Prompts in Practice: Using AI and Simple Math to Leverage Other Tech and Better Predict Election Resource Needs

A Practical AI Use Case from Arlington County, Virginia, General Registrar/Director of Elections Gretchen Reinemeyer

I’m catching up on watching the 2025 U.S. Election Assistance Commission Data Summit. One of the quotes colleagues highlighted was that “AI is about guessing – elections are about knowing.”

But I don’t think that’s true. Election administration is often about guessing. What will the legislature do next session? What parts of VVSG 2.0 will my state adopt? How much money will I need for an election that is 18 months away? How many ballots should I order? I’m left asking, “Can AI help me guess better?”

During the June 2025 election, I ran a test: How accurately could I predict turnout based on the first day of early voting? Turns out, pretty accurately. Or at least that’s what my AI assistant – Employee #10 as our nine-person office likes to call it – tells me.

Using Employee #10 as a guide, I was able to run a Historical Proportional Forecast based on five primaries worth of early voting data to predict turnout each day of early voting and at each precinct on Election Day.

My predictions got an R-squared of .990001. I have no idea what this means. My degrees are in anthropology, but Employee #10 tells me this was a very good analysis. Better AI-assisted guessing can help improve staffing and equipment allocations.

How I Did It

Virginia has 45 days of in-person no-excuse absentee voting (also known as Early Voting). The law allowing no excuse early voting passed in 2020. When I first asked Employee #10 to help me predict turnout, I needed 50 datasets to build a predictive model. Arlington County predictably has a Democratic Primary every spring, so I had five datasets. I suspect large data modeling will only ever be of limited use to a local election shop because not even Georgia has 50 elections a year.

AI usage restrictions in my jurisdiction prevented me from loading data into Employee #10, but I could ask for help. Through prompts, Employee #10 recommended how to structure the data using Excel and recommended an analysis I could run.

I first looked at trend lines to see if there was consistency between elections. The 2020 primaries stood out – data from those elections, administered during the pandemic, did not follow a pattern. I excluded these elections from the dataset to avoid skewing the predictions.

I then calculated the percentage of the total early voters who voted each day of early voting. I averaged percentages for each day (e.g. on the first day of early voting on average 1.37% of early voters cast ballots). Using actual turnout from the first day, assuming that represented 1.37% of early voters I could expect, I forecasted that 4,308 would vote early. I then forecasted the number of voters might come each day using each day’s average.

On day 3, my model was off by two. On Day 4, it was off by five. Strong start. I next decided to use it to forecast the number of Election Day voters at each precinct using a similar method of calculating averages of percentages. Two weeks into early voting I adjusted my ballot order because we were tracking higher than initial predictions based on the type of election and offices on the ballot. The model was right: Turnout was higher than anyone in the state predicted.

With the election over, certified, and audited, I’ve turned back to Employee #10 to tell me if my model was any good. With its help, it recommended calculating R-squared to measure how well actual turnout was reflected by the model. It provided steps to calculate R-squared in Excel (something I didn’t know Excel could do). Employee #10 helped me interpret this number, told me that I was very strongly able to predict turnout, and it gave me five stars.

I realize this isn’t groundbreaking research. Political scientists and analysts have been running these analyses for years. I suspect there are a few reading this who will politely comment on why this is analysis is bad or how it could be improved.

AI can’t make me an expert, but it can help me know I need to find a temp to cover lunches starting three Fridays out from an election as opposed the two I had planned.

Will this model hold for the fall? I’m not sure, but you can follow the election with me by checking out our dashboard where we publish early voting turnout daily.

P.S. If you’re wondering when the best time to vote early to avoid a line is, I recommend a rainy Wednesday afternoon. If you enjoy standing in line with all of your neighbors, we'll see you the last Friday around 3 p.m.

Note: This post originally appeared on LinkedIn. It has been edited slightly for tone and to update time references.

Prompts in Practice

Use Case/Purpose

Turnout forecasting for staffing, ballot ordering, and equipment allocation during early voting and Election Day.

Source

Election official (Arlington County, Virginia, General Registrar/Director of Elections Gretchen Reinemeyer)

AI Tool Used

ChatGPT

Example Prompt

I want to use Excel to build a simple model that predicts early voting turnout in Virginia. We have 45 days of in-person, no-excuse absentee voting. I only have data from a handful of past elections (about five datasets), and I cannot upload the data here because of restrictions.

1. Tell me how to structure the data in Excel to make it usable for forecasting.

2. Recommend what kind of analysis I can run with a limited dataset.

3. Explain how to identify anomalies (for example, pandemic-era elections that don’t follow normal trends) and how to decide whether to exclude them.

4. Walk me through how to calculate daily turnout percentages and how to use the first day’s turnout to project total early voting turnout.

5. Keep all explanations step-by-step and assume I don’t know advanced statistics, but do know how to enter formulas and use basic charts in Excel.

AI Use Safety Signal

🟡 Yellow (Caution)

This use case supports productivity by assisting with data structuring and forecasting methods. While outputs were highly accurate in this instance, predictions carry real-world consequences (e.g. ballot orders, staffing). Always validate models, confirm assumptions, and review results with subject-matter experts before acting.