IBM spent billions on Watson Health. The idea looked strong – a neural network that helps oncologists choose treatments. The technology was there, the data was there, the brand was there. The result fell far short of expectations. In 2022, the division was sold off.
The problem wasn’t just the model. Watson tried to solve a problem that no one had clearly defined. What decisions should it produce? Based on what data? How does it fit into a doctor’s actual workflow? These questions were asked too late.
Now a different example. Every day your email filters out spam. You don’t even think about how it works. But behind it is a neural network that processes billions of messages and learns from your actions. You don’t see AI. You see a clean inbox.
The difference between these two stories isn’t in the power of the technology. It’s in how precisely the problem was defined before development began.
Where neural networks prove their worth
The first situation is when people inevitably start making mistakes due to volume or monotony. Quality control on a production line, content moderation, document verification. Anywhere a specialist processes hundreds of items per shift and loses focus by the end of the day. Not because they’re a bad specialist, but because attention has a physical limit. A neural network doesn’t have that limit.
The second – when the task is unsolvable manually. Detecting fraud in a stream of bank transactions, personalization for millions of users, finding anomalies in large datasets. No analyst can sift through a million records and spot a pattern in a reasonable amount of time. A neural network can.
If a task doesn’t fall into either of these categories, implementation most likely won’t deliver results. You can spend months of development and end up with a “smart” feature that performs worse than a simple one. Like Watson – the technology was there, but the value wasn’t.
Our approach
At Elpixel, before every implementation we ask one question – can this task be solved without a neural network?
If it can – we solve it in a simpler and faster way. If it can’t – we build it so the impact shows up in the product, not on a slide. Fewer errors, faster processing, more accurate decisions, less manual work.
Nobody calls the spam filter in your email an “AI product.” But that’s exactly what good implementation looks like – the technology is in its right place, solving a specific problem, and the user gets the result without even knowing there’s a neural network behind it.
