At this point you have probably heard about the big TDX announcements related to Headless 360 and the Agentforce Experience Layer. Salesforce describes this as giving you the ability to build on Salesforce “any way you want.” This is partly true. You still need the Salesforce platform to handle the backend (business logic and data layers) but your frontend can be in a variety of different places. With Salesforce’s new MCP Servers you could theoretically have Salesforce users who never need to use the actual Salesforce UI ever again.
What is MCP?
Model Context Protocol (MCP) is a new (ish) open-source standard created by Anthropic that allows AI models to connect with other tools. There is what’s called an MCP Server which exposes data, and an MCP Client (like an AI tool) that connects to it. This allows you to connect your AI chat tool (such as Gemini, Claude, or ChatGPT) to anything that supports MCP. This means booking flights on Expedia, planning a hike on AllTrails, editing your trip pictures in Photoshop, and, of course, working in Salesforce can all be done from AI chat tools.
Currently Salesforce has eight different MCP servers that allow connections to various data and metadata. You can think of these like pipelines into your org for AI tools. You can connect them to external tools just like you would with an integration between Hubspot and Salesforce for example, except without all of the data mapping.
So, yes, your sales reps can close opportunities, update accounts, and create tasks all from ChatGPT. Your admins can audit custom fields that aren’t being used. Your managers can use a prompt to run reports on any piece of data that their permissions allow them to access and then have AI spit out a pretty chart or graph to their specifications.
But it’s not magic of course. For things like running reports you have to treat the tool like an intern that has a lot of data but doesn’t know much else. You will need to get good at writing prompts in order to be successful at getting exactly what you want. I had ChatGPT run a report of ARR by account and it told me that there was no ARR field on the Account. Once I explained that ARR was on the Opportunities and that it could just sum by Account it was off to the races, creating ARR reports every which way I could think to tell it. It also came back with suggestions of other related data points that it could analyze. Some were very good suggestions, some were less than useful. The key benefit here was that I did not have to create a single static Salesforce report. I would simply type in my request in natural language.
What will this cost?
Outside of your usual licenses for Salesforce and whatever platform you are using there is no hard cost to using MCP servers. However, all calls the MCP servers do use up API requests so if you go over your limit that will certainly be an added cost. This depends heavily on how often you will be pinging Salesforce from the external service as well how many integrations are already doing the same thing. This is something you certainly want to pilot first and monitor closely.
The bigger issue here is that once you hit your 24-hour API limit Salesforce just shuts you off. This can throttle systems with which you have frequent (and important!) syncs. But Salesforce will certainly be happy to increase your API limits for a price. Note that if you have Enterprise edition you get 100,000 daily API requests plus 1,000 per user. Moving to Unlimited gets you 5,000 per user. Again, there is no harm in piloting this and testing actual usage metrics.
Can we trust it?
I’d be remiss if I did not mention security here. Obviously you do not want to give an LLM the keys to your entire Salesforce org. A couple of things help with this. First when you use MCP and connect to Salesforce you are using your own login credentials, so whatever your user has access to limits what you can access through the AI tool. Second, each MCP Server has its limitations. Some connect only to records with read-only access. Others can give edit or delete access. Some give access to object schema which could be useful for an admin. The recommendation is for you to only set up connections with the subset of things in Salesforce you want to expose to the tool on a user by user basis. In addition, an MCP Server connection uses Oauth so it works like most other integrations with Salesforce. Lastly, most major LLMs like Claude and ChatGPT will give you the option of not having your data used for training purposes.
The jump to using AI to control aspects of your Salesforce org is not a trivial decision. The thing to remember is that agents and LLMs in general are probabilistic, not deterministic. They don’t behave in the exact same way every time, and can potentially reason their way to what Salesforce calls “unexpected outcomes.” So if you are in a heavily regulated industry or absolutely need scripted answers and static outcomes every time then you’ll need to determine if that outweighs the extreme flexibility provided by an AI layer on top of Salesforce.