A lot of AI automation is predicated on clean digital input
An email comes in, a model classifies it, a response is sent, a database is updated
Real world operations aren't clean.
A maintenance request might have a picture, a short description, equipment ID, location, years of service history. A repair request might need to reference parts, previous work orders, technician availability, and operating conditions.
The engineering challenge is not just adding an AI model.
It's connecting all of those pieces to produce a useful outcome.
The input is usually messy
Let's say we have a maintenance request like so
"Machine 14 is making a strange noise."
That sentence on its own probably doesn't tell us what we need to know to make a maintenance decision.
A more useful system would look something like this
Request
↓
Asset ID
↓
Equipment History
↓
Sensor Data
↓
Previous Work Orders
↓
AI Analysis
↓
Suggested Next Step
↓
Human Review
We're embedding the AI as a cog in a bigger information processing workflow.
Context is more valuable than a signal
Let's say we have a connected machine that's reporting more vibration than usual
There are a number of factors that could explain that.
Maybe the machine is under a different load, maybe a bearing is wearing out, or perhaps the sensor is just noisy.
Vibration on its own isn't necessarily a useful signal.
A better system would look at vibration in the context of temperature, operating conditions, maintenance history, run time, or other available sensors.
This is about data integration.
The challenge is not finding a new model, but rather getting the right information into the model at the right time.
AI doesn't need to own the workflow
There's another temptation to build systems where AI makes everything possible.
That can lead to unnecessarily complex systems.
A better approach is to think about the system in layers and let each layer do one job
IoT / Application Data
↓
Data Processing
↓
AI / Rules / Analytics
↓
Recommendation
↓
Human Review
↓
Business System
↓
Physical Action
For lower risk decisions, it's appropriate to make them fully automatic.
For higher risk decisions, the AI can make a recommendation but an authorized person has to take responsibility for the final decision.
Observability matters
When you embed an AI in an operational workflow, you need to be able to reason about what happened.
You need to ask yourself questions like
did the model produce an unexpected prediction?
was the prediction reviewed?
was the suggested action taken?
You can answer these questions by logging fields like
request_id
asset_id
timestamp
model_version
input_reference
prediction
confidence
review_status
final_action
It makes subsequent debugging easier and gives you a feedback loop to improve your models.
A prediction can be reviewed and either approved or rejected. That outcome can be used as a signal to improve future predictions.
It's also worth emphasizing the feedback loop
Prediction → Review → Outcome → Feedback → Improvement
Start with one workflow
If you're going to connect everything into a big graph, it's tempting to try and do it all at once
But you'll get much further by starting small and picking one workflow to optimize.
Here are some examples
- Classify maintenance requests
- Prioritize work orders
- Identify unusual equipment behaviors
- Extract information from inspection reports
- Match repair requests to asset history
Measure the current process, add the AI, and compare the two. What's faster? What's safer? Are there fewer steps? Are there new insights available?
Did you reduce response times or unnecessary work? Were the predictions helpful? Are people using the suggestions?
These are more interesting questions than "how accurate was the model".
The bigger engineering challenge is connecting everything to the model
AI becomes interesting when it can operate in the real world.
Sensors, machines, images, locations, maintenance, inventory, and human decisions all become part of the same workflow.
This is the domain space explored by the companies at Aperture Venture Studio, which is focused on AI and IoT companies working on real world operational software.
The lesson is simple: don't build an AI model and then look for problems to solve. Start with the problem and the workflow, identify what data you have and what you're missing, and then think about where AI can add value.
Top comments (0)