Get the newsletter
Microsoft · AI / ML · Filed Jun 20, 2025 · USPTO published Nov 19, 2026 · This week's release

Microsoft Patents a Screening Step for the Files and Chats Its AI Reads

Before an AI agent acts on your behalf, Microsoft wants it to sort through your meetings, chats and files and throw out the junk. This patent application lays out a two-step filter to do that.

An office application sends work documents across a network so AI models and servers can analyze relevant project details for coworkers. Drawing from patent filing US 2026/0368232 A1.
An office application sends work documents across a network so AI models and servers can analyze relevant project details for coworkers.All 13 drawings ↓
Publication number
US 2026/0368232 A1(US20260368232A1)
Applicant
MICROSOFT TECHNOLOGY LICENSING, LLC
Filed · Published
Jun 20, 2025 · Nov 19, 2026
Inventors
Sakshi MUNJAL, . CHAITANYA, Priyank RASTOGI, Sachin SWAMI, Rohan GYANI, Sheetal S SETHI, Sandeep Kumar SRIVASTAVA
Status when we published
Approved; patent expected soon
Document
20 claims
US classification
707/749
Examiner
ROBINSON, GRETA LEE (Art Unit 2163)

How Microsoft's AI sorts what it reads before acting

You ask your work AI to prep a project update. Behind the scenes it has to read last week's meeting transcript, a pile of chat threads, a few shared files and maybe some web pages. Some of that is useful, some is outdated, and some of it flatly contradicts the rest.

Microsoft's patent application describes a two-step sorting process for exactly this mess. First, the system casts a wide net, pulling anything that might be relevant from every source and labeling each piece with what it is, where it came from and how it was shared. Second, it ranks and trims: it scores each piece for relevance to your task, spots overlaps and conflicts, and drops snippets that are weak and unconnected to anything else.

The AI then works from a cleaned-up summary instead of a heap of raw material.

From the filing · CLAIM 1
ranking the artifacts according to their relative importance and relevance; identifying dependencies, overlaps, and information conflicts; and filtering out artifacts that do not connect with other relevant artifacts and are below a minimum relevance threshold.

Translation: The system sorts the gathered data by importance, checks for contradictions, and discards anything irrelevant.

How Microsoft's two-stage pipeline picks the facts

The patent application describes a two-stage pipeline for feeding AI agents (software helpers that carry out tasks) the right background information. Engineers call that background grounding: the facts an AI model relies on so its output matches your actual situation.

Stage 1 is built for speed and recall (catching everything that might matter). A meta grounding agent queries each source, such as meeting transcripts, files and web content, and returns grounding artifacts. Each artifact is a chunk of text plus metadata saying what the text represents, where it was extracted from, and in what context it was shared.

Stage 2 is built for precision. A synthesizer does three things:

  • Ranks artifacts by importance and relevance to the task (the description names task relevance and source credibility as the main criteria).
  • Identifies dependencies, overlaps and conflicts between artifacts.
  • Filters out artifacts that connect to no other relevant artifact and also fall below a minimum relevance threshold.

The filing also describes helpers for the details. A preprocessor agent splits long text into chunks and tags topics, a reasoning agent clusters chunks by topic and drops the weak ones, and a questioning agent asks for missing information. The result is joined into one cohesive block of text, plus an index agents can use to look up the original artifacts on demand.

Why this matters for Microsoft's workplace AI agents

If you use an AI assistant at work, the quality of its answer depends on what it read first. The filing's own background says pulling from meetings, chats, files and web content adds delay and raises the risk of using wrong or irrelevant data. A filter that cuts clutter and settles conflicts before the agent starts could mean faster, more accurate results. It also addresses a limit the filing points out: the pile of source material can grow past what a large language model (the AI behind chatbots) can hold at once.

The application has been approved, and once Microsoft pays the final fee it should be issued as a granted patent. The filing names Microsoft's Multi-Agent Runtime Service as the setting, but it does not say when or whether any product will use this.

One more thing to know: the patent office has already approved this application. Once Microsoft pays the final fee, it should be issued as a granted patent, and the final claims may read a little differently from the version described here.

That makes this Microsoft's 35th filing we've tracked since July on our AI agents watchlist, building on turning manuals into fixes and finding the right button.

Source: US 2026/0368232 A1 at the USPTO (PDF). Full patent text and figures are in the official publication.

Get our take in your Top Stories

Liked this breakdown? Add Patentlyze as a preferred source on Google, and our plain-English take shows up more often in your Top Stories the next time Microsoft patent news breaks.

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 · 13 sheets

13 drawing sheets from US 2026/0368232 A1 · click any drawing to enlarge

Gathering relevant files before an AI summarizes the facts, drawing sheet 1 from US 2026/0368232 A1
Sheet 1 Gathering relevant files before an AI summarizes the facts
Checking retrieved meeting data through automated questions and feedback, drawing sheet 2 from US 2026/0368232 A1
Sheet 2 Checking retrieved meeting data through automated questions and feedback
Sending verified background notes onward to finish the user's request, drawing sheet 3 from US 2026/0368232 A1
Sheet 3 Sending verified background notes onward to finish the user's request
Breaking long documents into smaller chunks for careful ranking, drawing sheet 4 from US 2026/0368232 A1
Sheet 4 Breaking long documents into smaller chunks for careful ranking
Choosing specialized search assistants based on incoming user requests, drawing sheet 5 from US 2026/0368232 A1
Sheet 5 Choosing specialized search assistants based on incoming user requests
Fetching raw files and transcripts from across different cloud storage, drawing sheet 6 from US 2026/0368232 A1
Sheet 6 Fetching raw files and transcripts from across different cloud storage
Settings directing the AI to your meeting files, drawing sheet 7 from US 2026/0368232 A1
Sheet 7 Settings directing the AI to your meeting files
Refining background notes until every missing detail gets answered, drawing sheet 8 from US 2026/0368232 A1
Sheet 8 Refining background notes until every missing detail gets answered
Show the other 5 sheets
Virtual chat rooms where automated agents coordinate complex tasks, drawing sheet 9 from US 2026/0368232 A1
Sheet 9 Virtual chat rooms where automated agents coordinate complex tasks
Steps for filtering out irrelevant or conflicting reference data, drawing sheet 10 from US 2026/0368232 A1
Sheet 10 Steps for filtering out irrelevant or conflicting reference data
Computer hardware components that process and store reference files, drawing sheet 11 from US 2026/0368232 A1
Sheet 11 Computer hardware components that process and store reference files
Feeding screened file data through AI models into desktop apps, drawing sheet 12 from US 2026/0368232 A1
Sheet 12 Feeding screened file data through AI models into desktop apps
Servers delivering screened reference information to phones and tablets, drawing sheet 13 from US 2026/0368232 A1
Sheet 13 Servers delivering screened reference information to phones and tablets
Patent filing page

More in AI assistant & agent

All 436 AI assistant & agent patents →

Watchlists this filing belongs to

All watchlists →
Reader comments

No comments yet

Start the discussion

Real name or a handle, either is fine. Comments are read by a person before they appear, so allow a little time. Keep it about the filing.