/MEZLb5EOZG4?t=900) of reasons why that can actually be a dangerous prospect. And so we make sure that folks, when they're using AI, are trained and understand that they, as staff, are individually responsible for the decisions of those tools as if it was their own work. It doesn't matter that you're using generative artificial intelligence or another mechanism. We individually have to take responsibility for the outcomes and the output of the work. And so that means making sure that there are reviews for anything that's going to be a deterministic decision-making, that we're reviewing the information that goes into these systems as much as possible, and that where anything is going to be automated, we're disclosing the use of those tools as well. So disclosure helps build trust, as well as make sure that people have some autonomy and control over the technology that's being used in the government context. And when it comes to anything that's going to be approaching decision-making, we have a whole separate review process that we've stood up using, our existing system architecture review process and IT mechanism that we have across the state for any new IT system, we try to fold all this into that so that the staff knows that the training is a part of, and the guidelines are a part of the same mechanisms that we use, not something that is additional to have to worry about.

[16:10](https://youtu.be/MEZLb5EOZG4?t=970) But again, we really focus on these two cases where a human review is essential and we can disclose the use case transparently. So what that looks like in practice, we have the New Jersey AI Assistant. It's an internal chatbot application that state employees can access. And there's a guideline that's built into that describing what our definition of responsible AI use is and providing to links for more information. So we're defining some terms about hallucinations, about how there could be inaccuracies or potential biases in the responses from artificial intelligence tools. If you wanna learn more about that, there's a guide that you can read to follow and understand about how these things are embedded in the application and how to make sure that you're controlling and testing your outputs to avoid some of the more difficult parts of it while then getting the value that the systems can provide. And as I mentioned, we also have the video-based responsible use for generative AI training that's available to all state workforce. And so there are links for people to access those trainings as well.

[17:11](https://youtu.be/MEZLb5EOZG4?t=1031) We, as part of the continual feedback process, when we build the AI Assistant, also wanna be human-centered in our own approach. And so we have surveyed the state public workforce about how they're using it. Some of the top reasons or the benefits that folks address when they respond to the survey is that the application can help them with saving time, with efficiency of repetitive tasks. Most commonly, that's in writing and editing. So if you're thinking about creating a new document and you need some brainstorming guidance, then AI can help play a role in that with the right context and prompting. And then editing content to perhaps edit for plain language or make sure that things are explained in consistent and clear language, AI can be very helpful with that. Also, we're seeing a lot of use around research and helping define research questions for checking accuracy of other information. So using some prompts to check the outputs of other prompts as a mechanism or an approach that we've seen as well. And then again, brainstorming, ideation, trying to help figure out how we can bring some elements of creative work into some of the tools that we're using.

[18:18](https://youtu.be/MEZLb5EOZG4?t=1098) We also asked for what are the concerns that people have with using these tools? We wanna make sure that we are addressing them, but also it helps us understand that some of the things that we're raising in the trainings are resonating and that folks are aware. So the top concern that we're seeing is around the potential accuracy and risk for hallucinations. This is something that we wanna make sure we address, but it's also a positive signal that folks who are using the AI tools that we're providing here are aware that these are real constraints that need to be managed. It's making sure that that message is getting through. Folks are also concerned about data privacy, storage, and retrieval, making sure that they understand any limitations about when and when not to use the technology that environmental and ethical concerns are addressed as best as possible, or at least disclosed where folks have questions about the use of this technology. And again, following in our values, we're seeing the same thing resonated back that folks are very interested in the transparency and attribution guidelines that they would need to meet when they use artificial intelligence tools.

[19:21](https://youtu.be/MEZLb5EOZG4?t=1161) Let's talk about some guidance here that I think we would offer as more general purpose, things that we're using on our teams that when we do training, that might be useful across the folks that are gathered here as well. So generative artificial intelligence tools are computers, they're software. I think it's very convenient for us to abstract them into the language of humans. They are not humans. And it's just very important to think about how we interface with them as machines, knowing that when we program machines deterministically, you know, using more traditional methods, we expect predictable output because it follows more common principles like math, right? There are certain single ways to answer questions. But when we move into artificial intelligence, particularly LLM-based artificial intelligence, we're not necessarily in the realm of just math anymore, where there's a predictable agreed set of rules that will have a consistent answer, but we're now dealing with probabilities. What is the likelihood that the response that follows should match the input that was given? So if you provide some context to the model, the model is going to try to come up with what it estimates to be the most accurate thing that would follow from what you've provided.

[20:29](https://youtu.be/MEZLb5EOZG4?t=1229) So it's predictive. It can seem highly accurate, but it's important to understand that it's not thinking in a traditional sense. It's just trying to guess through the training and information that it has and the checks that are programmed in what the most likely next outcome is. Use AI for the right use case. Very good for things like content and editing, content editing and summarization. But as I just mentioned, it's not necessarily a statistical analysis tool in the way that some more traditional machine learning algorithms may be better suited. Make sure to check outputs for mistakes. When you get outputs, there's a series of tactics and things that you could do to make sure that you're doing the work. But if you think about the work as being contributed by a colleague or some other system and you go through and make sure that you've really fact-checked, you don't wanna skip that step. That's an important part of using these tools as well. It can be easier to break down a complex task. So you might be thinking about a very large project that you want to undertake. And if you're using a chatbot LLM to try to get a result or response from that, it's often helpful to think about how can I break this task down so that it's manageable and ask a series of questions as opposed to one large task.

[21:41](https://youtu.be/MEZLb5EOZG4?t=1301) There are more advanced frameworks and more advanced services that use the agentic kind of interfacing that we've discussed to be able to take a complex task and then use a model to break down the task for you. But always better where you understand how to break down the task best to do that work oneself so that we can make sure that each of the individual prompts are thoughtful and that we're getting the best response from the model as possible. Very often, there's a huge trial and error phase to using these tools. And so you might, over time, develop tactics that work well for you, but there's a lot of testing and iteration that's normal as part of using LLMs generally. And you want to take the time to really iterate and refine your prompts and check them from time to time too to make sure that they're still accurate as the models might be updated with new information and new intelligence. And finally, use the availability of different models here to try to compare them and understand which model is best at which tasks. They're all built and programmed a little bit differently even if they follow the same general approach.

[22:45](https://youtu.be/MEZLb5EOZG4?t=1365) And so you might find that certain models are better for the kinds of tasks that you're doing and other models might be better at checking the work of that output. So you can kind of mix and match across different services or different model levels to optimize for the best outcome, the best output that you're getting, as well as cost where for some inferences or some prompts that you're doing, it's easier to use models that are a little lower cost. So when building generative AI, we look for the following in determining good use cases. First, again, establish your values and principles about how you do the work, what's important to you. And then from that, build out your AI plan. What are you going to use AI for consistent with those values and principles? Look for opportunities to save on things that are redundant or repetitive. It's very often the case when you talk to staff about what their day looks like that you can see, observe what they're doing. One might see, or even the staff might identify themselves areas where they have to manage multiple screens, go through physical or scan documents to pull out information.

[23:46](https://youtu.be/MEZLb5EOZG4?t=1426) And as you do this work more, you'll probably start to see patterns that are redundant. And those are areas where you might want to try some optimization with generative artificial intelligence to see maybe not to abstract the decision-making away, but can you get all the information synthesized from multiple documents into one place, into one form to make it easier for the agent then, or the analyst to review the information from that model against the source document, confirm it's accurate, and then help make the decision. So those kinds of things that are already happening through repetitive work can be prime candidates for automation or improvements there. And those use cases also, because of the human review, can have a lower risk for inaccuracies. So start with the less risky use cases and work your way up as you understand the tools. But of course, because of the public trust, because of the sector we're working in, we never want to sacrifice quality. We always want to make sure we understand exactly when a decision point is being made, when somebody is being included or excluded from a certain decision, so that humans reserve the autonomy over making that decision.

[24:46](https://youtu.be/MEZLb5EOZG4?t=1486) Thought it'd be worth sharing a couple examples here, and I'll pull from our team's work with Rob's team previously as well, to just give you some ideas on how this technology has been helpful. So a few years back, actually, after the first version of ChatGPT launched publicly, we were collectively with the Department of Labor in New Jersey and our innovation team working on improving the emails that go out from the unemployment insurance system. And those emails needed to make sure that people took a certain action, maybe provide more documentation or answer a question that somebody reviewing the claim had. And the emails traditionally had a focus on compliance and policy. They were written in a way that was very clear what the risks were if you didn't answer in the wrong way. But what we wanted to do is try to write them in a way that was focused on actions that the user needed to take and make sure that they are very clear plain language so that people could really understand. Our theory here was if we helped people understand what exactly they needed to do, they were much more likely to comply and to do so sooner.

[25:47](https://youtu.be/MEZLb5EOZG4?t=1547) And so doing this, we took the language from the existing emails with a series of prompts. First, we did this manually using some of the capacity and experience that our designers on our team had to do plain language rewriting. But then we tested taking that language and putting it into a chat GPT system and asking it to simplify or to explain to me what I need to do if I were to receive an email like this. That still then goes back to the policy experts. There's a review to make sure that nothing that the model generated is misleading or inaccurate. But overall, that increased response time by about 35%, which means agents are getting information faster from claimants. The claimants are able to get ultimately their benefits paid sooner because the issues are being addressed. Just to give you a visual of what that looks like here as well. Again, we have pasted in the language from the previous email. If your weekly benefit rate is less than the maximum weekly benefit rate, and you have dependents, you might qualify for the dependency benefit, a different benefit that stacks on top of your UI benefits here.

[26:48](https://youtu.be/MEZLb5EOZG4?t=1608) So rewrite this so that I can understand it. And you can see a plain language right there, which defines benefits a little bit more and helps lay out then the next steps of why somebody would wanna do this, bringing up that 7% additional benefit to the top. So this is just one example here, very early on experiment, where we fit this into our existing workflow through the human capacity in our team. We saw that this approach was already yielding improvements. And so then we can use the tool to help us scale it. So now we can go beyond what we might be able to do with the individual staff time that we have and be able to address dozens more email templates with a human review rather than doing the full rewrites from scratch. Another example is that the Department of Labor wanted to relaunch the unemployment insurance application in a more modern language, a more plain language application, as well as make sure that it was available in full Spanish translations as well. And so working with USDR and in-house talents across the state, the Department of Labor was able to put together a series of updates, including a Spanish language glossary that could be used to inform the translations.

[27:54](https://youtu.be/MEZLb5EOZG4?t=1674) So this helps us make sure we're reaching folks equitably and that there are less errors going into the systems because people understand the questions that are being asked. And if you have less errors than information being provided, it's easier to mark those claims up and get people paid faster too. So this is another example where using the AI to help scale our translations helped us complete the task faster. And then there's still a human review checking for the nuance of the translation and making sure that any individual differences in dialect could be addressed on top of the machine translation layer as well. A whole bunch of other use cases have emerged over time here just to give you a couple examples that we can unpack more in our discussion coming up. We help agencies improve their call centers. And a lot of that is having to make sure that the menus that people hear in call centers are also in plain language and responsive to the questions and the challenges that people are facing when they call in. And so we can actually use generative artificial intelligence to take scripts and to send those scripts along with some specs about the call center tools that we're using to generate the menus in a much simpler, much faster way.

[29:00](https://youtu.be/MEZLb5EOZG4?t=1740) And it helps reduce the technical barrier to entry for folks to make those menu updates as well. So we're getting faster information out to the end users. We also use generative artificial intelligence to extract information from documents. Very often people are uploading scanned documents which are not in a structured format and somebody has to manually then transcribe the information. AI tools, more modern AI tools that have computer vision built in are actually very good at understanding information on a form that's been scanned in and pulling it out of the structured data that can be used by an application for an agent to mark up and to verify. We've also been able to do this in bulk. So if you're batch processing something like a lot of public feedback on a rulemaking process where you might have lots of attachments, we can run the same data extraction process in bulk to get a spreadsheet of information as opposed to individual rows. Now we're building out a whole spreadsheet of information from those bulk documents. So it can help simplify the workload and make sure that we're addressing the most common themes that we see in those large corpuses. The next example we have here is around finding information across a large corpus of documents.

[30:07](https://youtu.be/MEZLb5EOZG4?t=1807) You might have lots of different case files and need to figure out where is the proof point amongst all those case files that helps us understand that somebody meets a requirement. That might traditionally mean an agent opening up file after file and looking for something. If through an AI-based search, we can give the agent an early warning signal, say, hey, the information you're looking for might be in this document, verify it. That can help reduce in that time. If the information is not found, the agent would still do the human review, the same as always. But by speeding up the process and giving an early warning signal, we can then cut out the cycles it takes to find that information manually. So these use cases are going to range from a variety of very simple, almost out-of-the-box use cases where you might have a chatbot, something that is tapping into a knowledge base and allowing people to interact with it, to more complex applications where you're building a custom application. We, on our team, provide support for this and have tried to contribute public resources. There are public resources through Innovate US, and I'm sure through the sessions to come in this series, you will learn more about this as well.

[31:08](https://youtu.be/MEZLb5EOZG4?t=1868) But as you think about the complexity, I would encourage everybody to reflect on how much complexity is necessary. How simple can you make the first use case to experiment, to make sure that you have the controls in place to learn, and then to build on top of those as well. When you get into custom-designed AI solutions, you have the opportunity to build exactly what fits your business needs, but it also introduces the complexity of needing to make sure that you're tracking all the different checks that you would for data, for privacy, for security, for acceptable, responsible use as well. So this is something that eventually will save time. In our case, by developing an AI-assisted chatbot on an open-source package and launching it internally, we were able to save really over a million dollars each year in the run rate cost we would have spent by a commercial tool. But we were able to do that because we had the ability to build a team that understood how to do this work over the course of years. So it took some years of investment to be able to do that quickly.

[32:09](https://youtu.be/MEZLb5EOZG4?t=1929) And so investing in digital services, human-centered design, agile iterative development, these are all common skill sets that are really important regardless of AI. At this point, I would love to hand it back over to Rob as our moderator here and hear any questions and discussion topics.

**Rob Asaro Angelo**

[32:26](https://youtu.be/MEZLb5EOZG4?t=1946) First of all, Dave, thank you for the presentation. As somebody who is a frequent user of our AI system in the state, I want to say thank you to your team for developing that. I think it was important for me as a leader of the department. I'm very proud, by the way, we're the first department to have all of our staff trained up on responsible use of AI, thanks to Innovate US. Mainly because, or one of the reasons why, not just the governor's support, but also showing our staff how I was using it on a daily basis. And you mentioned the regulatory process. We had a rule proposal with over close to 10,000 comments last year, and I use AI a lot in that process to sort of help discern which letters were for which side or the other, how many times an argument was being repeated, how many times was it a template form letter. So the use cases, I think, are endless. Just real quick, at the end there, you also said something about frequently used terms like agile development. I think, sadly, in government, that is still not a frequently used term. So I want to thank you and your team for helping us through our first agile procurement and iterative process and our UI modernization.

[33:32](https://youtu.be/MEZLb5EOZG4?t=2012) And I will proselytize to the end of the earth about the importance of using agile development as opposed to waterfall for everything we do. But I'll give you a little bit of a break on your voice, Dave. I'm going to ask Cass a question. Cass, you're working with human services leaders across the country. What is the top one or few barriers that they face to implementing AI use cases to assist their customers and staff?

**Cassandra Madison**

[33:56](https://youtu.be/MEZLb5EOZG4?t=2036) Yeah, happy to share. And it'll piggyback on a bunch of things that Dave said. I mean, first things first, I want to say that in the past year and a half, I feel like experimentation all across the country has really exploded. And certainly states are in very different places from a maturity and a resource perspective. But I think the number of states who aren't experimenting at all is really dwindling as people realize that, like, hey, this is, again, a tool that it seems like it's going to be transformative. And there's a lot of good here. And there's also some risks. And the only way we're going to figure it out is to start experimenting. So that's been really heartening to see. And I think there's a lot of communication and lessons learned happening across states about that. Some of the things that are barriers, things that Dave spoke to, I mean, capacity comes up in every conversation and HHS teams are typically already strapped for capacity. And so you start to layer on budget challenges, you layer on like technology systems that are end of life, the pressure to make, you know, update systems with new policies and rules on really short timelines, on existing systems that are already operationally backlogged and, you know, staff shortages and things like that.

[35:11](https://youtu.be/MEZLb5EOZG4?t=2111) And it just makes for a tricky situation. And so I think on everybody's wish list, it's like, I wish we had more time, resources and talent to be able to do this stuff. And I think the talent piece is really worth pausing on because I think like you see in New Jersey, it takes years to build up the internal bench and the muscles that are needed to do this tech in more modern ways. And at the end of the day, a lot of what we're talking about isn't, I mean, there are some things that are very unique to AI and there are some things that are just good tech practices like involving your users and, you know, focusing on product versus project thinking. And you just need people internally, whether you're buying or building the tech, who really understand how it works. And that that takes a while to develop and is really an important piece of investment for states. I think the other piece that is sort of circling around all of this is that unlike other technologies that HHS has used or experimented with or tried to implement, AI is in the public dialogue every single day.

[36:19](https://youtu.be/MEZLb5EOZG4?t=2179) So people are not only using it in their personal lives, but there's a whole bunch of discourse around it that that makes it and can make it an unsettling moment of transformation. What's the impact going to be on jobs? What is it going to mean for the day to day tasks that I undertake? And so there's also an undercurrent of fear, I think, for a lot of folks, especially frontline workers who have already dealt with so much change and frankly, broken promises about how technology was going to make their day to day jobs easier. And so there's real change management and cultural things that I think need to be addressed here and making sure that people, you know, have safe spaces for experimentation, that staff feel supported, that they understand that, you know, this is humans plus AI, not AI instead of humans. Like all of those things are going to be really important puzzle pieces to making experimentation and implementation viable. And then the last thing I'll say is that I think, you know, this letting a thousand flowers bloom, let's experiment and see what the tech is capable of is really powerful.

[37:28](https://youtu.be/MEZLb5EOZG4?t=2248) And also, it's a little bit of a different ballgame to go from experimenting in a perfect container to scaling something agency or statewide. And so I think over the next year or so, that's another challenge that states are going to face that need to be supported through, because this is where you get to the things that make AI a little different. And setting up the oversight, monitoring and evaluation systems that are needed to make sure that these technologies are running well at scale is still kind of nascent. No one has a magic answer for what that needs to look like. What's the staffing? How often should you be checking the answers? You know, what's your your beneficiary notification process? If they find something awry in a system, all those things need to be set up for these to work well at scale. And so I think that's a little bit of a journey that all states are going to have to go through together and excited to see what states like New Jersey are doing at the front end of that and how we can get those learnings back to the rest of the states that are moving a little bit more slowly.

**Rob Asaro Angelo**

[38:27](https://youtu.be/MEZLb5EOZG4?t=2307) Thanks, Cassidy. I know it was a great, I guess I would say relief maybe for us when we did we did a poll before our training about my generally of the public sector and the private sector in New Jersey. And I was encouraged the fact that public sector workers were less fearful about AI than private sector. I don't know what the reason are, but I as a bureaucrat, I think that it's because we see it as a way to help us out from underneath the bureaucracy made by policymakers that and processes that have burdened us for way too long. So I'm most excited about AI because and our use cases that we've been most successful with are ones that Dave showed, other ones that are taking the road and repetitive work off of our staff to give them more time to work with the harder cases, the edge cases to deal one on one with with claimants or constituents. So I was very helpful about that in New Jersey. We have a bunch of questions in the Q&A now, Dave, I'm going to go the one here first, which sort of touches on what Cass was talking about, my responsibility. Being that the AI solution is being used statewide, could you share what safeguards you may have to be put in place to comply with HIPAA or other privacy measures?

**Dave Cole**

[39:37](https://youtu.be/MEZLb5EOZG4?t=2377) So for the AI assistant chatbot that we put together, we are using the same IT platform that we're using for our word docs, email, all of that. So this was a way that we could leverage the same contracts that are negotiated that have the right security controls in place when we are. Well, I'll say I am not aware it may exist, but I'm not aware of a use case right now that touches HIPAA protected data using artificial intelligence tools. What we usually say is we consider the tool safe for PII as defined within our state information security manual. But when you have additional data requirements like HIPAA PACS data, for instance, which might be protected under FDI, then we want folks to consult with the folks in their agencies who are responsible for approving those use cases. So if we were to have to build something with HIPAA, we would want to make sure that the systems that we're using have the right business agreements in place to protect us on that, too. But we would be pretty careful about stepping through that.

[40:39](https://youtu.be/MEZLb5EOZG4?t=2439) I always encourage when I'm working with folks and agencies to think about only use the information that's absolutely necessary in your prompting to get the answer that you're trying to get to. So if somebody's name and address isn't useful to make the decision, then don't use it because that actually, in some ways, could potentially introduce unconscious bias in the response that you get. Right. Certain addresses might infer something that the model could pick up on versus others. So you don't have to use more information than you need when you're setting up these prompts. Think about the factors that are really relevant there. And if at all possible, avoid using HIPAA-protected data. If you have a use case that needs that, then there's probably a more specific review to make sure that you're meeting the legal terms there on top of having the right security practice in place.

**Rob Asaro Angelo**

[41:26](https://youtu.be/MEZLb5EOZG4?t=2486) Great. For Cass, maybe, but Dave also, we have a couple of questions here about balancing responsible AI or use of AI with the possible environmental impacts. Is this something that's come up with your work, other state leaders, Cass, around the country or any policies you've heard about in these regards?

**Cassandra Madison**

[41:43](https://youtu.be/MEZLb5EOZG4?t=2503) Yeah, it's just starting to. I mean, I think this is for state leaders, especially elected leaders, thinking about environmental impacts and thinking about workforce impacts are top of mind. And I think that's only going to grow over the next year or two. So it's definitely something we're trying to be in dialogue about. Unfortunately, I don't think there's any easy answers to it right now. And it's whether you're talking about environmental impacts or workforce impacts, it's going to look different depending on the specifics of your state and your workforce and your natural resources and your economy and energy mix. But it's an important thing to be in dialogue about. And so we're trying to figure out what's the right way to convene conversations so that as states are making decisions about this stuff, they can balance the things that they care about. And I think this gets back, Dave, to your comments about values and why they're so important for state leaders and agency leaders to be really thinking about what are those values and that and then you can create frameworks for decision making around that. And just the knowledge of AI's impacts are going to continue to evolve in rapid clip, just like the technology is advancing.

[42:51](https://youtu.be/MEZLb5EOZG4?t=2571) So we're all kind of like playing a little catch up here, trying to run alongside the train. But anchoring in values is the most important piece of the puzzle, I think.

**Dave Cole**

[43:01](https://youtu.be/MEZLb5EOZG4?t=2581) Yeah, I could just add a bit on that and appreciate all the points you mentioned there, Cass, that using AI to get responses, inferencing is very different than training. So if you're building out an LLM from scratch, as a lot of large companies are doing, the resource usage is going to be considerably more. But I think it's important to know exactly the usage that you're getting for your use cases, the environmental, the energy use case in particular. And as with anything, the government has buying power, we have leverage. And so I have, and we'll attest to this publicly, talk to all of our suppliers who provide us AI about what is the environmental impact in the different regions that we're operating. So specifically, what does the energy grid look like that's providing the energy behind the region that we might be in, if it's Virginia versus if we were to move our resources over to the Pacific Northwest where there's more hydropower, for instance. So asking the question, showing that we are very focused on this issue as folks who buy on behalf of public, with public dollars and public trust, I think can exert the right influence to make sure that the service providers are trying to build out efficient energy kind of infrastructure to support these use cases as well.

[44:10](https://youtu.be/MEZLb5EOZG4?t=2650) So, you know, and that's mostly on the electricity. It has to do a similar use case with the water consumption, too, which I know some of the folks are rightfully focused on, cooling the data centers is done by water, by and large. And that is also a very important national resource that we need to be cognizant of. So I think we can exert our leverage and just keep in mind that our usage may vary depending on use case, very different than a technology company training kind of foundational models.

**Rob Asaro Angelo**

[44:36](https://youtu.be/MEZLb5EOZG4?t=2676) Dave, does the Innovation Authority, there's a question from the chat, I'm sorry, from the Q&A, plan on drafting statewide governance artifacts for implementation of AI use cases within the state agencies that already exists? Is that under your shop or Office of Information Technology?

**Dave Cole**

[44:51](https://youtu.be/MEZLb5EOZG4?t=2691) Yeah. So we've done an in partnership with our statewide OIT, Office of Information Technology, and we've launched it, but as the goal, which is a state OIT O confuse This is a circular that goes out to all employees, it's basically the IT guidance. that the government in the OIT and agency that refers to it as a national We have focused on making that interactive, so there's actually a web-based FAQ that's publicly available that people can see, because we wanted to recognize the technology's evolving and people have very specific use cases. Can I use this program versus that program? How might I think about summarization versus data extraction? We put out those resources, and then on our team, too, we have a series of guides, which are basically when we work with agencies and consult on how to approach AI technology, we'll try to write up the guide for ourselves internally so we're consistent and then publish that wherever we can, so we can link those how-to guides as well from our website. Thank you, Nate.

**Rob Asaro Angelo**

[45:50](https://youtu.be/MEZLb5EOZG4?t=2750) We talk at length, thanks for the answer. We talk at length about making sure humans are in the loop, that we're testing things, and I heard a great phrase that in government, not just about AI, but everything we do, we don't get oopsie-daisies in government, that we're dealing with people's lives, benefits, protection, security, whatever it may be. Even though I'm in a lot of these panels and sessions, and I still haven't heard yet of anybody having a bad public-facing experience with AI or any kind of release of data or making a bad benefits determination. Just curious if either of you have or if there's one that's on the precipice that you're able to stop from happening. Just I know that this is something we're always concerned about in government, but I still haven't heard anybody having any problems on the public sector side of using AI for some kind of governmental function. Just curious if either of you have or how it was mitigated.

**Cassandra Madison**

[46:43](https://youtu.be/MEZLb5EOZG4?t=2803) I mean, there has been a couple of news stories here and there. I think this is what's really heartening to me about it is that I think actually most folks inside government are being incredibly cautious and focusing the experimentation on the back-end things. And I think that is really important because that's where a lot of the backlogs are. And that's where I think AI, I'm personally most excited about those back-end transformations because it takes some of the admin burden off of people's plates. So I think continuing to experiment there is really powerful. And I think the places where stuff's being rolled out, by and large, it's people are taking an incremental and cautious approach to it. Whether that changes and we see some sort of big failure over the next year or two is anybody's guess. I mean, I imagine something will happen, but I think right now we're really encouraged by how states have been approaching it in a more incremental and cautious way and focusing on those back-end things. I'm curious, Dave, if that's also been an experience.

**Dave Cole**

[47:46](https://youtu.be/MEZLb5EOZG4?t=2866) I think the most publicly visible examples of things gone awry are chatbots, the chatbot that might give somebody bad advice on how to treat an employee or a partner in question. So anything unsupervised like that, where the models are just trained on information and there's that potential for them to hallucinate a response, there's some risk. We do this in IT a lot. Think about classifying our systems about slow, medium, high risk. And think about that for AI systems, too, high risk being something that is going to make a decision about somebody's benefits or that could affect their lives, a low risk, maybe informational content not related to benefits and benefits information somewhere in the middle, just as a rough kind of classification. But then you can guide your experimentation based off of what the risk is and your tolerance along those lines. I don't want to dismiss the worry, though, that it's possible things could go wrong and it might just take some time to notice. Because part of the challenge with an AI system not being deterministic is we don't always have a lot of visibility into what exactly is happening when people put content or context into a prompt and they get a result back.

[48:53](https://youtu.be/MEZLb5EOZG4?t=2933) And there are certainly examples more prevalent early on when the technology was less refined, less guardrails, but still could exist where you could prompt it in a way that would trigger some bias in the training material. Folks aren't aware, LLMs are generally trained by pulling in a ton of content, much of which is just public information on the internet. And if you think about everything in the internet, probably not all of it's stuff that we would want to reflect as intelligence. Some of it's just there. And so if there are biases in the training materials and a prompt in some way triggers on something that was said on the internet that gets surfaced in response, it could then pass that bias into a decision, into something that somebody's reading. It's something that I think for us, the way we try to mitigate is go back to the training and make sure that we talk about these as potential risks, so that as staff are actually using the technology, they're focused on the kinds of things that could go wrong and what to look for.

**Rob Asaro Angelo**

[49:50](https://youtu.be/MEZLb5EOZG4?t=2990) I think that's a great segue, Dave. You said you wanted to answer that first question live. Do you want to address that on there about scraping the web? Yep.

**Dave Cole**

[49:58](https://youtu.be/MEZLb5EOZG4?t=2998) Yeah, so the question is about in one of the earlier slides when I described agentic use cases, I had scraping the web as one of the examples that we see there. I think that there's a comment about this. To give a more full definition, I think about agentic as really having two parts. The first is access to tools. So it can run software programs, one of which might be opening up a browser and downloading information. Another might be running a command line program, submitting to some API. And then it's largely second part doing so in an unsupervised fashion. So it's not coming back to the user every single step. You might have one prompt that builds your plan, other than all a dozen or 100 or more different prompts that are going to get executed. And so one of the tools that comes up is scraping. And I think this question rightfully touches on, I mean, scraping the web can be highly unethical if you are taking people's content and violating their copyrights. So that is not at all what I'm talking about. It's important to point out that there are questions and many lawsuits about the IP that is a part of that training documentation and training material that go into LLMs and certainly would want content creators and folks who own that IP to get compensated and have the right legal protections in place.

[51:07](https://youtu.be/MEZLb5EOZG4?t=3067) That's a matter of case law that certainly is not my area of expertise, but is being addressed and hopefully addressed effectively. And what I'm thinking more about is, you know, as much as there are challenges that AI can introduce, there are, if we're being honest, very real challenges with the way government technology works today. And one of those challenges is we often ask people to absorb a ton of information before they take a step. They might have to read lots of documents, lots of FAQs, sometimes PDFs before they're able to then get an answer about what information and documentation they need to apply for a program that they would otherwise be eligible for. And often that can happen at a time when they just lost health insurance or they lost a job and they need this help. So a lot of cognitive emotional load goes into these decisions. I think that appropriately deployed artificial intelligence can help ease that burden by synthesizing vast amounts of information rather than us just in the government saying, we put it all out there, we publish it, it's up to the user to figure out what to do with it. We try to really boil it down and help them get the information they need to get to the result they're looking for.

[52:09](https://youtu.be/MEZLb5EOZG4?t=3129) And one of the ways to do that is, in fact, scraping government websites. And I'll leave you with one example where I think this gets very real. As part of the human-centered research that we do, we run a site, business.nj.gov, and we've had folks who are applying and running businesses. In interactive sessions, we give them tasks and say, here's the task. We need to get this license for your new business. Show us how you do that. And we observe. And we expect they go to nj.gov and they look for the documentation and they read it and then they go and get the answer that they need or go to business.nj.gov and do it in one place. What we found is that people can't get the information pretty quickly. They open up ChatGPT or Cloud or a similar service and they ask the public model. And if nj.gov, as in some cases it does, is blocking all bots, including AI scrapers, from accessing that information, then the reality is people have third-party information in those chatbots, not first-party information from the government. So I do think in the government, we have to think carefully about how we want our data and our validated content information to be surfaced to chatbots as an interface for search now.

[53:11](https://youtu.be/MEZLb5EOZG4?t=3191) I mean, that's just the reality of how people are using it. It's not something we encouraged or asked them to do. It's just how they're using it. And if we don't meet people where they're at, we're abstaining from providing the most valid factual information in that place.

**Rob Asaro Angelo**

[53:23](https://youtu.be/MEZLb5EOZG4?t=3203) Yeah. And the vacuum will be filled by bad actors eventually. So I appreciate you bringing that up. I'd like to want to take a look at that second question under there. But in the meantime, there's a question about, is there a small project suggestion that a local government agency might use to help demonstrate the value of AI leadership? I'll just start off by saying just the basics of summarizing large documents, something I talked about in my email to all of our staff, or comparing documents, comparing memos, that's the most basic, I think, thing to show that there can be real value. But Cass, this is something I'm sort of like always focused on when we go to fantastic convenings around the country or go on a Zoom like this, and I hear about this great use case somewhere. I'm like, oh, I want to bring it back and do it in my state or my town or whatever it is. But there are some of those barriers you talked about before. But I feel like AI, in some ways, thanks to the Innovation Authority and setting up this system for us in the state, we're able to use it without having to go through a procurement necessarily, working with our partners at the Innovation Authority.

[54:23](https://youtu.be/MEZLb5EOZG4?t=3263) Have you seen something that's been replicated that you heard from one of your participants and is already being used somewhere else? And that goes to both Cass and Dave.

**Cassandra Madison**

[54:32](https://youtu.be/MEZLb5EOZG4?t=3272) Yeah, I mean, I think the main thing I find is that people are just so hungry to talk to each other and share ideas. So there's all sorts. You just put people who are experimenting in a room with each other across states, and they just want to talk and talk and talk. And I think because so many people can get this technology in their hands and actually experiment, it just creates a really nice environment for sharing that has previously been inaccessible. I think the document, the synthesis that you've talked about, and then content generation, like trying to generate plain language communication and things are two really promising, pretty easy use cases. I do think that there's all this. Part of what states have done out of the gate is unlock productivity, like these tools into the hands of caseworkers. Hey, you can use Copilot. You can use this. Gemini, figure out, experiment, how can you make your day-to-day more productive, which is sort of separate from, hey, let's find a problem and then figure out if AI is the right solution to that problem in a very specific standalone pilot or use case.

[55:37](https://youtu.be/MEZLb5EOZG4?t=3337) And so I think one of the things that I find interesting is what some states are doing to try to create different mindsets internal to their agencies, which is like, we're going to give people access to this tool. And we're going to create a space for them to play. And then maybe we're going to create space for a little competition, like who was able to come up with the most creative way to save themselves time this week? And sharing those lessons learned across people within an agency, I think there's some creative things like that that are really promising. Now, are those things going to necessarily scale across a whole agency? Maybe, maybe not. But you've created an environment in which people start to feel comfortable looking at their own day-to-day tasks and saying, hey, I feel like I'm doing a lot of administrative work over here. Let me say if something could make it easier. And it just shifts the power dynamic that is so typical around technology within agencies, which then I think can open up discussions to more targeted standalone pilots and sharing things across states and agencies.

[56:38](https://youtu.be/MEZLb5EOZG4?t=3398) So some of the most creative stuff I think I've seen in that space. Great.

**Rob Asaro Angelo**

[56:43](https://youtu.be/MEZLb5EOZG4?t=3403) We'll say that. I think it's a great way to end the session. As I said in the beginning, I was very, it came to my mind, we had to end on time. I want to thank Dave and Cass, both for being here today. And just personally, as somebody in the New Jersey government, how much you both helped us. And when I went through COVID, one of my promises to my staff was I'd do everything I can to make sure they never went through that again. And I feel that AI is something that's going to be a really important tool to take the stress, anxiety, threats, whatever else it may be, off of our staff who are doing everything they can every single day to help folks in the community and across our state and locality. So I'll end up there. Agata, do you have anything else we need to do for closing?

**Agueda Quiroga**

[57:23](https://youtu.be/MEZLb5EOZG4?t=3443) Yes, I will close the session now.

**Rob Asaro Angelo**

[57:24](https://youtu.be/MEZLb5EOZG4?t=3444) Oh, and thanks for all the awesome questions in the Q&A. Thank you, everybody.

**Agueda Quiroga**

[57:27](https://youtu.be/MEZLb5EOZG4?t=3447) Okay. Thank you, everybody. And we hope to see you soon in the next session. Thank you so much. I am going to close now. Sorry.

**Rob Asaro Angelo**

[57:35](https://youtu.be/MEZLb5EOZG4?t=3455) Thank you.

* * *

Original page: https://innovate-us.org/ai-101-for-human-services-leaders · Licence: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/), as stated on innovate-us.org