Salesforce Patents an AI Agent That Grades Its Own Answers
Most AI chatbots have no idea whether their answers are any good. Salesforce is patenting a system where a second AI model watches the first one work and grades every response it gives.
What Salesforce's self-grading AI agent actually does
Imagine you hired a customer-service rep but also brought in a supervisor whose only job is to review every reply that rep sends. That's roughly what this patent describes for AI.
Salesforce's system uses two AI models working together. The first one handles your actual questions and keeps a running summary of your conversation. The second one reads that summary and the full conversation history, then produces three things: a score for how well the first AI did, an explanation of that score, and a direct quote from the conversation to back it up.
The clever part is that both models are connected inside a single AI agent, meaning the quality check happens automatically every time, not as a manual review step added later. For businesses using Salesforce tools to handle customer interactions, this is a built-in way to catch bad answers before they become a problem.
How the second model rates, explains, and cites in one call
The patent describes a two-model architecture for building AI agents that can evaluate their own outputs in real time.
Model one (the primary language model) does two jobs: it answers user questions and periodically compresses the conversation into a short summary. That summary is important because long conversations become too bulky to process efficiently, so the summary acts as a compact stand-in for the full history.
Model two (the evaluator) is trained specifically on a dataset that pairs those summaries with full conversation histories. When queried, it returns all three of the following in a single API call:
- A rating of how accurate or complete the summary is
- An explanation of why it gave that rating
- A citation, an exact excerpt from the conversation that supports the explanation
Both models are wired into a single AI agent through separate API connections (a programming interface that lets two software systems talk to each other). The agent can run many conversations at once, and the evaluator checks the outputs from all of them. The key engineering claim is that the rating, explanation, and citation all come back in one network request rather than three separate ones, which keeps the system fast.
What this means for AI reliability in business software
For companies that route real customer interactions through AI agents, answer quality is a liability issue, not just a technical preference. A system that automatically flags weak or inaccurate responses and points to exactly where in the conversation things went wrong is a meaningful step toward AI that businesses can actually audit.
Salesforce builds a lot of its enterprise product strategy around its AI platform, Agentforce. A patent like this fits squarely into that direction: rather than asking humans to manually review AI conversations, the platform itself generates the evidence trail. If this approach ships in a product, it would make it easier for compliance teams and managers to spot problems without reading every transcript.
This is a practical, unglamorous piece of infrastructure work that solves a real problem enterprise software buyers actually worry about. The citation feature is the most interesting part: generating a quote that supports the evaluator's reasoning moves this beyond a simple thumbs-up/thumbs-down score and toward something that could hold up in an internal audit. Don't expect a press release, but do expect to see something like this in Agentforce within a product cycle or two.
The drawings
15 drawing sheets from US 2026/0220451 A1 · click any drawing to enlarge
Which company should we read for you?
We track 17 companies here. Pro is the same weekly breakdown for any company you choose, delivered privately. Type a name and we'll scope it and send you a quote.
Get one Big Tech patent every Sunday
Plain English, intelligent commentary, no hype. Free.
Editorial commentary on a publicly published patent application. Not legal advice.