For most of my career in technology, adding more capability was almost automatically considered progress.
A camera became better when it recognized faces. A photo app became better when it organized pictures automatically. Software became better when it knew more about its users and could anticipate what they wanted.
AI has accelerated that thinking. Today, almost every product team is asking the same question: where can we add AI?
After running Yogile, a safe photo storage for more than 15 years, I’ve started asking a different one:
Where should we deliberately not use it?
That question became surprisingly important to us because of something our customers started asking.
“Do you scan my photos with AI?”
At first, I found the question interesting because we had never really thought of the answer as a feature.
Yogile is a private photo storage and sharing service. We don't scan customers' private photo libraries with AI, and we don't use their photos to train AI models.
There wasn't a marketing strategy behind that decision. We simply didn't need to understand what was inside someone's photographs to provide the service they were paying us for.
But when enough people independently start asking about something, it's worth paying attention.
We realized that “no AI” was becoming part of the product.
The feature nobody asked us to build
This creates an interesting reversal in product development.
Normally, a feature is something you add.
You add automatic tagging. You add recommendations. You add facial recognition. You add an AI assistant.
But sometimes a product decision is valuable precisely because of what you choose not to add.
Imagine someone uploads 20 years of family photographs.
From a technical perspective, that's an incredibly interesting dataset. Modern AI can potentially identify people, objects, places, activities and patterns across those images. A product team could probably imagine dozens of features built on top of that information.
But there's another way to look at the same dataset.
Those aren't simply images waiting to become smarter.
They're someone's life.
Children growing up. Parents getting older. Homes they've lived in. Relationships. Weddings. Holidays. Birthdays. Ordinary moments that were never created with the expectation that software would try to understand them.
At Yogile, we need to store those files reliably and give people an easy way to privately share them.
We don't need to understand them.
That distinction has become part of how I think about product design.
Capability isn't the same as value
AI has dramatically reduced the technical barrier to adding intelligence to products.
But that can create a dangerous assumption: if something is technically possible and moderately useful, it should become a feature.
I don't think that's a good enough test anymore.
Every AI feature has a second side: what does the system need access to in order to provide it?
If an AI can automatically categorize someone's family photos, the benefit is convenience. The requirement is that software analyzes what's in those photos.
A user might happily make that trade.
Another user might not.
The important thing is recognizing that there is a trade at all.
As product builders, we should be able to explain not only what an AI feature does, but what it needs to know in order to do it.
Restraint can become differentiation
Technology companies traditionally differentiate by doing more.
I think AI may create categories where companies can also differentiate by doing less.
A writing application might promise not to use private documents for model training.
A children's product might deliberately avoid behavioral profiling.
A health application might minimize what it infers beyond what is necessary to provide its core function.
A photo service can decide that storing someone's memories doesn't require understanding them.
This isn't anti-AI. We use AI internally at Yogile, including in software development. It can make us faster and help us build better software.
The boundary is about purpose.
AI helping us write or understand code is very different from AI analyzing a customer's private family archive.
That leads to a rule I've found increasingly useful:
Don't ask only whether AI can improve the feature. Ask what AI needs to know about the user to improve it.
Then decide whether the improvement is worth that access.
Customers may tell you where the boundary moved
One of the more useful lessons from this experience is that product positioning doesn't always begin in a strategy meeting.
Sometimes customers tell you that something has changed simply by starting to ask a new question.
Years ago, “Does AI scan my photos?” wasn't a common question for us.
Now it comes up.
That doesn't mean everyone suddenly rejects AI. Far from it. Many people actively want AI-powered organization, editing and search.
But it does mean that AI itself is becoming part of the purchasing decision.
For some customers, having AI is a benefit.
For others, knowing where AI stops is a benefit.
That second group is interesting because traditional product thinking can easily overlook them. You can't measure demand for a feature that consists of something not happening by looking at feature usage.
You have to listen to what customers are worried about, what they ask before buying, and what they need reassurance about.
The next generation of products needs boundaries
We're entering a period where adding AI will become technically easy enough that “powered by AI” stops being particularly distinctive.
Almost everything will be capable of analyzing, generating, predicting and personalizing.
When capability becomes abundant, restraint can become meaningful.
The interesting product question may no longer be:
“What can our AI do?”
It may become:
“What have we deliberately decided our AI will never do?”
For Yogile, one of those boundaries is straightforward. People give us their personal photos because they want us to store and privately share them.
We don't need a second purpose for those memories.
We don't need to turn them into training material.
We don't need to understand who is in them.
Sometimes the product is better because the technology knows less.
And in an industry racing to make everything intelligent, I think being able to say exactly where the intelligence stops may become a feature in its own right.