How many months of clean, structured operational data do we actually have?

Most ML models need at least 12-18 months of consistent, structured data to find reliable patterns. If your history lives in free-text fields and disconnected systems, the investment is in making your data ML-ready before the model matters.

What is the one prediction — callback risk, no-show likelihood, denial probability — that would change how we operate daily?

Pick the single highest-cost operational outcome you currently react to after the fact. ML is most valuable when it turns reactive firefighting into proactive prevention on one well-chosen prediction.

If a vendor says their product uses ML, can they show us exactly what data it learned from and how they measured its accuracy?

Any credible ML vendor can explain their training data, accuracy metrics, and how the model performs on data similar to yours. If they cannot, the ML claim is marketing, not engineering.

How AI Works

Machine learning

By Mark Ziler · Last updated 2026-04-05

Machine Learning is how AI gets good at something without being explicitly programmed for every scenario. Instead of writing rules like "if claim amount is over 10K and provider is out-of-network, flag it," you show the system 50,000 historical claims and it learns which patterns lead to denials. It finds rules you did not know existed. The practical difference for a business: rule-based systems only catch what you anticipated. ML catches what you missed. A mid-market company does not need to hire data scientists to benefit from ML — it needs clean, structured data and a partner who knows how to apply the right model to the right problem.

Go deeper

Your HVAC company tracks 200 variables per service call — equipment age, part installed, technician, time of day, weather, customer history. Your best dispatcher has gut instincts about which jobs will require a callback. ML does the same thing, except it processes all 200 variables across 50,000 historical jobs and finds patterns your dispatcher never could — like the fact that a specific compressor model installed in coastal zip codes fails 3x faster when a particular refrigerant is used.

The trap most companies fall into is thinking ML requires a data science team and a six-figure budget. It does not. What it requires is data that is structured, consistent, and historically deep enough to learn from. If your service history lives in technician notebooks, free-text fields, and three different dispatching systems you have used over the past decade, the ML model has nothing clean to learn from. The investment is not in the model — it is in making your data ML-ready.

Questions to ask

How many months of clean, structured operational data do we actually have?
What is the one prediction — callback risk, no-show likelihood, denial probability — that would change how we operate daily?
If a vendor says their product uses ML, can they show us exactly what data it learned from and how they measured its accuracy?