Whether you’re using machine learning models built into your CRM platform, or you’re looking to have models you’ve already built to provide insights to your users, orchestrating a seamless user experience is critical to deriving the greatest value from your data science investments.
Customer data has always been essential for driving revenue and growing your business, and CRM platforms such as Salesforce have been a core part of many businesses for the past two decades. Making decisions based on your business’ data is crucial to providing the products your customers want, an excellent customer experience, and generally keeping your customers happy. As it is, AI and machine learning depend on smooth workload orchestration. AI and machine learning have become increasingly critical components to any analytics you are performing to support these decisions. But companies are finding that building AI models and getting them into the hands of users in order to realize value can be challenging.
The Challenge with Operationalizing an AI Model
Many companies have access to a team of data scientists, who are able to take existing data that businesses have generated and find insights that can be used to improve customer experiences and drive revenue for the company. The investment, however, in generating these insights and machine learning models can be substantial. And there is considerable risk in investing in machine learning models without a clear path to getting them operationalized with users.
If an AI is not helping users to prioritize their workload, or find insights that help customers find value from a company’s products or services, then even the most predictive or insightful model ultimately is not delivering value. Far too often, these models get built and fine-tuned, and produce statistically meaningful recommendations, but the people who work day-in and day-out with your customers, such as your sales or support staff, have no easy way to engage with these models to find those insights.
Multiple Platforms Confuse Users and Reduce Efficiency
So how can a company which has invested time and resources into gathering and analyzing its data get these results to CRM users? Historically, this involved either building a separate portal for users, with a machine learning model behind it or building out custom point-to-point integrations between a given CRM function and a specific machine learning model.
For someone in a support representative role, having to leave their primary CRM application in order to get insights on behalf of a customer can lead to an interruption in workflow. This is confusing to users and can be costly, especially in a call center environment where minimizing call handle times is a critical metric for controlling cost. It is highly likely that any value the machine learning model insights are providing will be taken away by the reduction in efficiency or loss of customer goodwill.
Similarly, in a Sales CRM scenario, machine learning models are often used to provide insights into things such as how to prioritize workloads (for example, indicating the best customer leads to follow up on or focusing on deals which are most likely to close) or assisting with functions such as forecasting and predicting revenues. Much like a support representative, if a sales user is forced to work in multiple systems in order to get insights or prioritize their day, they are much less likely to adopt the machine learning models companies have built for them.
Building Automated Integration Can Improve User Experience but Also Create Technical Debt
It is common for companies to invest significant IT resources into creating relatively seamless experiences for their users in order to avoid lack of adoption, reduction in efficiency, or general user frustration. This sometimes comes in the form of using AI or machine learning platforms native to the CRM tool being used, and in other cases is solved by building out extensive integrations.
There are many good reasons to leverage built-in or native AI or machine learning platforms provided by a company’s CRM vendor. They are often fast to implement and do not require the long-term expertise of a data science team to build or maintain, freeing the company’s data scientists up for other tasks. Also, because the CRM vendor knows their own platform extensively, they are able to build well-tuned models to fit many standard use cases.
However, there are challenges with this approach. Because the CRM vendor is building solutions that are a fit for all of their customers, they can only pre-build for standard general use cases. If your company has specific needs for its business model or relies heavily on custom CRM data elements and fields, these pre-built models cannot provide specific enough recommendations to provide sufficient value.
If you have a complex business or have customized your CRM, it’s likely you’ll need to leverage custom AI models. If you have to host custom AI models, and they are using custom fields or data elements (which often are not stored in your CRM at all), then the best approach is to use integration to call out to the model when a score or recommendation is required. This often comes with a great deal of complexity and technical debt and can be highly fragile — if a new iteration of a model changes factors even slightly, the CRM platform integration code will need to be updated and tested. At best, this will slow down the pace of innovation, and at worst can lead to user frustration as the models “are always broken.”
Using a Machine Learning Model Broker to Create a Seamless Experience
By implementing an intermediary solution, which we refer to as our Machine Learning Model Broker service, you can get a seamless user experience without the administrative burden. With minimal configuration of your CRM for each AI model, you are able to call out to the Model Broker and receive back an expected response in real-time or near-real-time. Users see insights on the CRM screens they need them on, and the data is retrieved on-demand without the need for complexity within the CRM itself.
So what are the benefits of this approach?
In short, the Model Broker is aware of all the custom (or even CRM Vendor-provided) models that have been created and understands from the callout made by the CRM platform which model is being requested (for example, a gauge of customer sentiment or a product recommendation). The CRM calls into the Model Broker with basic information about what is being requested, and the Model Broker has the rules needed to gather information from various systems (whether the CRM itself, or an ERP, a data lake, or even custom computed model features) as well as access to the model itself, and the parameters needed to call it.
By standardizing the pattern for calling into the Model Broker, the CRM platform maintenance is minimal, and changes to the model (including fields and features needed to make successful predictions) are not a problem. To the CRM platform and its users, the model(s) become “black boxes,” and as a result, companies can realize similar benefits to utilizing a built-in AI or machine learning model without the downsides of lack of flexibility or customization.
In this manner, the data science team can provide finely-tuned, highly specific AI models on the platform of their choice, and CRM users get seamless results and responses without a great deal of technical debt or poor user experience. If the CRM Vendor’s own AI models provide an API to access them, the Model Broker allows for leveraging the built-in models, while at the same time also providing insights from models built on other platforms. This is all easily configured by CRM administrators without custom code and gives companies the best path to maximizing the value of both their CRM investment and their data science investment.
A Better User Experience with Atrium’s Machine Learning Model Broker Service
We’ve built out pre-built and ready-to-deploy packages that connect your Salesforce solution to a Model Broker. Additionally, we’ve built out connectors for the Model Broker to leverage a number of standard platforms and services for hosting machine learning models.
If you’ve invested in building custom models, but have not found a good way to get those models into your CRM user’s workflow without a poor experience or building out a great deal of custom integration, we would be happy to walk you through our approach to solving for this. And if you’ve invested in your CRM vendor’s own AI or machine learning platform, but are struggling to get custom, computed, or external data to work within the framework, our Machine Learning Model Broker service can assist there as well.
Contact us to learn more about our Machine Learning Model Broker service.