Decoding the future of Salesforce: Insights from a test drive with ChatGPT

Salesforce continues to make its platform more accessible and customizable for individuals with less technical expertise. Their latest tool, Einstein GPT, promises to further strengthen access for this set of users. Based on OpenAI's natural language processing technology, Einstein GPT has the capability to turn English language into usable code. This ability, though, brings forward a key question: Will generative AI be just a novelty—or could this tool truly help Salesforce achieve its aspiration of a no-code/low-code future?

Since Einstein GPT is currently in a closed beta test, I decided to run my test of this question using OpenAI’s GPT-4, the tech underpinning Einstein GPT. The mission was simple: Determine if generative AI could tackle a routine task like creating a user interface (UI) component. Specifically, I tested the creation of a field-driven banner that displays a message on a Salesforce page layout.

I began by giving GPT-4 the following prompt:

“Build a lightning web component for a lightning record page that displays a warning message based on a flag on the record. While configuring the component, admins should be able to specify the field, message, and color. If the field evaluates as true, the corresponding message should appear on the component with the designated background color.”

After pressing Generate, GPT-4 churned out the necessary code in about 30 seconds. It produced an HTML, Javascript, and js-meta file—which, to my delight, was deployable on the first attempt to Salesforce. A quick test of the component did reveal one minor bug, but with a minor course correction, GPT-4 had a working version in no time.

Next, I decided to push the boundaries a bit more, asking it to modify the component to accommodate a higher degree of complexity: 

“Make this component more dynamic. Allow an admin to put in a list of fields, messages, and colors to determine more dynamically what to display in the banner.”

GPT-4 managed this request smoothly. This time, however, a small quirk with how the getRecord function works caused the solution to break. But the issue was easily revolved with my suggestion to create its own Apex Class to retrieve field values.

The results were quite impressive. GPT-4 managed to generate code that was about 95% complete with just a simple prompt. It’s worth noting that the 5% gap means that this tool is not immediately helpful for inexperienced developers, as it requires the know-how to fill that technical gap. But in a future with even more advanced generative AI, could Einstein GPT overcome that gap and push Salesforce even closer to being a no-code/AI code-generated platform?

It's certainly plausible. My experiment, while basic, suggests the potential impact of generative AI for Salesforce admins. It's conceivable that in a few years, we'll see a new UI-Builder tool from Salesforce that allows admins to type a brief into Einstein GPT—which then adds a custom UI component based on those specs to the page. It feels like we're just a bug or two away from this right now.

However, even with generative AI, I don’t think that Salesforce will become a true no-code platform—at least not in the next few years. The complex, unique needs of each business mean that AI won't have all the answers. While developers may be able to use Einstein GPT to generate starter code for basic problems (and customize as needed), for more complex challenges, there are limits to how Einstein GPT can help. 

Generative AI is simply not able to grasp the bigger picture. One cause of this is the token limit. GPT-4, one of the most sophisticated models today, can only process up to about 32,000 tokens (~25,000 words) in a conversation, according to OpenAI. Taking Apex code, flows, UI components, and Custom Object/Fields into account, a mature Salesforce org can easily surpass that limit. If users try to share all their Salesforce customizations in a single request, GPT-4 will error out due to the request being too large. While breaking the information into multiple requests seems like a solution, once the total amount of tokens/words in the conversation reaches the limit, GPT-4 starts forgetting everything that came before, leading to jumbled, unrelated results.

Not long after finishing my experiment, a new technology that could help relieve the token limit issue gained popularity. This technology is called a vector database, and it allows for supplemental context to be easily consumed by generative AI without counting against the token limit. While this could bypass the hard token limit, conceptually we are still left with the same issue: AI is not able to grasp the bigger picture.

Developers, architects, and the code they write play a nonreplaceable role in Salesforce development—whether it’s understanding how external systems interact with Salesforce or how new, complex features should interact with each other. The company and industry context that developers and architects provide—as well as their ability to comprehend and enhance large systems—means that access to custom code remains an essential element. This bridges the gap between technology capabilities and practical, effective solutions tailored to the unique needs of each organization.

Even without knowing the bigger picture, Einstein GPT still promises to revolutionize what admins and developers can achieve with Salesforce for day-to-day tasks. Armed with straightforward prompts, builders will be empowered to craft intuitive user experiences that effectively drive business value. Generative AI will help everyone, technical and nontechnical alike, get their work done more efficiently and free up brain space for more complex, creative problem-solving. Generative AI prompting will undoubtedly advance the low-code future of the platform.