Salesforce's Moirai model predicts staffing needs for contact centers by analyzing time-series data. Pilot tests showed 14% higher accuracy avoiding understaffing and nearly 40% higher accuracy preventing overstaffing. The model detects anomalies like holiday-week demand spikes and enables restaffing decisions in 15-minute increments.
Large language models generally are really good at predicting the next word, but when you need to actually predict the future and have a forecast of something, you're thinking about moments in time. That's what's called time series. These events are not necessarily text, and this is exactly where a model like Morae, which is a time series model, comes into play.
The time series element of this is so fundamental. How do you make sure there are enough service reps who can man a phone line or be on a chat and support inbound service requests? And Morae is able to, over time, measure and accurately predict how many reps you will need to actually staff those service requests.
We started working on, on Morae in 2023, and we've been the industry-leading model since then. Actually, we've had over 30 million downloads on Hugging Face, and consistently we've been The number one on several benchmarks for time series.
I am just really amazed at the results right now. You know, we tried it with a few customers, including our own service department here at Salesforce, and we've seen the needle move on 2 critical metrics: 14% greater accuracy on understaffing, and then nearly 40% greater accuracy on overstaffing.
Morae is really differentiated in a few different ways. The first one is one-shot forecasting at scale. That means that the first time that you're actually asking the model, it'll give you that forecast. So it works across different modalities, different types of data, both structured and unstructured. And finally, it's really, really good at anomaly detection. So understanding what is going to change in the future rather than it being the same as it has in the past.
The anomaly detection allows us to then rapidly respond to developing incidents, Restaff in 15-minute increments. We are able to quickly respond to ongoing issues.
As an example, we took a year's worth of data at a contact center, seeing all the different ups and downs and the ebbs and flows of the capacity. And then we'd ask the agent, what are some of the fluctuations that you see during US holidays? And we asked it for the holiday week in December, what is the anomaly that is going to happen there? Because it's not a week like any other week.
It's been really important To have that accuracy, because what customers are then trying to do from a business outcomes perspective is to start with accuracy so they can optimize their spend on payroll.
