Every vendor has a model now
If you're a patient access director in 2026, you've been pitched forecasting at least once a quarter. Every patient access vendor, every EHR add-on, every operations consultancy is selling some version of "AI-powered predictive volume forecasting."
Some of these models are real. Some are linear regressions with a marketing budget. The difference matters because if you actually staff against a forecast, a bad model costs money and patients in roughly equal proportion.
Five questions will tell you, before you sign the contract, which kind of model you're being sold.
Question 1: What data was the model trained on?
The right answer is two to three years of your specific facility's data, ideally including the period most relevant to current operations.
The wrong answers come in two flavors. The first is "industry benchmark data" — directionally correct but operationally unusable. The second is "the last six months" — meaning the model hasn't seen a full annual cycle and will miss seasonality entirely.
If the answer is vague, the rest of the questions don't matter.
Question 2: What's the confidence interval, and how does it scale?
Every forecast should come with an uncertainty range. The question isn't whether the model produces one. The question is whether the range is honest.
A model that says "Tuesday will see between 47 and 51 patients" is making a strong claim. A model that says "between 30 and 70" is making a weak one. Both might be calibrated; the second is much less useful for staffing.
Honest models have substantially wider intervals at four weeks out than at one week. Models that show similar confidence at all time horizons are usually overfitting recent history.
A forecast without an honest confidence interval is a guess with a graph. The interval is the model. The point estimate is just the marketing.
Question 3: What's the granularity?
Daily volume forecasts are a baseline. Most staffing decisions are made at the half-day or hourly level.
For most facilities, the honest answer is "daily forecasts are reliable, hourly forecasts have wide intervals." That's fine. The wrong answer is "yes, we forecast hourly with high precision." High precision at the hourly level is suspicious unless the facility has very high volume.
Question 4: How does the model handle exceptions?
Holiday weeks. Snow days. Conference weeks where physicians are out. Marketing campaigns that drive sudden volume.
A good model treats them as known exceptions: it lets you flag them, learns from past instances, and produces forecasts conditional on the exception state. A bad model averages them in with regular days.
Specific question: "If I tell the model that next Wednesday is a snow day, how does the forecast change, and how does the model know what 'snow day' means at this facility?"
Question 5: What's the override and feedback loop?
Forecasting models, when deployed against real operations, are wrong. The question is what happens after the model is wrong.
A good model accepts overrides — the operations team can adjust the forecast based on local knowledge — and learns from the corrections. A bad model produces a forecast and disclaims responsibility.
Ask specifically: "When my staffing manager overrides the forecast, does the model see the override and the actual volume, and does it incorporate both into future predictions?" If the answer is no, the model will be exactly as good in two years as it is on day one.
Why these five
Each question separates a model that was built by people who understand patient access from a model that was built to clear a feature checkbox.
Training data tells you whether the model knows your patterns. Confidence intervals tell you whether it's honest about what it doesn't know. Granularity tells you what decisions it can support. Exception handling tells you whether the people who built it have ever staffed a clinic. The override loop tells you whether the model improves or stagnates.
Don't trust a forecast you can't interrogate. The interrogation is the value.