Off-the-shelf AI trained to forecast anything from weather to stock prices turns out to be pretty good at predicting how a company's workflows will change too.
Researchers evaluated time series foundation models - large models pre-trained on generic time series data from unrelated domains - on Process Model Forecasting, the task of predicting how the sequence of steps in a business process shifts over time, based on the "directly-follows" relationships between activities in event logs. Using real-life event logs, they compared these general-purpose models, applied with zero additional training, against traditional statistical baselines and specialized machine learning models built specifically for the task. The foundation models generally produced lower forecasting errors than the purpose-built alternatives. Fine-tuning the foundation models on process-specific data improved accuracy further, but only modestly, and those gains sometimes disappeared on smaller or more complex datasets.
That's notable because forecasting process changes has long been hampered by sparse, uneven data - different processes generate wildly different volumes and patterns of activity, and models trained from scratch on a single log have struggled to generalize across that variety, per the study. It's also another data point in the broader pattern of general-purpose foundation models transferring surprisingly well to niche domains they were never trained on, echoing what's already happened in forecasting and coding tools.
Zero-shot beating custom-built is a low bar worth taking seriously, not a verdict that process forecasting is solved - fine-tuning's inconsistent payoff suggests there's still room for a model built for the job, if anyone bothers to build one well.