Salesforce Patents a Self-Teaching System for AI That Handles Back-and-Forth Conversations
Training an AI agent to handle a real customer conversation, the kind where the user changes their mind or asks follow-up questions, requires enormous amounts of labeled data. Salesforce's new patent describes a way to generate that data automatically, using AI to teach AI.
How Salesforce trains AI agents to hold longer conversations
Today, building an AI assistant that can hold a real back-and-forth conversation is expensive: someone has to write out thousands of example conversations by hand, label each step, and check that the AI did the right thing. That process is slow, and it breaks down every time a company wants to support a new topic or add a new tool.
Salesforce's patent describes a system where AI models do most of that work for you. One model plays the role of a task planner, inventing realistic conversation scenarios, complete with the correct steps and correct answers. A second model acts as both a simulated user and a trainee agent, practicing those conversations. If the practice responses match what the planner marked as correct, the exchange gets saved as a training example.
A third model then studies all those examples to become the actual deployed AI agent. The whole loop is designed to run automatically, so a business could feed it their own internal tools and policies and get a trained, domain-specific assistant without building a large labeled dataset from scratch.
… building, at a server, the AI agent employing the third neural network based language model after the training is completed; and generating, by the AI agent, a command to an autonomous driving system based on a user request.
Translation: The system builds the final AI helper and uses it to send commands to self-driving cars.
How three language models split the training workload
The system uses three separate language models (large AI programs that read and generate text) in a coordinated pipeline.
- Model 1 (the planner) receives a description of a company's available tools (called APIs, the software connectors that let an AI take actions like booking a meeting or looking up an order), plus the business domains and any specific rules or policies. It generates a multi-turn task as a structured bundle containing the goal, the correct sequence of actions, and the correct final outputs. Think of this as writing a detailed answer key before the test exists.
- Model 2 (the simulator and trainee) invents realistic user questions based on that goal, then tries to answer them and take the right actions. If its answers and actions match the answer key, that entire conversation gets saved as a verified training example.
- Model 3 (the student) is trained on all those verified examples. It learns to produce both natural-language responses and concrete software actions from multi-step prompts.
The training process uses a training objective (a mathematical score that measures how far the student's answers are from the verified correct ones) and adjusts the model's internal settings to reduce that gap over many rounds.
The claim specifically ends with the trained agent generating commands for an autonomous driving system, which signals that Salesforce sees this pipeline as applicable well beyond customer service, though the core method is domain-agnostic.
… generating, by the first neural network based language model, a multi-turn task including a tuple of (task intent, groundtruth actions, groundtruth outputs); …
Translation: The first AI model creates conversation goals and correct answers to train the system.
What this means for AI-powered business software
For businesses using Salesforce's AI products, a pipeline like this could mean faster setup and more reliable responses when a customer's request spans several steps, such as checking an account balance, filing a complaint, and then scheduling a callback. Today, getting an AI agent to handle that kind of chain without going off-script usually requires a lot of hand-tuned examples.
Salesforce's bet on agentic AI shows up clearly here. The company has been pushing AI that doesn't just answer questions but takes real actions inside business software. A self-generating training loop directly supports that direction by making it cheaper to adapt an agent to new tasks. Whether this specific pipeline makes it into a product as described is a separate question, but the problem it targets is real and widely felt.
Salesforce's 23rd filing we've tracked since May in our AI agents acting for you watchlist builds on earlier work covering teaching AI to read images and pulling live data before acting.
Claim 1 covers a three-model system where one AI invents realistic customer scenarios, a second AI practices handling them, and a third AI gets trained on that practice until it can hold full conversations and take real actions on its own. The claim spells out each handoff in the chain, meaning the protection runs to the whole pipeline, not any single piece of it.
If granted, that scope would block others from assembling this specific arrangement, where AI systems generate their own practice material, verify it, and pass it downstream to produce a deployable agent. The claim names autonomous driving as a concrete application, which grounds the method in a real product context without caging it there.
That breadth matters because Salesforce is claiming ownership of a training architecture, the way these systems are organized and sequenced together, rather than any one feature inside them.
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
16 drawing sheets from US 2026/0300690 A1 · click any drawing to enlarge
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →
Be the first to weigh in