Into the Swamp: American Sound AI Integration for AV Systems

‍American Sound is formally launching an integration practice for agentic AI in AV systems. We're doing that as an Anthropic Services Partner, as a member of the OpenAI Partner Network, and as a contributor to Swamp, a deterministic agent harness that makes agents safe enough to put in production environments. Three separate relationships (with more on the way), and one practice.

‍Great, what does that mean?

‍AI suffers from a bit of over-saturation and dubious product marketing online, so for the rest of this post I'm going to be as specific as I possibly can. Our role here is integration. We connect standalone Claude, GPT, and other agent models to client AV systems such as Crestron, Q-SYS, Lightware, Cisco, and more. It's a software integration practice derived from our existing offerings around digital signage and collaboration, and because control and management APIs have matured to the point where this is actually achievable, we felt it was time to formalize the delivery process so that any customer anywhere looking for a support partner for agentic integration with an AV system has somewhere to go. The reason those three partnerships are listed separately is that we're deliberately model agnostic. We are not in the business of telling folks which frontier model to standardize on, partly because people have probably already made that decision at the enterprise level. If someone shows up with Claude, we'll integrate Claude. If someone wants GPT 5.6, same. We've even had success with open-weight models like Qwen or OLlama. Regardless of model choice, the work looks essentially the same. The hard part of this process is the AV system and the workflows and governance built to integrate other systems with it. It's also very important to point out that this is not just a consulting engagement that ends in a slide deck, a workshop, and an invoice. At the end of working with our team on AI projects, customers have agents running against actual room systems, doing actual things, with activity logs that can be handed to a security team.

‍Is this actually possible?

‍Yes. In true American Sound fashion, our first case study was ourselves.We've integrated hosted agents with several internal Microsoft Teams Rooms and Webex systems, plus several digital signage platforms across our own offices, and those agents can raise and lower system volume, push firmware to endpoints, update control system UIs, change TV inputs, and handle a growing list of small tasks that otherwise would generate a help desk ticket and, possibly, a truck roll. The mechanism, roughly, is treating control systems as API endpoints for AV device orchestration. You can now have Jarvis in your conference rooms, and we can help. By the way, if you work in IT and you just read "push firmware to endpoints" and felt something clench, good! That leads me to my next point, which is:

‍What does deterministic mean, and why should you care?

‍A frequent complaint about production AI use, and a valid one, is that AI has a tendency to hallucinate, or functionally invent an answer when it can't find one for itself. This has improved considerably over time thanks to the work of Anthropic, OpenAI, and everyone else pushing the frontier, but "improved" isn't really good enough when something has write access to the same AV system running an all hands meeting or a broadcast feed. So, we don't just hand an agent open ended access to gear and hope for the best. We work with Swamp, an open source harness for deterministic agent execution that we contribute to and run in production ourselves, to define workflows ahead of time. Our deployed agents don’t decide what they are permitted to do in the moment. That is pre-decided when the workflow was written, by a person, and it's enforced by the harness as opposed to really great "prompt engineering" or a complicated agent skill markdown file. Literally anything outside the predefined path doesn't happen. All actions sit behind approval gates, and every workflow execution produces an audit trail showing what ran, when, on which device, and under whose authority.I honestly do think that is missing from most of what people are being sold right now, and that's a really big part of why we were willing to put this in our own environment for long periods of testing, because the industry needs practical and responsible AI integrations if AV and IT are to successfully adopt this tooling as an accelerator.

‍This is cool, how do I find out more?

‍Good news, we love talking to people. Reach out at marketing@asei.com and we'll set up a call to talk through what you're trying to accomplish, which in our experience is usually a more productive conversation anyway than just a generic one about AI.  We take our responsibility as an integration partner very seriously, and the team at American Sound is very excited to be in the position to help organizations everywhere with AI for AV and IoT.

Doug Schaefer
Chief Technology Officer
dougschaefer@asei.com

‍American Sound is a US-based audiovisual and information technology systems integrator and managed services provider.  For more information on our software development and integration practices, please visit https://www.asei.com/solutions and check out our Github at https://github.com/American-Sound.  For more information on Anthropic and their frontier models, please visit https://www.anthropic.com.  For information on OpenAI and THEIR frontier models, please visit https://openai.com.  And finally, please check out swamp-club.com to learn more about Swamp and how determinism affects successful AI adoption.