Salesforce Patents a Way for AI Agents to Pull Live Business Data Before Acting
Salesforce has filed a patent for a method that stops AI agents from acting on stale or missing information by making them look up relevant business data first, every time, before taking any action.
What Salesforce's context-aware AI agents actually do
Ever asked a work chatbot for help and gotten an answer that clearly had no idea what your account, your customer, or your situation actually looked like? That gap, between what an AI says and what's actually true in your company's records, is exactly what Salesforce is trying to close here.
The idea is that when you interact with an AI assistant inside a business app, the system doesn't just guess or pull from a general knowledge base. Instead, it looks at a map of your company's data, figures out which records are relevant to your specific question, fetches those records, and only then decides what to do. The AI is essentially doing its homework before answering.
For a salesperson asking an agent to draft a follow-up email, for example, the system would first pull that customer's deal history, recent activity, and open support tickets before composing anything. The action is grounded in real, current data rather than a generic template.
receiving a data model for the agent to determine one or more data sources to generate context for the user interaction; generating a prompt to determine relevant information for the agent for the interaction based on at least one selected from a group consisting of: an action for the agent based on the data model, and a directive for the agent based on the data model; …
Translation: The system uses a blueprint to figure out what live information sources to check and what questions to ask the language model.
How the agent queries an LLM to fetch the right data
The patent describes a server-side method that kicks in the moment a user triggers an AI agent inside an application. Here is the sequence:
- Data model lookup: The system consults a structured map (a "data model") that tells it which data sources exist and what kinds of objects they contain, like accounts, contacts, or cases.
- Prompt generation: Using the agent's assigned task (an "action") or a standing instruction (a "directive"), the server builds a query and sends it to a large language model (LLM), the same class of AI that powers tools like ChatGPT, to figure out which specific records are actually relevant.
- Contextual data retrieval: Based on what the LLM identifies, the system fetches those records from the data sources. This step is called grounding, meaning the agent's subsequent decisions are anchored to real retrieved data rather than the model's internal assumptions.
- Action determination: With that grounded context in hand, the system decides what the agent should actually do next, whether that's drafting a message, updating a record, or escalating an issue.
The architecture keeps the LLM in a consultative role. It doesn't act directly on your data; it helps the server figure out what to look at. The actual retrieval and action happen on the server side, which keeps the company's data behind its own access controls.
Context data may be retrieved for one or more first objects from the one or more data sources based on a received response from the LLM system based on the transmitted prompt.
Translation: The AI pulls real business records from internal systems after the language model identifies what data is needed.
What this means for AI assistants inside business software
Business AI tools are only useful if they know what's actually going on inside your company. A generic AI that doesn't know your customer's name, purchase history, or open complaints is little better than a search engine. Salesforce's approach ties agent behavior to live CRM records, which is the difference between an assistant that sounds helpful and one that actually is.
For buyers and users of Salesforce products, this is a blueprint for how Salesforce's recent investment in agentic AI could translate into agents that don't require you to paste in background context every time. For competitors building AI assistants on top of business software, this filing draws a line around a specific method of doing it.
That makes this Salesforce's 21st filing we've tracked in our AI agents that act for you watchlist since May, building on earlier applications like the voice checkout assistant and the one that interviews you first.
Claim 1 is written broadly. It covers any server that invokes an agent, uses a data model to pick data sources, sends a prompt to an LLM to identify relevant records, retrieves those records, and uses them to decide an action. That sequence describes a large share of how production AI agents in business software already work, or are being built to work.
The claim does not require a specific type of LLM, a specific kind of data model, or a specific retrieval method. That breadth is a double-edged thing for Salesforce. A broad claim, if granted, could give them leverage over a wide class of AI-agent architectures inside enterprise software. But that same breadth makes it an easy target for prior-art challenges, because retrieval-augmented generation (fetching real data to ground an AI's answers) has been a well-documented technique for several years.
The patent examiner will likely push back hard on novelty. What Salesforce may be banking on is the specific framing around "directives" and "actions" as structured inputs to the prompt-generation step, which is a narrower idea wrapped inside a very wide claim.
There are more where this came from
We read every patent application Big Tech publishes and send you the ones worth knowing. Plain English, free, every week.
The drawings
5 drawing sheets from US 2026/0267696 A1 · click any drawing to enlarge
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →