At the Ad Council, one of my guiding principles is simple: technology should be used as a force for good. Sometimes that means using technology to meet people where they already are. Other times, it means using it to remove friction between people and the information they need.
That’s exactly why we wanted to pilot conversational AI for “Agree to Agree,” our youth firearm injury prevention campaign. Firearm injuries remain the leading cause of death among children and teens ages 1–17 in the U.S.. Through our research, we know that many parents and caregivers want to talk about firearm safety with other parents, but the fear of being misunderstood, a lack of personal experience, or just not knowing where to start can be obstacles. Together with our partners, we began with a question:
What if AI could help people feel more prepared to have those conversations?
Why conversational AI?
People are increasingly engaging with AI tools like ChatGPT, Claude, Gemini, and Copilot: asking questions and seeking clarity, particularly when trying to understand something complicated. According to a 2026 Pew Research Center study, nearly half of U.S. adults now say they use AI chatbots, and nearly one in four (24%) use them every day. That’s a meaningful shift in how people are finding and interacting with information.
For marketers, that behavior change creates an interesting opportunity, especially when the information someone needs is sensitive, nuanced, or difficult to discuss. Conversational AI enables people to engage with information more naturally, asking questions in their own words and receiving guidance tailored to their specific situation or concern.
We wanted to bring that experience directly into the campaign website. Someone might want to know how to ask another parent whether there is an unsecured firearm in their home. Someone else might be looking for information about secure storage or guidance on how to start an age-appropriate conversation with their own child. Through the chat experience, they can describe their situation and be guided toward relevant, trusted, and vetted campaign information hosted on the website.
The goal wasn’t to build AI for AI’s sake. It was to test whether a conversational interface could lower the barrier between a person and information that could help them take action.
Building the partnership
Creating a conversational AI experience around a topic as sensitive as youth firearm injury prevention required more than technical expertise. We needed partners who could help us think deeply about trust, safety, user experience, and governance from the very beginning.
- Microsoft provided the technology stack, including Microsoft Copilot Studio as the conversational orchestration layer alongside Azure OpenAI, Azure AI Search, Azure Blob Storage and Power Automate. The technology also provided enterprise-grade security, governance and controls that were critical for a use case like this.
- RSM US LLP, a Microsoft implementation partner, architected and built the conversational experience.
- DEPT® served as a strategic funding partner and advisor, bringing expertise in conversational AI experiences and helping us think through how the technology should show up for consumers.
Working side by side, we all agreed on something important: the experience couldn’t just be useful. It also had to be safe, transparent, and trustworthy. Users needed to know they were interacting with AI rather than a person and the system needed clear boundaries around what it could and could not discuss. Additionally, given the subject matter, safety considerations couldn't be something we just layered on at the end; they had to be a part of the architecture itself.
What we learned about building conversational AI responsibly
One of our biggest lessons was the importance of grounding the AI in vetted content rather than the open web. We use Azure AI Search to retrieve information from a curated knowledge base made up of approved campaign resources and content hosted on the campaign website. The AI chatbot isn’t searching the internet and improvising an answer; the information available to it is intentionally constrained to content our teams have reviewed and approved.
That distinction is important. Retrieval architecture like this doesn’t magically eliminate every risk associated with generative AI, but it significantly reduces the opportunity for unsupported answers and gives organizations far greater control over what informs a response. For a sensitive issue like youth firearm injury prevention, that control was essential.
We also learned how important it is to guardrail both topic and tone. Not every question should receive a generative answer, so we categorized different types of inquiries and created specific conversational pathways, including resource-search topics, informational redirects, fallback responses and crisis flows. Questions outside the scope of the campaign are redirected rather than allowing the AI to wander into areas it wasn’t designed to address.
The result is intentionally a narrow-purpose AI experience rather than a general-purpose chatbot. I think that’s an important distinction for marketers: more capability isn’t always better. Sometimes a more constrained AI experience actually creates a safer, clearer, and ultimately more helpful user experience.
Another major consideration was designing for the highest-risk interaction, not just the average one. When you’re building technology around a sensitive subject, you have to think seriously about what happens when someone says something you weren’t expecting. We built dedicated crisis detection and response pathways for signals involving threats of violence, self-harm, and other high-risk situations, along with content filters and predetermined crisis messaging designed to guide users toward appropriate support resources. Those situations may represent a small percentage of interactions, but they deserve an outsized share of the design thinking. Responsible AI isn’t only about how the experience performs when everything goes according to plan; it’s also about anticipating what happens when it doesn’t.
Transparency was equally important. We were very deliberate about making it clear that users are interacting with an AI-powered experience, and the tone of the AI chatbot is intentionally neutral, clear, and solution-oriented. It isn’t overly emotional, overly familiar, or designed to make someone believe there is a person on the other side of the screen. As conversational interfaces become increasingly human-like, I believe that distinction is only going to become more important.
Test intentionally before you launch publicly
We intentionally put the experience into market in early 2026 and spent months testing, iterating and, in many cases, actively trying to break it. We tested prompts, edge cases, redirects, tone, resource retrieval and safety mechanisms, continuously refining the experience based on what we saw.
That quiet runway was invaluable. There can be pressure with emerging technology to announce the shiny new thing as quickly as possible, but this experience reinforced the opposite for us: give yourself room to test, learn, and iterate before you launch publicly. The weeks and months spent testing were not separate from building the product—they were how the product got built.
Ahead of Cannes Lions 2026, we began sharing more broadly what we were learning. I joined our partners Jennifer Kattula at Microsoft and Carryn Quibell at DEPT to walk the audience through a visual demo and talk about the thinking behind the work—a conversation I wrote more about in my Cannes recap.
What we’ve learned so far
Because the chatbot is a relatively deep-funnel action, we’re evaluating not only how many people opened the chat, but how they engaged with the experience and if it is actually helping them find the information they need.
So far, we’ve seen a 3.3% conversion rate of people that open the AI chatbot and click on a recommended resource. These resources include conversation guides and content from the "Agree to Agree" campaign’s trusted resource library, which provides resources for parents, concerned adults, and firearm owners. Each resource click represents someone taking an additional step to learn more about preventing firearm injuries and having potentially important life-saving conversations. Among users interacting with the chatbot as intended, the median conversation includes about 2.5 messages.
The AI chatbot was designed to help parents and other concerned adults ask questions in their own words and receive guidance tailored to their situation. Qualitative review of the conversation transcripts provides another window into how people are using the experience. People are asking about secure firearm storage, how to raise the topic with another parent, and how to find conversation guides to talk to their own children.
We view this pilot not simply as an AI-powered chatbot exercise, but as an opportunity to to better understand conversational interfaces when they are actively seeking information and guidance. The questions people ask can also provide valuable insight into the resources, concerns, and moments of decision-making that matter most to our audiences.
Start with the problem, not the technology
For me, the biggest takeaway from this work has little to do with any particular AI platform. Technology wasn’t the starting point; the human barrier was. We had people who wanted to have important conversations but often didn’t know where to begin, and we explored whether technology could make taking that first step a little easier.
If you’re a brand, marketer, or nonprofit considering whether conversational AI belongs in your user experience, I think that’s the place to start. Ask what is hard for the person you’re trying to serve today, where friction exists in their journey and whether a conversation could make that experience easier. Only then should you start thinking about conversational AI technology.
We spent an enormous amount of time thinking through AI disclosures, guardrails, crisis pathways, knowledge architecture, placement, tone, security, governance, and testing. On a sensitive and nuanced issue like youth firearm injury prevention, every one of those decisions matters. Ultimately, the technology is only valuable if it helps create a meaningful human outcome.
In our case, that outcome might seem remarkably small: a parent feels a little more confident with steps they can take to help to protect their child, a caregiver accesses secure-storage resources, or someone finally feels prepared to start a firearm injury prevention conversation. Those small moments are the point. When technology helps someone get information or take actions that can help keep a child safe, that's technology for good. That's technology being used exactly as it should be.