It is staggering what ChatGPT can create with just a few prompts: Validation Rules, Apex Code, LWC, lyrics for a song, or even blog posts. This article could have been written by ChatGPT – you wouldn’t even realize!
First, what is ChatGPT? It is an AI (Artificial Intelligence) chatbot or LLM (Large Language Model) that uses natural language processing (NLP) to interact with you as though you are talking to a person. The company behind it is OpenAI, which is heavily backed by Microsoft. The results are jaw dropping, but not necessarily 100% accurate – but more on that later.
The scope and potential of ChatGPT is still being explored by the million people who jumped into the pilot during the first five days of its launch. What is clear is that this technology has reached a level of maturity that is becoming disruptive. And OpenAI is not alone; Google has similar technology which it has been unwilling to launch just yet due to concerns about the quality of the results, as well as its reputation.
Let’s explore the potential for the Salesforce ecosystem, including the potential winners and losers, the best use cases, and the considerations when utilizing the results.
Be Careful What You Wish For
The way you interact with ChatGPT is by typing a request into the input field just as you would ask a person. ChatGPT then scans its enormous database of information gleaned from millions of websites, and formulates an answer before presenting it back in a remarkably natural, conversational style. It is not creating new ideas or content, but it is doing an amazing job of synthesizing and representing historical information.
The more accurate you are with the request you enter and the narrower the scope, the better the result. Also, bear in mind that results are not 100% accurate. The implications are that you need to understand the problem you are trying to solve and be able to validate the answer. Think of ChatGPT as a very efficient and capable assistant or consultant.
If the result is not accurate, you need to modify the request and it will correct the answer. It is training you to ask questions in the right way, while you are training it to give better answers.
In terms of a Salesforce example, ChatGPT can create a Validation Rule, Formula, Apex Class, Lightning Web Component (LWC), or a Unit Test for a LWC. It cannot create a flow or any declarative result as the answer presented as text. But it could create the XML that is generated from declarative action, such as creating an object and fields.
ChatGPT works well for a narrow scope that is well specified. For a broader scope the results are less convincing. Do not expect it to be able to write the XML to configure Salesforce for a car dealership, no matter how much detail you give it.
This discussion on Reddit includes a number of examples using ChatGPT for Salesforce.
Winners and Losers
ChatGPT will not be replacing Salesforce Admins anytime soon. In fact, ChatGPT will make them more effective and productive; those with great analysis and admin skills can exploit ChatGPT to accelerate much of the configuration. But they need to learn how to work with ChatGPT to improve the quality of the results.
It’s possible that those at greatest risk are junior developers. ChatGPT has learnt from millions and millions of lines of code. Therefore, it is very effective at developing code. And whilst the code may not be 100% correct, it is a great starting point for an experienced developer. As a junior developer, perhaps you can use ChatGPT to help you improve your coding skills.
If you are a consultant, ChatGPT can help you become more effective. Ask ChatCPT to write a proposal for you – it will do a decent job of writing some generic, core paragraphs, which is a great start. But it is only a start. Be very wary of leaning on it too much.
For example, you could ask it about the process for merging orgs. It will give you an answer that, at first glance, appears very credible. We tried it! But it is not accurate enough. ChatGPT doesn’t understand the nuances.
Don’t be tempted, when asked a question in an area where you have no experience, to ask ChatGPT and simply present ‘the answer’ to a client. This is the fastest way to destroy any trust you have built with your client. ChatGPT cannot give you experience that you have not earned.
Business analysts could use ChatGPT to suggest process steps, but again, these tend to be generic. So, it is useful to prepare before a mapping workshop or validate that you haven’t missed an angle.
The value for architects is less clear. The reason that the path to architect certification is long and arduous is that it requires experience applied in context. This is not something that ChatGPT excels at.
Finally, documentation. It could be marketing, training, or a business case. ChatGPT does a staggeringly good job at creating the first draft of these documents, but it is not a silver bullet.
When you are documenting your org, you need to specify “why” you made the changes, rather than “what” you created. ChatGPT is never going to know that. And by the time you have described the situation to ChatGPT, you might as well have typed it straight into your favorite metadata dictionary (i.e. Elements.cloud)!
Limitations and Considerations
Although LLMs like ChatGPT produce seemingly elegant results, they also have well-known problems. They amplify social biases found in their training data, often denigrating women and people of color.
In addition, results can often be false or misleading in response to queries. Maybe this is because ChatGPT is building its results from millions of websites – some of which include questionable data. This may also be why ChatGPT is so good at writing code, as the training data is made up of millions of lines of code that work.
Users have also found that ChatGPT ‘lies’ about a wide range of issues, from making up historical and biographical data, to justifying false and dangerous claims. This is less relevant to the Salesforce ecosystem where the immediate application of ChatGPT is unlikely to be requesting opinions, but generating specific content, such as a formula.
ChatGPT – ‘Down the Rabbit Hole’
Here is the link to log in to ChatGPT. You may struggle to get a login just now as the demand is so great. But once you do, you will be hooked and may find you have spent hours exploring different scenarios and use cases. Careful there!
Rory Plewes and Walter Bril from Elements.cloud have had early access to and benefitted from ChatGPT. Both have contributed to this post.
We can all work together to share and improve the ‘prompts’ that we put into ChatGPT, which will in turn improve the results we receive. What’s clear is that, with just a few refinements to the words we use (and the order), results are dramatically improved.
Rather than everyone learning in isolation, we have launched a collaborative site called SalesforcePrompts, where anyone can share their prompts and results. People can save their prompts privately (so it’s a great place to store them), but we also encourage users to make them public so we can all learn together.
When you first see ChatGPT generate results, you will find that it is powerful, amazing, and truly jaw dropping. There is clearly huge, untapped potential, and in the coming months we will begin to identify the use cases where it really excels.
What is clear is that ChatGPT cannot replace experience; you cannot ask it to produce something you don’t understand – firstly, because you cannot phrase the question to get the best result, and secondly, because you cannot validate the response.
OpenAPI CEO, Sam Altman, who is clearly managing expectations, recently tweeted:
I think he is underplaying the potential it has to become an incredibly fast, efficient, and cooperative junior developer or admin. We will have to see what the future holds.