The Growing Strength of CRM AI
In a twenty Data/AI career I’ve seen evolution in many ways. One way involves the tools and environment used to build use case solutions. The big change was from on-prem to cloud and, until recently, there’ve been the obvious giants in Azure, AWS, GCP, Databricks and Snowflake pretty much dominating. Similarly, algorithm packages evolved from being mainly open source to including closed source more recently.
However, while all are still dominant providers and likely will be for a long time, I’ve recently seen CRM solutions start to emerge as serious contenders. Undoubtedly Salesforce is the biggest player here. While AI solutions have existed in CRM systems for a while, in reality they haven’t been anywhere near as good as the other environments and tools discussed earlier. For example, when I explored SAP’s machine learning K-means it was clunky and tedious compared to something like Scikit-Learn or Spark ML. Also, Einstein within Salesforce has historically had mixed reviews within the online community. Things have dramatically changed over the last few months though, as Salesforce in particular has matured its overall data ecosystem quite noticeably. Sticking with Salesforce let’s dive into this a bit.
Salesforce and AI
From an AI standpoint the capabilities available are almost exclusively no code, which is naturally designed to simplify development and deployment. The core offerings are grouped into Einstein and, more recently, Agentforce.
Einstein is rather rich in terms of services available. They span most of the palette of data science, specifically prediction models, recommender systems, forecasting and natural language modelling. In truth, plenty of the pre-baked models are for a bunch of specific solutions to common CRM-related use cases, such as establishing likelihood of a lead to convert. In examples like this, the models are pretrained on lots of historical data independent of this organisation’s data, such that the final model only uses variables already deemed powerful. The benefit is it cuts development time; the drawback that other variables may prove influential and other important analytics is cut out. However, Einstein does also have a DIY approach, which is still no code, but offers somewhat more flexibility to build from scratch. It also has ability to connect to external models, e.g., Google’s.
It’s vision deep learning capabilities have been deprecated due to frankly being quite poor. Also there were language model capabilities, but these have more recently been repositioned; this segues nicely into Salesforce’s other (powerful and impressive) weapon in its arsenal: Agentforce.
Agentforce is already on its Version 3. It’s fundamentally powered by a mixture of true experts including GPT and Claude; we all know how great these foundational models are, so imagine a combo, where the best one for a specific task is called on in the background. I’ve already used it on a few projects and have been impressed. What I like is not just its accuracy and intelligence in its completion, but how flexible it is as a result of being ridiculously modular. It’s so easy to build workflows, extend to agentic workflows and importantly customise. I’m seldom a fan of no code over pro-code (I’m old school), but I do like Agentforce. It’s also simple to interact with for a user, appearing as a friendly assistant, and provides pretty sound reasoning. In essence, all data used for Agentforce should be in Salesforce’s customer data platform (CDP) which is exactly a data cloud.
Salesforce Analytics
Then for classic fundamentals of analytics, dashboards, there are really three options: CRM Analytics, Tableau and very recently Tableau Next. Salesforce acquired Tableau a few years back and CRM Analytics came first and enables real-time analytics directly on Salesforce objects. Tableau is an amazing tool which I’ve loved for as long as I’ve been in industry. It offers beautiful dashboards and reports, created simply. What’s more, Salesforce has gone one further and built Tableau Next, which leverages Agentforce as its ‘face’. The user can build dashboards and generate insight using natural language by talking into Agentforce as its copilot assistant. Again, the data comes from the CDP, but any table linkage also needs to be prepared carefully via semantic/data modelling.
Salesforce Governance
Also worth noting that Salesforce now owns Informatica, one of the leaders in relation to data governance monitoring. Where CRM data is particularly sensitive, organisations now have an ideal tool at their disposal to track data lineage, data access, and end-to-end data management with no stone unturned.
Continued Importance of Talent
Of course, this by no means is an excuse not to have strong data experts ensuring the likes of responsible AI usage, intuitive data science, sound quality assurance and careful monitoring and maintenance of tools in production. Salesforce, for one, is a huge juggernaut of a tool, for which there are several roles needed to manage it completely. There is a need for the traditional data roles like data engineers to see
how their responsibilities overlap with these Salesforce roles; it won’t always be a one-to-one mapping but instead a one-to-many. Similarly it makes sense for all data roles working in this specific arena to familiarise with the fundamentals of Salesforce, e.g., terminology like objects and records. Salesforce through its Trailhead platform has excellent training here to bridge the gap.
Final Remarks
In reality, while possible for organisations to build solutions outside CRMs it’s neither easy nor cheap, and adds complexity to integrate those external engines back into the CRM for use. That is not to say though that this should be ruled out – there are several scenarios when it makes sense to have solutions outside, the most obvious being because the solution simply doesn’t exist inside. While there are certainly areas to improve on, such as data modelling, I’m personally excited to embrace and explore what’s available, to help our customers reliant on CRMs build everything organically within one data world, and shape their entire data strategy with this view. I expect things will only get better.
