Rethinking Employee Satisfaction and Productivity Measurement with LLMs

Measuring the impact of the workplace on employee satisfaction and productivity has been a persistent challenge for corporate real estate for as long as I can remember. Answering this question has usually been limited to surveys. And everyone hates surveys.

Over the years, I've reviewed methods used by a number of companies trying to understand the impact of workplace on employee productivity (or satisfaction, or engagement...). I've run statistical tests on large datasets (n>25,000) from a single company, reviewed the methodologies of survey providers, and tested new surveys with questions informed by frameworks such as Gallup's Q12 and Harvard's research on healthy workplaces. So far, I'm not aware of any survey that accounts for more than 52% of the variation in employee satisfaction with their work environment. That means we're flying blind half the time!

It's clear that a new way is needed. This next part might sound like I'm jumping on the LLM bandwagon. But I'll try to stick with specifics. After I read this paper on how researchers created a prediction model for the presidential election using such tools (LLM Generated Distribution-Based Prediction of US Electoral Results) [1], I got intrigued that perhaps we could learn something that could break the log jam in measuring and improving the impact of workplace on satisfaction/engagement, even productivity.

Summary of the Election Prediction Approach

To give you a gist of what's in the paper, here are some key concepts:

The LLM creates synthetic personas based on demographic data (e.g., age, gender, education, location) and simulates how these personas would respond to survey questions or vote in elections. This is done by embedding demographic and historical trends into structured prompts.

Two main prompt types are used:

1. Role-Play Prompts: Simulate voter behavior purely based on individual demographic characteristics.
2. Structural Prompts: Combine individual data with contextual factors, such as a state's historical voting trends, to improve predictive accuracy.

Then the LLM/Generative AI model uses the personas and population demographics to create data representing simulated votes.

By aggregating the simulated votes across states and comparing them to historical data, the model demonstrates its ability to replicate past election outcomes and forecast future trends, with adjustments made for demographic shifts and state-level political leanings.

Applying the Election Prediction model to workplace satisfaction

This method of combining rich contextual understanding with predictive modeling could be directly applied to workplace satisfaction and productivity. In this adaptation, here are some of the tasks that would need to be completed:

1. Employee Personas: Develop detailed personas based on job roles, demographics, and workplace preferences. (Been there, done that! But this approach puts those personas to great use.)
2. Baseline Surveys: Collect initial data on how personas value specific amenities or workplace conditions (e.g., ergonomic furniture, collaboration spaces).
3. Predictive Modeling: Use the LLM to simulate satisfaction or productivity based on workplace resources, layout, or policies.

It's not quite as simple as this might make it sound. We'd also have to look at incorporating data on how interactions between personas—such as cross-functional collaboration—amplify the value of shared amenities, exceeding individual valuations.

But the use of LLMs and generative AI basically takes the place of what we might otherwise use a statistics-driven regression analysis..

I also think if the personas include details about what productivity looks like, the model can be extended beyond satisfaction to productivity!

Possible Advantages of the LLM Approach

I think there would be many advantages of using an LLM + Generative AI approach to measuring, modeling, and improving employee satisfaction and even productivity.

1. Dynamic Predictions: Unlike static regression models, LLMs adapt to new data, predicting satisfaction and productivity across evolving workplace scenarios.
2. Nuanced Insights: LLMs can analyze both structured and unstructured data (e.g., survey responses, open-text feedback), uncovering subtle relationships and emergent patterns.
3. Avoid Surveys!!: By simulating satisfaction and productivity impacts, LLMs reduce the need for frequent large-scale surveys, enabling continuous improvement.

Corporate real estate managers can use this approach to optimize their investments, identifying which amenities and policies deliver the highest value. Beyond improving satisfaction, this framework enables workplace strategies to actively enhance productivity, offering a transformative way to manage real estate as a strategic business asset.

[1] Bradshaw, Caleb & Miller, Caelen & Warnick, Sean. (2024). LLM Generated Distribution-Based Prediction of US Electoral Results, Part I. 10.48550/arXiv.2411.03486.

[2] For additional information about birds carrying coconuts, flesh wounds, and bringing out yer dead, watch Monty Python and the Holy Grail. TL;DR read this article: https://www.pastemagazine.com/movies/monty-python/monty-python-and-the-holy-grail-best-quotes