Get the newsletter
IBM · AI / ML · Filed May 16, 2025 · USPTO published Nov 19, 2026 · This week's release

IBM Patents a Panel of AI Experts That Quiz Each Other About Broken Machines

IBM describes a software setup where AI characters playing a reliability engineer, a quality engineer and other experts interview each other, then turn the answers into a draft failure checklist for industrial equipment.

Simulated industrial experts gather facts about broken equipment and question each other in repeated rounds to find the cause of a breakdown. Drawing from patent filing US 2026/0368237 A1.
Simulated industrial experts gather facts about broken equipment and question each other in repeated rounds to find the cause of a breakdown.All 13 drawings ↓
Publication number
US 2026/0368237 A1(US20260368237A1)
Applicant
International Business Machines Corporation
Filed · Published
May 16, 2025 · Nov 19, 2026
Inventors
Dhavalkumar C. PATEL, Christodoulos CONSTANTINIDES, Vivek SHARMA, Shuxin LIN, Nianjun ZHOU, Bharathi CHAUDHURY
Status when we published
Approved; patent expected soon
Document
20 claims
US classification
707/736
Examiner
HOANG, SON T (Art Unit 2169)

What IBM's panel of AI experts actually does

Ever tried to train a new coworker on how a machine can break? The expert knows the answers, but the newcomer doesn't know which questions to ask.

That gap is what this IBM patent application goes after. Companies building AI to predict equipment failures need expert knowledge, and today data scientists spend a lot of time working out what to ask the experts. IBM describes a system that skips that step. You tell it what kind of equipment you care about, say a wind turbine gearbox, and it creates a small cast of AI characters, such as a quality engineer, a reliability engineer and a subject matter expert.

Those characters question each other. One answers, another builds a follow-up on that answer, and the system weeds out repeats and weak questions along the way. What comes out is a catalog of questions and answers that other software can use, including a draft checklist of how the equipment might fail.

From the filing · CLAIM 1
… collaboratively generating, by the plurality of AI agents, a second set of questions based on refinement of the first set of domain-specific questions and interactions among the plurality of AI agents …

Translation: The artificial intelligence agents work together to improve the questions by interacting with each other.

How the AI personas pass questions along

The system starts with a domain context: an asset class (a type of equipment), a problem type, or a data source. The filing's example is a wind turbine gearbox in predictive maintenance, meaning predicting when a part will need service. From that, it picks AI personas, each with a fixed role such as data scientist, subject matter expert, summarizer, question writer or question router. The filing says each one can run on a large language model (the AI behind chatbots).

A first batch of questions is generated from the input. Each question goes to the persona best suited to it, chosen by a role-specific classifier (a small AI model that decides who takes which question). The core claim is the next step: one agent's answer is used to seed a follow-up question from a different agent, over several rounds.

Quality checks between rounds, as the claims describe them:

  • Filter out near-duplicates using semantic similarity (how close in meaning two questions are)
  • Rank what is left by relevance to the domain
  • Rewrite for clarity
  • Validate with pre-trained classifiers that test whether each question is specific, unambiguous and new

Answers are stored next to their questions. An agent can then draft a failure modes and effects analysis (FMEA), a standard industry document listing how equipment can fail, what happens when it does, and what to do about it.

Why IBM wants machine-repair knowledge on tap

If you work near factories, power plants or fleets of machines, the pitch is faster groundwork for AI that predicts breakdowns. Instead of an engineer repeatedly interviewing a veteran technician, software could produce a first draft of what can go wrong with a pump or turbine. Fewer surprise failures would mean less downtime and lower repair costs, which is the benefit the filing points to.

The patent office has approved this application, and once IBM pays the final fee it should be issued as a granted patent. For now the filing describes software only, and it gives no test results, so how well the AI experts perform on real equipment is still unproven.

That makes this IBM's 26th filing we've tracked since May in our AI teams working together watch, following filings like one turning problems into math and one matching AI to each job.

Source: US 2026/0368237 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 IBM 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 2026/0368237 A1 · click any drawing to enlarge

Network setup linking factory sensors and computers to the AI, drawing sheet 1 from US 2026/0368237 A1
Sheet 1 Network setup linking factory sensors and computers to the AI
Connecting AI models into each phase of data analysis, drawing sheet 2 from US 2026/0368237 A1
Sheet 2 Connecting AI models into each phase of data analysis
Setting up specialized AI personas to study equipment types, drawing sheet 3 from US 2026/0368237 A1
Sheet 3 Setting up specialized AI personas to study equipment types
How AI personas interview each other to diagnose machine faults, drawing sheet 4 from US 2026/0368237 A1
Sheet 4 How AI personas interview each other to diagnose machine faults
AI assistants generating maintenance alerts from equipment records, drawing sheet 5 from US 2026/0368237 A1
Sheet 5 AI assistants generating maintenance alerts from equipment records
Filtering and refining diagnostic questions through AI experts, drawing sheet 6 from US 2026/0368237 A1
Sheet 6 Filtering and refining diagnostic questions through AI experts
Pairing generated technical questions with specialized AI experts, drawing sheet 7 from US 2026/0368237 A1
Sheet 7 Pairing generated technical questions with specialized AI experts
Removing poor questions to build a better test pool, drawing sheet 8 from US 2026/0368237 A1
Sheet 8 Removing poor questions to build a better test pool
Show the other 5 sheets
How AI personas quiz each other across multiple rounds, drawing sheet 9 from US 2026/0368237 A1
Sheet 9 How AI personas quiz each other across multiple rounds
Cloud servers and devices that run the AI, drawing sheet 10 from US 2026/0368237 A1
Sheet 10 Cloud servers and devices that run the AI
Computer hardware that powers the machine diagnostic system, drawing sheet 11 from US 2026/0368237 A1
Sheet 11 Computer hardware that powers the machine diagnostic system
Steps for AI experts to generate and polish questions, drawing sheet 12 from US 2026/0368237 A1
Sheet 12 Steps for AI experts to generate and polish questions
Training AI assistants to diagnose faulty equipment together, drawing sheet 13 from US 2026/0368237 A1
Sheet 13 Training AI assistants to diagnose faulty equipment together
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.