Maximizing Revenue with Machine Learning

There are four machine learning models that every technology, commercial or RevOps leader should develop to drive higher marketing ROI, close rates, cross-sell and NRR – spend potential, spend propensity, ideal customer and churn risk. Yet, according to Alexander Group’s latest Artificial Intelligence in the Go-to-Market Organization research, in 2025, 83% of companies did not have any machine learning models deployed to support go-to-market (GTM) planning for segmentation, territory design or quota setting.

In this article, we’ll provide an overview of the four models and how to begin the planning process regardless of which is selected. The full details of each model, including requirements, approach and GTM applications, will be released in a whitepaper later this year.

Four Foundational GTM Machine-Learning Models

Each of the four models represents a different component of customer “opportunity”. Three approach tenets apply to all models:

1. Includes current customers (expand) and prospects (land)
2. Produces results at a product or product-group level
3. Produces results at the most granular customer or install base definition level available (e.g., child account, site level)

Model 1 – Spend Potential (Serviceable Addressable Market or SAM)

Spend potential models estimate a realistic wallet size ($) for a company’s products and services based on historical spend patterns of customers using a combination of clustering and regression techniques. The automation of the underlying modeling allows for the process to be repeated across an unlimited number of product and industry combinations. Account-level output is primarily used to optimize segmentation and territory design. Product-level outputs can help optimize sales and marketing engagement.

Model 2 – Spend Propensity (Serviceable Obtainable Market or SOM)

A propensity model estimates the likelihood (0-100%) of an account realizing their spend potential for existing customers, or of converting a prospect across each product. When combined with spend potential estimates, companies can create a ‘targeting matrix’ of propensity vs spend to help focus marketing and coverage efforts. Companies can also multiply SAM $’s by Propensity % to calculate SOM, the most realistic short-term opportunity estimate available. SOM is used for more short-term GTM planning activities such as territory and quota planning, as well as insights for day-to-day sales enablement and activity prioritization.

Model 3 – Ideal Customer

Ideal customer models identify key variables of high-spending or high-potential customers and create breakpoints to define cohorts using a decision tree model. Total spend ICP results can be built into lead qualification system while product-specific ICP results can support expansion plays and ABM motions.

Model 4 – Churn Risk

Churn risk models estimate the level of risk of a customer not renewing their subscription or no longer purchasing products or services. This model focuses on current customers; however, prospects should be monitored after onboarding to begin generating risk results. Technique involves analyzing historical churn patterns to identify KPIs, assessing the impact of KPI timing and frequency, and developing customer and product-level risk scores.

Planning For Success

Before building any of these models, technology leaders should prepare to address multiple questions:

Technical Architecture

• What internal company data sources are available to support the model?
• What external data enrichment sources are available to support the model?
• Where will prospect data come from and what is the selection criteria?
• How should accounts, industries and products be defined for modeling purposes?
• What KPIs/variables should be included in the final model?

Business Application

• What exactly will the model be used to support? One or multiple use cases?
• What business logic and qualitative input also need to be factored into the model?
• Which stakeholders are critical to providing design input and output validation?
• How will the model be integrated into systems and leveraged by different functions?

Technical Architecture Planning

One of the critical differences between more intuition-based modeling and machine-learning-driven modeling is the ability to include an unlimited number of account-level variables at the beginning of the process. Rather than assuming a specific set of variables are important and assigning them directional weightings based on leadership alignment, or gut instinct, algorithms conduct variable selection to identify the right factors to include, limiting the inclusion of unhelpful or duplicative data. The presence of variable selection means companies can, and should, load as many relevant data sources as possible into the model. 

While the requirements for each model type vary, there are a number of standard internal data sources that support opportunity modeling:  

• Customer sales with revenue/ARR by product or contract/usage if consumption  
• Pipeline stage history and sales activity  
• Contacts/buyers  
• Marketing engagement (first-party intent) 
• Customer success or service interaction  
• Product usage data  
• Any current segmentation or prioritization insights (e.g., major customer flag) 

Sales and pipeline history are most critical to modeling efforts, with engagement data being a form of internal enrichment if able to integrate. Many companies also leverage modeling efforts as a way to reevaluate data hierarchies and “clean” their CRM prior to model build.  

External data enrichment decisions are also critical to modeling success. Thankfully, in the technology industry, there is an array of data sources that can be included to help inform opportunity models: 

• Firmographics: baseline company data such as industry, number of employees and annual revenue (e.g., Zoominfo, D&B) 
• Technographics: data on technologies in use at companies (e.g., HG Insights, Govspend) 
• Financials: for public companies, standard financial reporting metrics such as QoQ or YoY growth, gross profit, inventory turns, R&D Expenses  
• Workforce: company hiring, role, leveling, pay and training/certification data that may be indicative of buyer potential (e.g., Aura Intelligence)  
• Third-Party Intent: data that tracks marketplace signals, such as social media and hiring, for a company buying interest in a specific product or service (e.g., Intensify, 6sense) 
• Web-scraping or mining: the creation of custom data sets specifically designed to support the insights a company needs (e.g., which VARs offer integration services for my product). Scraping implies the ability to automate the generation of the data vs mining which requires human/manual dataset creation. 

Since tech integrations and associations are often such a critical part of tech industry buying decisions, the availability and granularity of technographic data often add tremendous value to modeling. This is the most critical data source outside of basic firmographics, which is required to help scale opportunity estimates. An example of how technographics can be used: an inventory planning software company has better product fit with 2 ERP systems than others. Technographic enrichment would allow the model to build this relationship into modeling results by looking at account-level ERP usage and intensity. This would help identify better prospects that use ideal ERPs. Financial data enrichment could also be useful – inventory-focused metrics such as ‘Inventory Turns’ could help optimize modeling results further. More examples of how data enrichment can improve modeling will be included in each future model-focused article. 

Industry and product categorization decisions for the model should be output and use case oriented. Industries with materially different buying behaviors should be modeled separately from groups that behave more consistently. For instance, a company can decide to isolate government accounts from the rest of the install base, due to a unique procurement process. Product categorization decisions should be based on utility. While it might be interesting to generate results at a SKU level, too many may lead to modeling challenges, false precision, and difficulty interpreting.

Once all useful data sources are acquired, they need to be integrated into a single source of truth for variable selection. This is an algorithmic process (e.g., AIC stepwise selection) that looks to limit the number of variables to a highly impactful set. This allows companies to statistically validate the importance of certain factors to spend, propensity or churn (or that in combination with other variables do not help improve accuracy). Once algorithmic variable selection is complete, companies should evaluate results and add additional relevant variables. These may include variables that are critical to validation or business or rule inclusion.

Business Application Planning

Business application planning is as critical as technical planning. The precise objective and intended use of model outputs need to be clear. Even a model with as seemingly obvious a use case as churn risk still needs to be evaluated. Is the goal to understand the overall churn risk for the business? To support key account coverage decisions for the next 6 months? To build a rep enablement tool for active churn management across thousands of customers? The answer impacts the detailed modeling approach and assumption decisions throughout the process. Companies should always be developing models with the objective in mind to maintain practicality and avoid becoming fishing expeditions. This is especially the case for spend potential and propensity models which have a wider set of applications. 

To support practical business applications, models must also incorporate business rules and qualitative inputs not reflected in the data. Even with the most advanced machine learning techniques, the underlying model is ultimately limited to patterns identified in the training data. New products, pricing models and/or shifts in buying dynamics should all be evaluated and built-in using assumptions to yield more “accurate” results. If not, the risk of field rejection is high. 

To generate the right business rules for the model, input is needed business leaders and customers. Leadership engagement is critical to model adoption from multiple perspectives.  

• Executive leadership to gather input on key business and market changes that may impact modeling.
• Sales leadership typically helps validate customer and product-level outputs to help refine the model.
• Product leaders to validate categorization and total product opportunity.
• Marketing leaders to provide input on ideal customer and prospect targeting.
• RevOps leaders to validate data usage and integration planning.

Voice of Customer (VoC) information can help further clarify data enrichment needs, validate results and develop business rules. For example, if a relatively new software add-on is showing high opportunity results based on year 1 of spend, but recent VoC feedback is indicative of product issues that will lead to high churn, the model can be forced to reduce potential or propensity for said product. 

The last planning component to consider is technical integration and usage. The integration plan should align with the broader business usage plan. Consider which functional groups need access to which outputs of the model, how they will access them, and how outputs need to be translated, if at all. Translation in this context means taking raw output like potential spend dollars, or propensity percentage, and synthesizing into a more directional and adoptable output. For example, using a 1-10 scoring system for propensity instead of detailed percentage scores. Companies may also create new attributes in systems using modeling insights, such as product-specific expansion target flags. 

Model update and data refresh cycles, by data source, should also be planned in alignment with usage. For example, churn risk models may need to be updated monthly to identify new risks whereas spend potential estimates do not need to be updated as frequently to support re-segmentation. Pilot vs full roll-out decisions should also be evaluated. Applications with more direct connections to action (e.g., using product propensity to drive marketing campaign selection and messaging) necessitate longer testing and refinement periods than directional applications (e.g., including opportunity as a factor in territory design).

Ready to Build?

Now that the options and planning questions are clear, consider which model or models are most critical for your organization, and which you are most prepared to execute from a technical and business application standpoint. Begin to consider supporting investments including data collection and cleanup efforts, new enrichment sources, development time to build and integrate the model and leadership engagement in the project. If all systems are go for launch, your organization should be building one of the four models. In the next article, we’ll go into more detail on how to build spend potential (SAM) models and their unique planning considerations and share real-world examples.