I've been thinking about this for a while: we have gotten pretty good at connecting things to the internet.
A machine can tell us its temperature.
A tracker can tell us where an asset is.
A sensor can tell us when something moves.
A camera can tell us what's happening in a particular area.
The technology for collecting all of this information has come a long way.
But there's still a pretty basic problem.
What do we actually do with all that data?
That's where I think the next stage of IoT gets interesting.
A dashboard isn't the same thing as intelligence
A typical IoT setup isn't necessarily complicated.
You put sensors on something, collect the data, send it somewhere, and eventually show it on a dashboard.
That can be extremely useful.
But imagine you're managing a large warehouse with thousands of assets.
Knowing where everything is sounds great.
Then you start getting thousands of location updates.
Then status updates.
Then temperature readings.
Then movement events.
Then alerts.
Pretty soon, the system has more information than a person can reasonably look at.
The problem isn't a lack of data anymore.
It's figuring out which parts of the data actually matter.
That's one reason I'm interested in the combination of AI and IoT.
AI can potentially help turn a huge stream of individual events into something a person can actually use.
The interesting question isn't "what happened?"
It's "why?"
This is where things start changing.
An IoT system might tell you:
Asset 247 moved 18 times today.
Useful information, but not necessarily a useful conclusion.
A smarter system could potentially look at that activity alongside other information and ask whether the pattern is unusual.
Maybe that asset normally moves twice a day.
Maybe it's being sent between two areas unnecessarily.
Maybe the movement is connected to a bottleneck somewhere else in the operation.
Now the data becomes more interesting.
We're moving from simply recording events to trying to understand them.
That distinction matters.
AI doesn't magically fix messy data
There's also a part of AIoT that doesn't get nearly as much attention.
The data can be messy.
Really messy.
Sensors don't always behave perfectly.
Devices can go offline.
Different systems may use completely different formats.
Some equipment might be brand new while other equipment has been running for years.
You can end up with missing readings, duplicate events, inconsistent timestamps, or data that doesn't quite line up.
Putting an AI model on top of that doesn't make the underlying problem disappear.
If anything, it makes the quality of the data even more important.
So before talking about sophisticated models, there's a lot of fairly ordinary engineering work to get right.
Device management.
Connectivity.
Data pipelines.
Storage.
Security.
Monitoring.
Integration with existing systems.
None of that sounds particularly exciting when you're talking about AI, but it's often what makes the whole thing possible.
Not everything needs to go to the cloud
Another thing worth thinking about is where the processing should happen.
Sending data to the cloud is convenient, and for many applications it makes perfect sense.
But imagine a system that needs to react almost immediately.
Or a site producing enormous amounts of sensor data.
Sending every single raw reading somewhere else first might not be the best approach.
Sometimes it makes more sense to process at least part of the information closer to where it's being generated.
That's where edge computing comes in.
A device might be able to recognize a particular event locally and only send the useful information to another system.
It's not necessarily about choosing "edge" or "cloud."
In many real systems, it's going to be a combination of both.
The interesting engineering question is deciding what belongs where.
Then there's the physical world
This is where AIoT starts overlapping with what people are calling Physical AI.
Software normally operates in an environment where things are relatively predictable.
The physical world isn't like that.
Machines wear down.
People move around.
Sensors get dirty.
Objects aren't always where they're supposed to be.
Connectivity disappears.
Conditions change.
Something that worked perfectly yesterday might behave differently tomorrow.
If an AI system is going to interact with the physical world, it has to deal with that uncertainty.
That's a much bigger challenge than simply adding an AI feature to a web application.
And it also explains why things like robotics, computer vision, sensing, digital twins, and industrial automation are becoming part of the same conversation.
They're all different pieces of the same larger problem:
How do we make software understand and respond to what's happening in the real world?
The hardest part may not be the AI model
This is probably the part I find most interesting.
When people talk about an AI-powered industrial system, it's easy to focus on the model.
Which model are you using?
How accurate is it?
How fast is inference?
What framework are you using?
Those questions matter.
But imagine having a great model that can't access reliable data.
Or a great prediction that isn't connected to the workflow where someone actually needs to use it.
Or an intelligent system that can't communicate with the equipment already sitting on the factory floor.
The model can be excellent and the overall product can still fail.
That's why integration is such a big deal.
You might have to connect sensors, databases, cloud infrastructure, edge devices, enterprise software, AI models, and people.
And most industrial environments aren't starting from zero.
There's usually a lot of existing infrastructure that can't simply be thrown away.
Start with the annoying problem
I think this is a better way to approach AI projects in general.
Don't start with:
"We should use AI here."
Start with:
"What is currently wasting time, money, or attention?"
Maybe people spend half their day trying to find equipment.
Maybe maintenance teams only discover problems after something breaks.
Maybe nobody has a clear picture of where inventory actually is.
Maybe there are already hundreds of sensors installed, but the information isn't being used very well.
Those are real problems.
Once you understand the problem, you can figure out whether AI, IoT, automation, or something much simpler is actually needed.
That problem-first approach is also part of how Aperture Venture Studio approaches opportunities around AI, IoT, sensing, and industrial technology.
Where I think this is heading
I don't think the future of IoT is simply going to be "everything gets a sensor."
We're already pretty far down that road.
The more interesting direction is what happens after the data is collected.
A system sees something.
It understands what it might mean.
It decides whether something needs to happen.
A person or machine takes action.
Then the system gets another piece of information about the result.
So instead of thinking about IoT as:
Connect → Collect → Display
we can start thinking about it as:
Sense → Understand → Decide → Act
That doesn't mean humans disappear from the process.
In many situations, the opposite is true. Good systems should make it easier for people to understand what's happening and make better decisions.
For me, that's the part of AIoT worth paying attention to.
Not how many devices we can connect.
Not how much data we can collect.
But whether all those connections actually help us solve something that was difficult to solve before.
Top comments (0)