At a high level, how does the Org Intelligence agent work?
The agent/model has knowledge about Salesforce and Salesforce best practices. It is augmented with tools, subagents, and databases that it can search.
A “tool” is a deterministic function that the agent can use to get an answer to a specific question (like giving a person a calculator).
A “subagent” is similar to a tool, but it also uses AI under the hood (so it’s like giving a person an expert in a particular area).
Tools/subagents we currently have:
Searching the org’s metadata XML (supports fuzzy matching to find relevant results without precise search terms)
Listing types, items in the org, and counting them (AI isn’t very good at counting).
Getting active flows in the org.
Getting the content for a specific item (this will soon be replaced by a subagent that can answer questions about an item)
Searching Gearset documentation (note: this is currently a static snapshot of Gearset docs, so no new docs will be ingested until we do another manual import).
Getting configured permissions for an item (i.e. what’s configured in a profile/permission set).
Generating links to the item in org intelligence (so that the items listed in the output are clickable and will go to the item view).
What topics is the Org Intelligence agent currently optimised for?
Understanding flow/process automation, e.g.:
If I change <x>, what happens?
Why isn’t the <x> alert/flow firing?
Why does <x> happen when I do <y>?
Permission configuration
Which permission sets or profiles give access to <x>?
Which fields can someone with the profile/permission set <x> see?
Understanding custom objects
What values are in <x> value set?
What happens if I delete <x> from <y> value set?
Find everywhere that field <x> is used.
What is the Org Intelligence agent not optimized for?
Anything about data - The Gearset agent does not interact with the data from a Salesforce org.
Standard vs Custom objects - Questions about standard objects or anything relating to relationships between standard objects. However, this is currently being worked on.
The Salesforce Metadata API does not return full XML for standard objects by default.
Can be weak on e.g. “what currency fields do I have”, as some of them are standard fields, and so not reported.
A query to return a long list of items (e.g. all the dependencies of a complicated item) - This is better served by using the deterministic UI.
“Effective permissions” (i.e. what the permissions actually do vs what’s configured in the profiles/permission sets) - due to a lot of confusion in the Salesforce community around permissions, the training data for the GPT model isn’t always reliable, so it can provide misinformation (e.g. confusion about what permissions are required to run a flow). We’re hoping that we can improve this via tooling, but we’re not there yet.
However in Q1 we released a ‘Permissions Explorer’ feature within the Org Intelligence permissions tab which will help you to answer questions like ‘Why can’t X person see X flow?’
The Org Intelligence agent does not communicate with any of the other agents. So users will not be able to ask questions that link to Observability or UI testing agentic capabilities. It also does not have knowledge of other product’s capabilities (e.g. Code Reviews functionality). This is being improved in Q2 2026.
Comparisons between orgs - Org Intelligence can only answer questions about one org at a time.
How can you get better responses from the agent?
We recommend ensuring prompts are clear, specific, and provide enough context for the model to understand what is being asked. Avoiding ambiguity and being as precise as possible will encourage accurate and relevant responses.
As with all agentic interfaces, iterative refinement of prompts can be necessary to achieve the most useful answers, e.g. starting with an initial prompt, reviewing the response, and then refining the prompt based on the output. It may be necessary to adjust the wording, add more context, or simplify the request as needed to improve the results.
What are some good examples of prompts for the Org Intelligence agent?
If I change Account.Industry, what happens? Here is the user story for this task, along with associated acceptance criteria.
Walk me through what triggers when a Case is set to 'Escalated'.
Explain the relationship between Object1, Object2, Object3, Object4 objects in this org.
Which profiles have edit access to the Custom_Quote__c object?
Show me where the custom field Discount_Tier__c is referenced in flows and validation rules?
What are some bad examples of prompts for the Org Intelligence agent?
Explain the impact of For Salesforce Spring '26 release on this org. What are the potential and/or known downstream impacts?
List all metadata that references the Status__c field
What currency fields do I have across my org?
Why can't John Smith see the Approval flow?
Are there any UI test failures related to this metadata change?