Skip to main content

Amplifying Human-to-Human Connection in the Face of Mental Health Crisis Using Agentic AI

Summary

  • AI-Powered Crisis Response: Crisis Text Line leverages agentic AI to amplify human connection at scale, supporting text-based mental health crisis intervention with over 300 million messages since 2013.
  • Responder Training Simulator: An AI agent workflow reproduces texter behavior to train responders on conversations ranging in difficulty, preparing them for situations where texters face imminent suicide risk or self-harm.
  • Efficiency and Quality: The platform includes automated clinical quality monitoring and predicted summarization to capture key context before conversations transfer, reducing burden and enabling responders to focus on human support.

Amplifying Human-to-Human Connection in the Face of Mental Health Crisis Using Agentic AI

Watch: Amplifying Human-to-Human Connection in the Face of Mental Health Crisis Using Agentic AI
Talk by Margaret Meagher, Chief Impact Officer, Crisis Text Line ; Mateo Garcia Pepin, Lead Data Scientist, Crisis Text Line
Crisis Text Line has been innovating for ten years in text-based mental health crisis intervention and is now leading the next wave of GenAI use cases in the space. With over 300 million messages exchanged since 2013 and a decade of expertise, Crisis Text Line is unlocking the potential of AI to amplify human connection at a global scale. We will discuss how we leveraged our bedrock application to co-navigate crisis care through a set of early AI agent workflows. First, a simulator that reproduces texter behavior to train responders in taking conversations ranging in difficulty where the texter is in imminent risk of suicide or self-harm. Second, a tool that automatically monitors clinical quality of conversations. Third, predicted summarization to capture key context before conversations are transferred. Through the power of suggestion, this compound system aims to reduce burden and drive efficiency, such that our responders can focus on what they do best — support people in need.
Databricks Named a Leader in the 2025 Gartner® Magic Quadrant™ for Data Science and Machine Learning Platforms: https://www.databricks.com/blog/databricks-named-leader-2025-gartner-magic-quadrant-data-science-and-machine-learning
Build and deploy quality AI agent systems: https://www.databricks.com/product/artificial-intelligence
See all the product announcements from Data + AI Summit: https://www.databricks.com/events/dataaisummit-2025-announcements

FAQs

What is Crisis Text Line's mission and scale?

Crisis Text Line offers free, 24/7 confidential support to people experiencing moments of duress. Founded nearly 12 years ago, the organization has exchanged over 300 million messages and uses technology to help people receive support from real human volunteers more rapidly and effectively.

How does agentic AI support responder training?

Crisis Text Line uses an AI agent workflow simulator that reproduces texter behavior to train responders. The simulator presents conversations ranging in difficulty, including scenarios where texters face imminent risk of suicide or self-harm, preparing responders for real interactions.

What AI tools help Crisis Text Line maintain service quality?

The platform includes automated clinical quality monitoring to assess conversation quality and predicted summarization to capture key context before conversations transfer to other responders. These tools help reduce burden on responders and maintain service effectiveness.

Why is human connection central to Crisis Text Line's approach?

Crisis Text Line uses technology as a tool to strengthen human-to-human connection, not to replace it. Research shows volunteer responders report improved mental health and resilience through participating in the program, making it an eminently scalable intervention for mental health prevention.

Full transcript

Welcome everybody. Excited to see so many people here today. Thank you for joining us. I'm the chief impact officer at Crisis Text Line and I am delighted to be sharing the stage today with my amazing colleague. Um, and we're going to share with you more information about what it is that we do at Crisis Textline and how we're using AI and data bricks to address the global mental health crisis at scale and make the world a better place every single day.
Um, Crisis Text Line was founded nearly 12 years ago at a time when uh 73% of teens in the US had access to a smartphone and we're sending about 100 texts a day on average. We stepped into this space to address the lack of textbased mental health support and crisis response. And since that time, we've been offering free 247 confidential support to people experiencing moments of duress. Our aim is to support anyone, anywhere with whatever they're navigating.
We use technology to help people in distress receive support from another real human being more rapidly, efficiently, and effectively. Texters in crisis are supported in our bespoke platform by trained volunteer responders who work under live supervision of clinical trained staff. We don't use technology to replace human beings. We use technology as a tool to strengthen humanto human connection. And this is super important because one of the most common questions that we get from those who are reaching out to us in a moment of crisis is whether the person whether they're being responded to by a bot or a human being.
And they want a real human being who genuinely cares for them to be on the other end of the line when they're in distress. Texters in crisis um to texters in crisis that matters enormously. and on the response side of the equation of being effective in what we do. On top of that, our research also shows that the volunteer experience that we offer through our program is an incredibly powerful and eminently scalable intervention on the side of mental ill health and prevention and crisis prevention itself.
This is important because brain disorders and mental ill health cost the global economy $5 trillion dollar a year right now. And that figure is projected to soar to over 16 trillion within five years. So it's extremely important for us to be working not just on response but also on prevention. Over 95% of our volunteer crisis responders say that their own mental health and resilience has improved as a result of this experience of engaging with us and supporting others in crisis in their moments of need.
And our volunteers say that they've also gained skills to help someone else in need of mental health support um uh off our platform in their families, schools, work, and community. and over 90% say they have done so at least once in the last six months. 42% say that they do so regularly. We conservatively estimate that for every crisis conversation supported in our platform, our volunteers are supporting or preventing another crisis moment in their everyday lives.
So here are some numbers to illustrate the scale of our impact in the US where we've supported over 11 million conversations with people in crisis over the past decade. In 2024 alone and I I should have started with saying we're a nonprofit. We're a tiny organization relative to the sea of need in the world. And in 2024 alone, our team of approximately 80 clinicians and 150 other staff building our platform, running our impact assessment and so on, together with 16,000 active volunteers, supported nearly a million and a half crisis conversations in the US for people struggling with everything from loneliness and anxiety to depression and suicidal ideiation from across every single state and every county in the country.
We also estimate that our volunteers who um supported or prevented another million and a half conversations or crisis situations um across the communities we served over the course of the same time period. That's three million mental health crisis crises responded to or prevented in the US last year alone with this tiny little nonprofit and the power of AI, data bricks and machine learning. Our model, platform, and impact are expanding worldwide.
Here's some numbers on our international impact. Our technology platform and model have supported over 15 million conversations with people in crisis worldwide to date across four countries and in three languages. Together with our affiliated partner network, uh we operate perhaps the largest volunteer crisis responder training program in the world, having trained over 100,000 volunteers to date. Our model of free 247 confidential mental health support and crisis response has been accessible to 5% of the world uh of the world's population so far.
And that scale will now triple within this year with our recent integration of an organization in Latin America that mimicked our model and then asked us to help take them over because they were not prepared to deal with the scale of everything that they need to navigate in terms of the need in that context. So we'll be supporting all the Spanish speaking world by the end of the year and we aim for at least 25% of the world to have free access to this critical service within the next three to five years.
Let me talk a little bit about how our service works. As I mentioned earlier, cris textures in crisis are supported in our platform by trained volunteer responders under live supervision of trained clinical staff. Anyone, any age experiencing any type of crisis can reach out to us by SMS, WhatsApp or web chat. And we triage conversations at the outset to move the most acute cases to the front of the queue. Much like an an emergency room, our trained volunteer responders engage in active collaboration with the texter to deescalate and help determine healthy next steps.
All of our conversations are monitored in real time by a trained professional mental health staff member who gives real-time feedback and can provide additional support as needed. Here's a highlevel illustration of how our crisis care technology brings together the critical elements of our conversation platform, our volunteer ecosystem, and our data ecosystem. The nature of our service and our approach to using AI is serious business.
It is literally a matter of life and death every day. That's why we keep humanto human connection at the forefront. We use AI to amplify humanto human connection to do a number of things. First, to maintain quality at ever greater scale. Second, to increase efficiency so that we can respond to people in need as fast as possible when they need that help. We also reduce administrative burden uh so people can focus on other people in their moments of need instead of the administrative tasks of filing, coding, whatever it is that we need taken care of.
And finally, we we also use AI to understand the features within the conversations that are associated with the clinical and well-being outcomes that we seek so that we can continuously improve our support for very very diverse and growing populations. I will now turn it over to my colleague to walk us through our model development process and some use cases. Thank you, Marad. And thank you everyone for stopping by. I will now continue the discussion and go more in detail into the use cases if this wants to work.
Okay, more in detail into the use cases uh that we're building uh with data bricks uh in AI. But before I go into a specific use case discussion, let me go a bit uh into some of the details and the interesting learnings that we're discovering along the way as we build these models. Okay, the first thing that I wanted to show is this diagram uh that basically summarizes pretty well how we work through the data bricks ecosystem to bring these models to life.
Okay, this is a diagram that should be pretty standard for anyone that is familiar with the MLOps life cycle. As you can see, we are mostly downloading uh open source models from the hagging face hub. Okay. And we're working within databicks to test and iterate on these models until we have a version that is uh good enough and ready for production. Okay. So in this picture you can see that we use several of the databicks features uh to again test and iterate till we have model ready then we we have the models ready in particular we use unity catalog to store base and fine-tune models we are also tracking and storing relevant data and relevant model results for batch use cases in the data lake we are using ML flow to track fine-tuning experiments and also notebooks to create our own uh bespoke uh automated testing and of course after having done that several times uh we ultimately deploy our models via the model serving functionality for realtime use cases or we just stand up data bricks jobs uh for those use cases that are batch okay so now let me focus for a second on model deployment I wanted to share this high level uh picture of our architecture for model deployment because it has taken us actually a a few iterations to get here and it's really like the most uh efficient way that we found to deploy these models.
At first we would have a single endpoint that would handle the whole element inference process. Okay, so it was responsible not only for the inference of the large language model but also for the business logic associated with it. Right? So the processing of the input and output data, the product guard rails and so on. And as you can imagine, every time that we needed to make changes to the prompt, we needed to redeploy the model again.
And that made sort of like our iteration cycles a bit cumbersome and really didn't make our work uh very efficient. Right now with the two this two-step architecture that you can see we basically have um an LL endpoint a naked L endpoint that is deployed via the datab bricks model serving functionality and on top of that we have an AWS lambda that handles all of the rest of the things right the prompt architecture the processing of the input and output and also the product guard rails that we might introduce to control for some of the model responses okay this ultimately this architecture comes from the realization that iteration cycles on prompts and product guard rails are faster than a model fine tunes.
Right? So whenever I'm experimenting with a prompt or some product guardrail that I need to introduce I like the idea is not to have to deploy redeploy everything but just redeploy the lab and still hit the the element. Okay. One of the best things about it also is that this decoupled structure is really helping us to test and iterate easily while also abstracting away all of the complexity from the app developer. Right?
So the app developer ultimately just has to send a simple payload to get a model response. Okay? And they're not concerned about all of the complexity that is uh behind the curtain. And ultimately it doesn't also change version to version. So that means that we can experiment on the prompt, we can do different things and experiment with the models, but the app developer doesn't need to change anything. They just send a payload, they get a response from the model.
Okay. Um, let me Okay. So another thing that I wanted to share as we iterate more and more on use cases and and build new models is that we're finding out that LLMs are indeed great at generalization. Okay. And this is making our work simpler and our deployment easier. And I will explain in a second one. Okay. So generally due to privacy and security reasons we rather use midsize LLMs. Normally we use llama models of eight billion parameters and we rather do that use those models that we manage in house and we rather do that than use larger models that are hosted elsewhere because of the very sensitive nature of our data set as you can imagine right so when taking these uh opensource models and fine-tuning them what we found out is that we can actually fine-tune the same model several times for different use cases and instruct it differently and the performance doesn't drop.
So that means that we can have a model that fine-tuned several times and instructed differently for each use case ultimately is uh giving model responses for all of our different use cases. Right? So if we go to the picture that uh I was showing before the ultimately like our deployment architecture generally what it looks like is that we have this single element endpoint of a model that has been fine tuned several times and several lambdas on the okay so as you can imagine uh this is making our deployment process simpler and also cheaper of course.
Now to the next slide if it wants to go. Okay, let me end this section before I go into the use cases showing a high level picture of our testing and rollout strategy and go a bit in detail. Okay, so it's basically a three-step a three-step process. Okay, first of all, we have the offline testing and experimentation stage. Then we deploy things to the volunteer training environment and I will explain why in a second.
And then finally, we move things to the conversation platform. Okay, so for those use cases that go into the conversation platform, we always deploy them first into the training environment. Okay, first of all, let me go into the offline process. Okay, during the offline fine-tuning and experimentation stage, we do all of the expected evaluation and experimentation uh that is a standard for these types of uh models. And on top of that we added different layers of bespoke testing and evaluation on bias security and also general use case alignment.
Okay. Another critical layer of the offline stage is the qualitative testing and review that we do together with subject matter experts. Okay. As Margaret mentioned, we have clinical staff working on our platform every day overseeing the conversations that volunteers are having with textures. Okay. So we work with these clinical trained professionals to make sure that our models do align with the reality of our platform and do align with the needs of our users.
Okay. And actually we bring subject matter experts and stakeholders very early into our product development process. So they are an integral part of the whole pipler. Right. So they not only provide help with data tagging and testing but they are also an integral part of uh product discovery use case validation and risk assessment. Okay. And actually for risk for the risk validation process let me stop for a second because that's a process for which we have stood up a formal uh process.
Okay. And we do so because we want to make sure that we identify all of the risks associated with model failure and implementation and build the right guardrails for them before we even deploy uh to the platform. Okay. Why? Because of the high stakes nature of our service and the high stakes nature of the environment that we operate. and on that uh after offline testing we don't go directly to the conversation platform for those use cases that end up going to the conversation platform what we do is that we still need to consider the feature roll out with care so we what we do sorry um and we do so because we want to make sure that we have validated our models at scale before putting them in the conversation platform so we basically need a handbox environment to be able to generate a strong feedback loop that really gives us model performance at scale.
And that for us is our training environment. Okay. And this is a great feat because our training environment is not only free of uh the texture side of things and texture data, but it's also our community of volunteers is a very engaged community of volunteers that is willing to help us and is willing to sort of iterate with us into make the making this models the best possible. Right? So with them with the community of volunteers we create a robust feedback loop that allows us to test this at scale before we even take it to the conversation platform.
Okay. Once the model has been validated at scale in an environment that is not consequential we can deploy to the high highly consequential environment that is our conversational platform. Even when doing so, we do so in a way we are applying safe design patterns and practices and we make sure that we keep human in the loop across our implementations. So again, the idea with uh these AI use cases that we're building is not really about um replacing uh or fully automating entire workflows, but rather augmenting our users so that they can be more effective and more efficient in the work that they do every day. both responders, volunteers and supervisors.
Okay. So now after discussing this, let me go into the use cases. I will now begin to discuss the AI enabled training uh that we're working on to provide more confidence to our volunteers before they go into the platform and also add some details into how we're using this environment to generate a sandbox environment that is able to uh is allowing us to test at it and iterate on use cases that end up going into the platform.
Okay, so let me start by saying a few words on the model that is supporting the texture simulations that are the really the backbone of our AI training, right? Uh these simulations are available in English and Spanish and it's basically a fine tune lama 38 billion. Uh and we provide the model with enough context by fine-tuning with a set a set of synthetic role plays that we have built together with our clinical. Okay.
After fine-tuning the model, we prompt it to the task of roleplaying a textter who's experiencing a specific mental health crisis. Okay, so with this approach with different prompts, we can create different experiences and different types of exercises without having to fine-tune every time. But at the same time, because of the fine-tune, we have we have given the model enough context so that the model has a good understanding of how a crisis conversation feels and is able to role play that every time.
Okay. So to expose these users to these AI models in our training and coaching life cycle, we created a standalone app with this different simulated scenarios that is open to the entire US volunteer community and is recommended by coaches and training specialists. Okay. Right now it is recommended for optional practice. Okay. As of today more than 4,000 unit users have used the app. about 25% of those are returning for additional training and simulations last about 35 minutes.
So this is already telling you that people are generally engaged with this tool and do spend a considerable amount of time practicing with uh with the simulated scenarios. Okay. By the end of uh Q2 we are working to further enhance the footprint of these models by delivering the first set of simulations that would be fully integrated within our volunteer train protocols. So, not as optional practice anymore, but as an integral part of the volunteer training experience.
Let me just spend a minute showing a few real examples of qualitative feedback that we're getting on the tool. Uh, this one in particular is from someone who just finished their training, is getting ready to go to the platform and speaks about how these immersive experiences can help him boost uh their confidence and increase their skills. Okay. The second one, the second one is from someone who already has platform experience and is using the tools as a skills ref.
Okay, this is someone that as they reactivated was probably directed by your coaching team to do some simulated practice before going to the platform. Again, these are real quotes uh from users of our training tool. It's interesting how gathering so much feedback is making us learn new ways about how people are interacting how people are interacting with this this tool that we haven't uh accounted for. So for instance we received a few messages already of people using the this uh simulator tool to warm up before taking a shift in our platform.
So having a few practice conversations and simulated conversations and then go and handle the real life stressful situations. Okay. So as you can imagine this is very interesting for us and it's really feeding our product ideation process. Okay. So after showing how we we utilize AI to create training experiences, the other thing that I wanted to share is our work on user augmentation. Okay. I want to preface this by saying that our intention is not to build these agents to fully automate things but rather augment user capabilities by providing suggestions and guidance. and with that reduce burden and drive efficiency.
Okay, so let me show you a couple of examples. So let me first talk about summaries and in particular summaries for conversation transfers. What is a conversation transfer? So let me give you a bit of context. Volunteers handling conversations might need to transfer a conversation to another volunteer because they need to log off the platform or any other reason. Okay. So in these cases the volunteer who receives the conversation needs to go ahead and pick up the conversation where it was left and start interacting this with the text.
Okay. So right now manually writing a summary in the conversation transfer uh form is encouraged but it's rarely done by conversation by the people that request the transfers. So the receivers do feel overwhelmed as they receive these conversations and need to read in detail through the whole transcript uh to really have an understanding of where the conversation was left, what was the issue that was being discussed and all that.
Okay. So automatically generating a brief summary can be very useful in providing the necessary context for the receiver so that we improve workflow efficiency but we are we are also maintaining quality on the transfer conversations. Okay, the interesting thing here then for model building is that since LLMs are great at summarizing, we have been able to develop an accurate model with just promptly generated. So we are not fine lama 3 for this use case, but rather utilizing a detailed prompt on how to summarize the conversation with the right set of guidelines.
Let me now show you two quick demos to sort of uh guide you through how we are uh deploying the summaries use case to the conversation platform. Um so this first demo is just a practice simulation and it shows how the summaries model is integrated within this so that we can test it with the help of our volunteers. As you as you could see, I'm selecting a scenario. I can select English or Spanish. I selected a Spanish scenario.
And this scenario particularly drops me in the middle of the conversation. And this is why and that's um and we do this because this scenario is meant to practice a specific thing which is the safety assessment model that uh responders uh go through when the text has confirmed that they're having thoughts of suicide. Okay. So this conversation already drops you at the place where the texter has confirmed thoughts of suicide and now you're able to sort of uh ask concretely about their suic the suicidalation of the texter to better understand uh whether the text is imminent risk imminent risk of suicide or not.
The model is responding back of course we introduce some delay so that it feels more realistic because the API has already responded. And now let me close the conversation. just for example purposes the text the user sorry is able to opt in to testing the summaries feature and we just show them the summary we are we allow them to tag to tag it for accuracy and also the next question that you will see it's a question that is this one it's specific to how the summaries model will be integrated within the conversation platform so we're not only we're not only trying to get uh we get tested model alignment but also whether the implementation that we're thinking for the model is aligned with uh the needs of the users or not.
These two these other questions that you see here are more related to the practice itself. So as you can see we're gathering a bunch of feedback from users and that is really for uh our processes. This second demo is really how the summaries workflow that we're working on right now is going to be integrated into the conversation platform. Let me Okay, so I'm just having a conversation with a texter. I request a a transfer of that conversation and as I request the transfer, the conversation transfer form opens up and I do need to fill out uh some questions.
Okay. And now I'm able to either manually write a summary or automatically generate. Okay. So there is there is optin to the feature. Once the summary is generated, I'm able to review it and edit and I'm forced to approve it. So as you can see, we're introducing explicit human in the loop friction so that we account for model failure. And ultimat ultimately LLMs are probabilistic machines, right? So there is all there will always be an uncertinates.
So we do need to introduce that friction to make a an experience a product experience that is forgiving to model failure and that's it. This feels small. It's just a summary, right? But the environment in which this AI model is deployed is high stakes. It's life or death, right? So this is why we're taking all of these steps and making sure that we're doing things right so that we can really empower our users but we're not putting anyone at risk.
Okay. Um now let me show you the last two use cases that I wanted to discuss that I think are relevant for what we call user augmentation. Okay. The first use case is one that continues to be in the training realm and it's about enabling our staff to provide effective and efficient feedback to trainees and volunteers. So basically the problem that we're facing is that in order to provide feedback or grade exercises, our coaches and training specialists need to manually review conversations and look for the specific parts of the conversations where the skills that are being graded are showcased uh by the volunteer.
Right? So this manual review process takes time and creates burden for our staff that could be greatly reduced by some level of correctly implemented automation. Okay. So the idea behind the model is simple. Since language models are great at understanding nuance, let's have a language model answer questions about the conversation. In particular, questions about whether the trainee is adhering to best practices or not, right? so that the specialist grading the conversation the conversation exercise can then take that pre-anser question rubric to craft a specific and relevant feedback to the trainee without having to go and do all of the work uh themselves.
Right? So again when correctly implemented this reduces conversation review time while also maintaining the human uh review element which is critical. Currently we have this model deployed as part of workflows for coaching staff for coaching staff focused on recent graduates and volunteer retentions retention and by Q3 the idea is to be done to begin to support exercise grading that is part of our initial uh training protocols.
Okay, this is a batch use case by the way one interesting thing that I wanted to briefly mention that you can see here is the two-step nature of the inference process. Okay, so we we basically have what we call the face classifier model and then the question answering models and as you can see the question answering models might be differing might be different depending on the ropic question both predict sequentially to generate the most robust prediction and really this structure comes from the realization that the signal that contains the answer to the different questions of the train rubric has a local component to it right so it displays within a specific stage of the conversation.
Okay, so breaking apart the conversation uh into different phases and just feeding the relevant pieces to each question reduces noise and makes the work of the question answering model easier. Okay, so we basically do that to ultimately increase the accuracy of the models and I wanted to mention this because this is an example of even in the world of very powerful LLMs having an understanding of the problem space is critical and contributes to the to develop the best possible right uh this is a qualitative uh realization that uh gets to you by talking to subject matter experts and by understanding the use case Okay.
And uh let me go to the last slide and present the last use case. This is actually just a model for now. We're thinking about uh use case implementations and it's uh our work on identifying suicide problems. Okay, let me give you a bit of context on the safety assessment model that uh our responders perform when the texter is in has confirmed that they're having thoughts of suicide. Okay. So in order to gauge the true risk of a texter dying by suicide, responders work with the texters to better understand their suicidal ideiation.
And in particular they look at four different dimensions that you see here. The desire, the intent, the capability and also the buffers. Okay. And understanding this and formally going through this process is critical to really gauging the risk and to understanding whether we need to call emergency services or not. Right? So this is a complex policy that it's difficult to enforce and difficult to understand as a volunteer.
Right? So you can imagine that building a model that is able to autotag messages in training for these four uh for these four dimensions can really generate more engaging experiences and allowed to better understand the volunteer. uh the the model and how to apply. Okay. Another use case that uh you might think with this is that when coaches or training specialists are reviewing uh conversation transcripts for grading, you could actually uh send them a a transcript that's that has been pre-tagged for these different dimensions.
So that when the person grading the exercise, the coach or the training specialist is going through the conversation, they already know where to look for, right? So we already kind of tell them the answer about whether the uh volunteer or trainee is uh questioning about intent is here because these are the intent related messages or these are the capability related messages. Okay, so this is a very powerful model that we're thinking about how to implement so that we can really make our training experiences better and empower our volunteers to be uh the best responders possible.
With this I end by saying that we are hiring. So if you have talent that want to work with us, come work with us. We do need volunteers. So uh think about becoming a crisis counselor. And we uh also uh are looking for donations. Thank you so much for attending. Uh this has been a pleasure. If you have any questions, we'll take them.

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.