Microsoft Patents an AI System That Figures Out What You Actually Meant
When you ask an AI to "book me something for Friday," it usually takes you at face value. A new Microsoft patent describes a system that goes further working out what you'd actually prefer before it acts.
How Microsoft's AI reads between the lines of your requests
Ever asked someone to "just handle it" and ended up with results that were technically correct but totally wrong for you? That gap between what you say and what you mean is the problem this patent is trying to close.
Microsoft is describing an AI system that, when you give it a task, does some background research on you before it acts. It looks at your past behavior and relevant data, asks its own follow-up questions under the hood, and builds a picture of your unstated preferences before deciding how to complete the task. You don't have to explain your tastes every time.
The idea is that AI assistants working on your behalf should do more than execute instructions literally. If you always pick economy seats but ask it to "book a flight," a well-designed system should already know that, without you having to spell it out each time.
… inferring a non-explicit user preference based on the input and the learning; and sending the non-explicit user preference to a task agent of the multi-agent system to control execution of a task according to the non-explicit user preference.
Translation: The system figures out what you want without you asking and tells another AI agent to handle it.
How the multi-agent pipeline infers unstated preferences
The patent describes a multi-agent AI system where several specialized AI sub-agents work together before a task is ever executed.
When you provide input, it goes to a dedicated input augmentation sub-agent. That sub-agent takes your request alongside your user metadata (things like past behavior or profile data) and instructs a generative AI model to write and run a query against a dataset. Think of it as the sub-agent asking the AI to write a small program to look something up, then collecting and summarizing what comes back.
The results of that lookup are bundled into an embedding (a mathematical representation that captures meaning and relationships) of both your original input and the newly gathered context. The system uses this combined representation to infer a non-explicit user preference, meaning a preference you never directly stated but that can be deduced from the evidence.
That inferred preference is then handed off to a separate task agent, the part of the system that actually does the work. The task agent uses the inferred preference to shape how it carries out your request, rather than acting on the bare literal instruction alone.
- Input augmentation sub-agent receives your request and metadata
- Generative model writes and runs a query to gather context
- Results are embedded alongside your original input
- A non-explicit preference is inferred and passed to the task agent
The query is formulated by a generative machine learning model (GMLM) in response to a code generation instruction.
Translation: An AI writes computer code to search a database based on what you typed.
What this means for AI assistants acting on your behalf
The practical payoff is that an AI assistant acting on your behalf would make fewer decisions you'd want to undo. Today's AI tools are often blunt: they take your words at face value and miss the context that a human assistant would pick up naturally. A system that infers your preferences before acting could mean fewer corrections, fewer do-overs, and less babysitting of AI output.
The stakes are higher as AI agents take on more consequential tasks, such as scheduling, purchasing, or managing files on your behalf. An agent that gets your implicit preferences wrong on a simple email is annoying; one that gets them wrong on a contract or a purchase is a real problem. Microsoft's steady investment in agentic AI filings signals this is a priority area, not a one-off experiment.
This is the 11th Microsoft filing we've tracked since July in our AI agents that act for you watch, following one on meeting attendance and summaries and one on AI presenting slides for you.
The cost of an AI acting on a misread preference is small when you catch it once. It compounds fast when the system is making dozens of calls on your behalf, each one slightly off from how you actually think, and you only notice the drift after the damage is done.
What this filing targets is that gap between what you said and what you meant, specifically by gathering context about you before a task runs rather than waiting for you to correct mistakes afterward. That timing matters. A correction fixes one thing; upfront context shapes everything downstream.
The honest question is whether pulling in data about someone actually captures their judgment, and those are not the same thing. This approach narrows the distance in a believable way, but the problem is big enough that no single mechanism closes it completely.
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
12 drawing sheets from US 2026/0267868 A1 · click any drawing to enlarge
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →