It all started when Saleforce missed internal sales goals one quarter.

This took everyone by surprise. It was caused by a lack of data transparency combined with human bias. Our sales pipeline lacked vital information that would have highlighted the risk of missing the quarter, while optimism created tunnel vision that led to the incorrect belief that the excellent results of the prior year would continue. In the end, the sales forecast was completely off. This event propelled us to build an AI-driven forecasting tool to take the emotion out of the process and ultimately establish a transparent sales culture to keep this from happening again. Here I’ll share how we took on this challenge, what we learned, and what your organization can take away from the process.

When we decided to build our own forecasting technology, it started with domain understanding. By that, I mean understanding the sales realm, sales pipeline, teams, structures, and processes. We sat down with sales leaders, reps, and all of the constituents to dig into the sales pipeline as a framework. We asked very specific questions such as:

  • Which fields on an opportunity are used by reps and what do they mean?

  • How does an opportunity relate to an actual sales deal — that is, the real interactions?

  • How is an opportunity created?

  • What are the stages in an opportunity? What do they represent? How are they updated?

With these answers, we were armed with enough domain knowledge to understand the problem on a deeper level and, by extension, to develop a product that would meet the needs of our users.

To refine our understanding of the problem, we asked sales leaders and the sales strategy team, “When you’re faced with this problem and there’s a shortfall in your business, what questions come to mind?” Their answers guided how we shaped the forecasting tool as a product.

We learned from the feedback that if there’s a shortfall, someone in the management chain is associated with it. Who is that? Is someone further down the sales team responsible for it? Once the problem is localized to an individual within the business, the next step is discovering what is driving the shortfall. A data-driven forecasting product needs to help that individual understand how to fix the problem.

With a strong understanding of these product-shaping questions, we moved on to start designing the actual product. This meant designing a data pipeline to source and process the input data, designing models to transform that data into a forecast, and finally designing a user interface (UI) to present the model outputs as actionable intelligence to the end user.

Developing a data pipeline to feed the forecasting models required a deep understanding of how the sales pipeline data is stored in the physical database. Specifically, we needed to understand what format the opportunity data was stored in and how the history of deal execution is captured and represented in the actual tables. It was also essential to understand the processes used to capture and store the data — in particular, when and how often these tables are updated, what kinds of errors occur when capturing this data, and how we could detect these errors. Finally, we needed to write code to extract and transform the data into a format suitable for feeding into the forecasting models.

A key insight that enabled us to build accurate forecasting models is that the results of a quarter are driven by patterns associated with cohorts of deals. The nature of those cohorts is different from business to business. However, they drive quarterly results in every business. For example, Salesforce has two deal cohorts characteristic of a subscription-based business. The first cohort consists of deals executed with new customers. This cohort is characterized by a deal cycle of several months with this time being used to deeply understand the customer’s problems and discover the right solutions in our product offering. The other cohort has a short deal cycle of a few days. These are deals where existing customers are simply buying additional licenses as their businesses grow.

Our models detect the seasonal and growth patterns in the number of bookings generated by those cohorts and use it to predict the quarterly results. The best part is that the models automatically construct these cohorts by eliciting patterns directly from the data. This completely eliminates the pain of figuring them out manually.

For the UI, we wanted to ensure the questions we asked earlier in the product-shaping phase were answered in an intuitive way. This required picking data visualizations that would convey answers at a glance and navigation aids to quickly guide users to the right information. Terminology and semantics proved to be essential to the design of the UI. Very specific terms might mean one thing to a layperson, but would convey something completely different to a sales leader or to different business units. We had to be careful to use the right terminology on the UI to make it intuitive and useful for reps and sales leaders.

Once we had a minimal viable product, we worked with the sales strategy team to release it. There was a constant flow of direct feedback into the final phase — and that is ongoing to this day. The product is always getting better.

There were big roadblocks to overcome in the process of building this forecasting tool. In the end, we were able to really solve the problem for us — and now for our customers.

Sales leaders are busy people and need fast, actionable intelligence about the weak points in their revenue or their bookings. Our tool automates the process of creating this intelligence. It achieves this by using AI to distill the essence of what’s driving the revenue in a sales pipeline, abstracted away from the complexity of a customer’s business and CRM implementation. This layer of artificial intelligence continues to learn about and adapt with the business to keep the product improving over time. Our goal is to minimize the effort on your part, especially after we went through this entire development process ourselves — whether it’s configuration, domain understanding, or extracting patterns from the sales data.

If organizations decide to build their own forecasting tool, it’s important to keep in mind some of the lessons we learned and the tremendous amount of resources and time that go into getting it right (our forecasting product took over a year to develop). In the end, we were able to build and deliver a forecasting tool that is now used throughout the organization and has proved to be an incredible way to solve the core problem of identifying the holes in the sales pipeline.

Sales leaders are busy people and need fast, actionable intelligence about the weak points in their revenue or their bookings.”

Robin Glinton | Vice President of Data Science Applications, Salesforce