AV control has spent decades perfecting a simple promise: Press a button and get the expected result.
Hit “Presentation” and the displays turn on, the right source appears, lights shift, shades move, and audio lands where it was programmed. It's predictable, fast, and hard to beat.
The challenge is that “Presentation” may not mean the same thing every time or to every user. A videoconferencing update may change the workflow. Staff may use the room differently than expected. Or the person walking in may simply have a different idea of what “ready for a presentation” looks like.
AI-based control introduces another option. Instead of requiring users to work within predefined presets or translate their needs into buttons and sliders, they can describe the outcome they want. Take the command, “Get this room ready for our executive presentation.” Now the system interprets the request, looks at what is happening in the room, and determines which approved actions will get there.
That doesn't mean the familiar button goes away. It suggests that future AV systems may need two ways of getting to an outcome, a deterministic path when the action is already known and an AI-based path when the system needs to interpret intent. The question becomes: Where should AI decision-making end and deterministic control begin?
Two Paths, One Control System
The first path is deterministic: actions whose meaning is already known. A TD who wants Camera 2 on program does not need an AI system to think about it. Pressing “Camera 2” should make that switch immediately, every time. The same is true for an emergency mute, blackout, push-to-talk, direct source selection, or frequently used preset.
The second is an AI-based intent path for objectives that require interpretation. Getting a room ready for “our executive presentation” may involve cameras, displays, microphones, conferencing, lighting, shades, and routing. The user should not need to understand every matrix crosspoint, DSP preset, or camera position required to make that happen.
The distinction is fairly simple: If the action is already known and defined, take the deterministic path. If the outcome is known but the steps are not or are extensive, AI can help determine the path.
Understanding the words is only the beginning. The harder part is understanding what those words mean in this room, at this moment, for this user.
Take a seemingly simple request: “Put the laptop on the main display.” Which laptop? Which display is “main” in the current room configuration? Is the source connected? Is something already playing that should not be interrupted? Does this user have permission to change the route?
That is the context an intent-driven control system requires. It can include device state, source availability, routing, active recordings, schedules, recent actions, room configuration, user permissions, and even user preferences.
Allow AI to Think, Not Break Things
Context turns a natural-language request into something the control system can actually evaluate. Once the desired action is understood, the next question is whether it should be allowed to happen.
Consider the request, “Make it louder.” The AI may determine from the room state and conversation that the user means program audio. But that does not mean it should have unlimited control over the audio system.
AV systems already use deterministic protections such as maximum audio levels, protected DSP parameters, amplifier limits, interlocks, and user permissions. In AI-driven systems, those guardrails become even more important.
The AI can determine that the audio level should increase. A deterministic limit needs to cap how far it is allowed to go. The same principle applies to motorized equipment, protected shutdown sequences, routing restrictions, and other high-impact functions.
A useful control loop is: Interpret → Validate → Execute → Verify.
Validation checks the proposed action against permissions, device availability, operational rules, and safety constraints. If something is unclear or potentially disruptive, the system should ask rather than guess.
A novel requirement today can become a predictable one-button workflow tomorrow.
After execution comes verification. Changing a route is not the same as confirming the expected source arrived. That confirmation can come from device feedback, signal detection data, other system-state reporting, or (when necessary) the user confirming that the intended result actually occurred.
This is also why simply connecting a general-purpose AI model to a handful of AV APIs is not enough. Reliable intent-driven control requires architecture around the AI that provides context, enforces boundaries, validates actions, and confirms outcomes.
Build With AI, Recall Without AI
Here is where the two paths become especially interesting: A workflow doesn't have to stay on the AI-based intent path forever. Consider a multipurpose room being prepared for a quarterly town hall. A user asks for the presenter on the confidence monitor, slides on the main displays, remote participants on side screens, the lectern microphone active, and presentation lighting.
AI can build that configuration without someone manually working through every subsystem. The user or technician can refine it until the room is right, then save it as “Quarterly Town Hall.” Next quarter, there is no reason to reinterpret the request. Simply recall the approved preset and take the deterministic path.
A novel requirement today can become a predictable one-button workflow tomorrow. That may be one of the more interesting roles for AI in AV—not simply executing commands differently, but helping a control system adapt to new use cases over time.
Of course, not every AI action deserves the same level of freedom. A practical rule is that the consequence of being wrong should influence how much autonomy AI receives. Turning off idle equipment after hours within predefined rules is relatively low risk and reversible. Changing a critical route during a live event is very different.
Some actions can happen automatically. Others should ask, “Proceed?” Some should only be available to certain users. Some should remain entirely deterministic.
There is another practical reason to keep the deterministic path available: AI should not become a single point of failure. If, for example, an AI service becomes unavailable because an internet connection or model endpoint goes down, the technology system should not suddenly become unusable. Familiar presets and direct controls can remain available underneath the intelligence layer. AI should expand what the system can do, not take away what already worked.
Commissioning the Decision-Making
AI doesn't just add functionality to commission or program. It adds a new kind of commissioning.
The traditional question is, “When this button is pressed, does the right thing happen?” Now there is more to consider. What context can the AI use? Which actions can it take? What guardrails apply? What requires confirmation? What happens when instructions conflict? What if equipment does not respond? What happens when the AI layer is unavailable?
Designing the decision-making is one part of the job; making it serviceable later is just as important.
Six months after turnover, “the AI did something weird” is not a useful service ticket. Good systems should provide enough logging to reconstruct what the user requested, how the request was interpreted, what action was taken, and what state came back afterward.
That added flexibility may also make the ongoing relationship between a facility and its service provider more important than ever. These systems will need to remain reliable, understandable, and current as both the AV environment and the AI layer evolve.
The next generation of AV control does not need to choose between deterministic systems and AI. It will likely need both.
When the action is known, take the shortest route. Press the button, recall the preset, execute the command. When the user knows the outcome but the system needs to determine how to get there, let AI interpret the intent, consider the context, and work within commissioned boundaries. Better yet, let a successful AI-created workflow become a deterministic one the next time around.
The goal is not maximum AI. It is a control system that knows when to think, when to repeat, and when simply to do exactly what it was told.
