AI Trailblazers: Matthew Paupst on Integrating AI in Technical Support

In our ongoing AI Trailblazer Series, we speak with forward-thinking leaders who are successfully integrating AI into their operations. This interview features Matthew Paupst from Crestron, a global leader in workplace technology, automation, and control systems.

As the Vice President of Global Technical Support at Crestron, Matthew is at the forefront of transforming customer experience through cutting-edge AI and automation. With over 15 years of expertise, he leads initiatives that not only streamline support operations but also drive customer and technical support innovation, ensuring Crestron remains a pioneer in the industry.

Matthew has been at the forefront of evaluating and implementing various AI solutions at Crestron, taking an experimental and pragmatic approach to finding the right tools for specific business needs. In this candid conversation, Matthew shares valuable insights on selecting the right AI tools for different tasks, gaining organizational buy-in, and planning future AI capabilities.

Adapting Different AI Models for Varied Business Needs

Q: Can you tell us about Crestron's journey with AI and how you've approached different tools and models?

Matthew: We've evaluated several different tools and native LLMs, but we found that no single model fits all our needs. Sometimes, we do code evaluation, which works best with Gemini and Llama. For emails, Claude is more effective. For specifications and data analysis, ChatGPT often performs better. So, for us, being able to switch between different LLMs based on the specific need has been what's most attractive overall.

Q: How did you begin your AI implementation journey before partnering with Dashworks?

Matthew: We have been investigating various platforms. We saw quite a few demos of a few all-in-one CRM platforms, and they have their pluses, particularly suggested answers, as an interactive LLM is live streaming data coming into cases. This experience gave us a foundation to understand what was possible with AI and what we wanted to achieve with our implementation.

Q: How are you personally working with AI tools daily?

Matthew: I've been tinkering with several platforms. I have accounts with Claude, ChatGPT, Perplexity, and Gemini. Each has distinct strengths for different use cases. I've been particularly impressed with the deep research capabilities available in some AI platforms. When using these advanced research features, they dig into even simple prompts, listing out their thinking process and providing comprehensive results.

Promoting AI Adoption Within the Organization

Q: How have you helped your team adopt Dashworks and other AI tools in their workflows?

Matthew: I've been sharing effective prompts in our Slack channels, telling the team, "Hey guys, this is a good one, keep this one." I even created one of the pre-configured prompts in Dashworks and shared it, saying, "By the way, in case you ever want to know what I look up when I get a case escalated to me, put the case number at the end of this prompt." This way, everyone can ensure they have everything needed for a case escalation. The team was quite impressed, and it helped them understand how I use Dashworks to quickly get background on cases before even reading the entire email. Now, they can see precisely what managers and directors do when a case is escalated, allowing them to prepare information before escalation better.

Q: What approach did you take when introducing AI tools to your team?

Matthew: I started with a select group of my team, many of whom had years of tribal knowledge, and others were new hires. I worked with people I know have the answers, and people I know will struggle to find answers. Then I had them both try the system and monitor their performance to see if the people who didn't know the answers were now getting them and if the people who did know the answers were validating that the AI responses were good. It's the whole premise of "How do you eat an elephant? One bite at a time." If somebody thinks they're just going to turn this thing on and—poof—the magic happens, it doesn't work that way.

Overcoming Concerns

Q: What challenges have you faced when integrating AI into different workflows?

Matthew: The first thing is ensuring the Legal and InfoSec team is on board. The technology is the technology, but educating the people who aren’t comfortable with this emerging technology is key. The biggest challenges are getting buy-in and making sure the stakeholders are educated accordingly. 

Q: How do you address the broader concerns about AI security?

Matthew: I use this analogy: if you lived in the state of New York and watched the news every night, you'd never step foot in New York City. You'd think you're going to get mugged, shot, stabbed, pushed in front of a train, lit on fire, or punched in the face. You'd never go to New York City! But in reality, it's one in a million that something terrible would happen to you.

The hurdle is that legal and InfoSec teams are often focused only on AI's negative aspects. All the good it's done—medical research, synthetic vaccines—means nothing to them if they're fixated on potential data breaches. So education is crucial.

Lessons from a Single Platform Dependency

Q: Have any experiences shaped your approach to AI implementation?

Matthew: We learned a valuable lesson about a month ago. We had a problem with a critical system that crippled us for a week. We had to reorganize how we did business internationally. They had a team on it but could not work fast enough to get the system back up and running.

A bug entered a recent release but didn't surface until the ingested volume increased. Once we crossed this threshold, it just fell apart. What made it worse was that it wasn’t a quickly deployable patch even when they had a fix. That's not agile.

This experience raised concerns about going all-in with one platform. Going all-in on a single platform leaves a potential for disaster; we're completely stuck. It's opened our eyes to the risks of putting all our eggs in one basket, forcing us to look at different solutions for things like CPQ and voice systems. We're making it so we have a core—like the central database—but everything else is a bolt-on. If the primary database goes down, we can still log into these individual bolt-ons independently and continue to do business.

Future AI Priorities and Capabilities

Q: What AI capabilities are you most interested in developing for Crestron in the near future?

Matthew: Our priorities are:

1. Proactive trend analysis - This is what's been killing us. I would love a system that doesn't just tell me we had 20 cases on a product but identifies why there's a trend and what's in common. For example, if we push a new firmware version for a processor, are we starting to see an influx in calls since we released it? And if so, why?
2. External-facing chat systems - Allowing people to self-serve information.
3. Deep research capabilities—particularly for supporting technical document understanding in C#, HTML5, and CSS. Our new panels are all coded in HTML5 and C#. While we have documentation available through authenticated login and a WYSIWYG software platform, many of our dealers aren't C# programmers or are novices because they've been forced into doing this. Achieving more complex applications is difficult.   Having AI that understands our specific technical documentation to guide users to the desired result would be invaluable.
4. Multimodal capabilities - Especially for technical drawing analysis. I'd love to be able to upload a technical drawing and have AI create a bill of materials based on the blocks it sees, or identify errors in connections, like if somebody has plugged an HDMI cable into a network port on a drawing.

Q: Could you elaborate on what you mean by proactive trend analysis?

Matthew: I envision doing two scans a day for trend analysis. For example, if we've gotten 1,000 cases in, and 200 of them are for a particular product, and 150 of those have been since we pushed a specific firmware version, I want deeper analysis: What are the trends? What are the symptoms? What is the resolution?

This is crucial because our business operates globally. When the US goes to sleep, Australia, Asia, and India are waking up, then Europe's waking up, and then it's back to us again. So instead of having each region entering and troubleshooting the same issues independently, we could identify trends across regions much faster.

Q: You mentioned supply chain management as a potential AI application. Can you explain that use case?

Matthew: I'm interested in being able to say, "Based on the open opportunities, compare this to the run rate of the attached file or location, and do the predictive analysis on forecasting supply chain based on the identified SKUs over the next 30, 60, and 90 days."

The system could come back and say, "Based on the open opportunities with a 50 percent or higher anticipated close rate, you should order 5,000 of these, 2,000 of these, 10,000 of these, and for safety stock, order 50, 50, and 50 of these based on possible surge orders." There are models out there that do this, but they charge a hefty price for it.

Implementation Advice for Organizations

Q: What advice would you give to companies looking to integrate AI into their operations?

Matthew: First, make sure your legal team is on board and well-educated about the technology. Send them to conventions and seminars where they can learn about this stuff, and keep them away from just listening to sales pitches.

Second, start with a small, targeted group—your most product educated performers and your newest hires—before rolling it out more broadly. Have them compare notes so you can find the similar threads of what helped and what didn't, then feed that back in as you bring the remaining team members into the fold.

Third, when implementing new systems, consider the "Greenfield" approach. This was the biggest thing we did when we went live with CRM and our knowledge base. We did not import 20 years' worth of data. I would recommend that if possible, you keep your old system, either export it or bring it in as a separate data set somewhere that you can filter. This makes the transition much cleaner.

Finally, maintain an experimental mindset. Be willing to try different tools and platforms to find what works best for your specific needs rather than trying to find a one-size-fits-all solution.

Key Takeaways

Matthew Paupst's experience at Crestron offers valuable lessons for organizations at any stage of their AI implementation journey. His pragmatic approach—using different AI models for different tasks, starting with small targeted user groups, gaining buy-in through education and demonstration, and maintaining system independence—provides a roadmap for effective AI adoption.

Most importantly, his insights highlight that successful AI implementation isn't about finding a single magical solution, but rather building a thoughtfully integrated ecosystem of tools that address specific business needs while maintaining security, agility, and user adoption.