# Why "the model isn't accurate enough" is rarely the real problem

_When an AI feature gives a wrong answer, the reflex is to blame the model and go shopping for a better one. Most of the time the model was doing its job on inputs that were already wrong or missing context it never had._

Two anonymized stories, one from a chronic-care health platform and one from an engineering leader who scaled large consumer systems, on why bad data and missing context masquerade as model accuracy problems, and what to check before you swap models.

# Why "the model isn't accurate enough" is rarely the real problem

Here is the pattern I see most. An AI feature produces a wrong answer, someone declares the model not accurate enough, and the team starts evaluating a bigger one. Weeks later the new model gives a different wrong answer, because the problem was never the model.

## The wrong answer was correct given the input

A founder who runs a chronic-care health platform told me a story that has stayed with me. His system scans patient records for risk signals, and one of them flags likely diabetes. A woman kept asking a question his team came to know well: why do you think I'm diabetic? They looked at her record. She was listed as taking diabetes medication, which is a strong indicator. Except she was not. Her husband was. The pharmacy had changed, and the new pharmacy assigned his prescription to her identity number. The old pharmacy had it right.

The model was not wrong. It read "patient takes diabetes medication" and concluded diabetes, which is exactly what you would want it to do. The failure lived one layer up, in a data pipeline that attached the wrong person to the right prescription. No amount of model shopping fixes that. He said it plainly: with hundreds of thousands or millions of records and events happening every day, you have to find ways to catch these wrong assignments, because the model faithfully believes whatever the data tells it.

This is why "the model isn't accurate enough" is usually a misdiagnosis. Around 95% of pilots stall on integration and trust rather than model quality. The output looked like a model error and was really a data-lineage error wearing a costume.

## Missing context looks the same as low accuracy

An engineering leader who scaled systems at several large consumer platforms, mapping, gaming, and dating among them, gave me the other version of this. He was building a health sensor product and described a walk with a friend whose smartwatch was recording the activity. They got on a bus. The watch, still reading a slow speed, kept insisting they were walking, then kept asking if the workout was finished when they stood at the stop. His point was that the device tried to predict in the moment without the context to do it well. Collect the data and interpret it afterward, he argued, and you get less intrusive and more accurate results.

Nobody would call that a model accuracy problem, and yet it produces exactly the symptom teams blame on model accuracy: confident, wrong output. The fix was not a smarter classifier. It was giving the system the context, in this case a second sensor and a moment to look back, before it decided.

## What to check before you blame the model

When an AI feature is wrong, work upstream before you swap models:

- **Trace the input.** Is the data the model saw actually correct and attached to the right entity?
- **Check for missing context** the model would have needed to get it right.
- **Look for a pattern rule** the way his team eventually required more than one signal before flagging diabetes, not a single mention.
- **Reproduce the exact failing case** end to end before you conclude the model is the weak link.

## How we approach it at Density Labs

In the AI Readiness Assessment, our $2,500 front door, when a feature is giving bad answers we map the path an answer travels before we touch the model. Where the data comes from, what gets joined to what, what context the model is and is not handed. In most engagements the "accuracy problem" turns out to be a broken join, a stale record, or missing context, and a better model would have inherited the same fault at a higher cost.

Buy a bigger model and you often buy a more expensive way to be confidently wrong. Fix the input first.