Yes, you can now connect an LLM like Claude or ChatGPT to your Salesforce instance. But why would you want to do that?

Here are four use cases that make this a worthwhile endeavor from a revenue operations perspective.

1. Claude can handle your Complex Reporting needs. How many times have you had to create multiple reports in Salesforce, downloaded them as Excel files and then married them up using vlookup formulas? How many times have you said, “I wish I could outsource that?” That’s where Claude comes in. All you need to do is set up a Salesforce MCP Server and connect it to the Claude desktop app. The sobject-reads server will suffice. This gives Claude access to whatever records and objects that you (your Salesforce user) have access to. From there you can run any report you like just by typing in a prompt such as “Show me how all of the sales reps are performing vs. sales target for Q2.” It will give you something like the below screenshot depending on how your org is set up and what additional instructions that you may have given it before. 

If you want to make changes to the format to the data that is included, just tell it!

No more tweaking reports from inside of Salesforce. And you are no longer limited by what charts in Salesforce reports can do. Claude can handle any chart or graph that you can think of.

2. How a sales rep is performing against his or her target has always been something that was tricky to do in Salesforce, especially if you had a complex compensation model. In the above example we merged datasets from two completely separate objects (Targets and Opportunities). But what if you want to add your compensation model into this? Ideally you would have a document that spells out what counts toward a sales target for each type of sale and sales role, and then combine that with your Salesforce opportunity data(or orders or quotes or assets). With Claude connected to Salesforce records and a detailed commission structure document loaded into Claude as a Skill you can do just that.

According to Claude itself, Skills are “curated, environment-specific playbooks that keep Claude’s output accurate and consistent for defined task types” (yes, technically I used AI to generate that sentence). Skills can be any type of document that you would use to maintain anything that Claude would need to know in order to generate accurate and relevant answers to your prompts. This is where you state things like how to calculate ARR, what your fiscal year is, your terms and conditions, color schemes and of course a detailed commission plan.

Once you run your prompt you can readjust as needed.

3. Claude can handle document creation. You probably already know that Claude (like many AI tools) can create PDFs, Word Docs, and of course spreadsheets. When you add Salesforce data to this mix that means you can create a quote document right from within the chat. Just tell it which quote or opportunity record to look at and what you would like on the quote. This is one place where the documents you have in Claude’s’ Skills become extremely important; you want your quote docs to be consistent and follow your business’s rules. You certainly don’t want AI randomly generating Terms and Conditions for you. In the example below I have done much of that upfront work – Claude knows the specifics on our docs to be generated – so we get a reasonably consistent output every time (this is a demo org… you’ll want it to be better than “reasonably” consistent so spend the time prepping).

4. Last but certainly not least, entire PowerPoint decks can be built from your Salesforce Data! This one takes a bit of work and you’ll want a lengthy prompt that works consistently over and over. Once you have that you can just copy and paste it into the chat any time you want to run it again. Or better yet, save it as a Skill with details on each of your reporting metrics spelled out. In the example below we have a quarterly board deck for the board. Once you have all of the details in Claude you can just ask it to create the deck, tell it which quarter, and let Claude do the heavy lifting.

Some (pretty important) caveats:

I have used the word “just” in a few of the examples in this post but this is a gross oversimplification of what goes into making sure this works right every time. I should say “once you have Claude set up properly” then you can “just type in your prompt” and expect a reasonable answer. Remember that LLMs use statistical patterns to guess what the answer to your question is. The less room for guessing that you give them, the better.

To ensure that Claude is going to give you the results that you are looking for:

1. Create documents with all of your business logic and have these set up as Skills in Claude
2. Test heavily. Claude will often give slightly different answers to the exact same question so you want to know where your variances are and then adjust and standardize your instructions.
3. Have clean data. Dirty data is just going to add to the guesswork that Claude has to do with each prompt and thereby decrease the chance of producing useful information.