Get the newsletter
IBM · AI / ML · Filed Mar 26, 2025 · USPTO published Oct 1, 2026 · Oct 1 release

IBM Patents a Way to Search Databases With Plain-English Questions

Most databases speak only one language: rigid code that takes years to learn. IBM has filed a patent describing a system that trains an AI to accept plain-English questions and translate them into real database results, no SQL required.

A sample relational database table with customer transaction data, including customer reviews. Drawing from patent filing US 2026/0300325 A1.
A sample relational database table with customer transaction data, including customer reviews.All 4 drawings ↓
Publication number
US 2026/0300325 A1(US20260300325A1)
Applicant
INTERNATIONAL BUSINESS MACHINES CORPORATION
Filed · Published
Mar 26, 2025 · Oct 1, 2026
Inventors
Bulent Abali, Rajesh Bordawekar
Status when we published
Approved; patent expected soon
Document
20 claims
US classification
707/739
Examiner
SANA, MOHAMMAD AZAM (Art Unit 2166)

How IBM wants plain questions to replace database code

A company's customer records sit locked inside a database. Only the people who know how to write precise code commands can ask it questions. Everyone else has to wait in line for a developer's help, or make do with pre-built reports that never quite cover what they actually need.

IBM's patent describes a system that bridges that gap. It takes the raw text entries already inside a database, runs them through an AI model to figure out what they mean, and then uses those insights to train a second AI that can answer natural-language questions. You type something like "which customers complained about shipping delays last quarter" and get back a clean, structured answer.

The key step is automatic labeling: the system reads your database entries, generates numerical "fingerprints" called embeddings that capture the meaning of each entry, and uses those to assign category labels. Those labeled entries then teach a machine learning model to handle new queries without a programmer writing custom code every time.

From the filing · CLAIM 1
… reading a query for the relational database, wherein the query comprises at least one of a query language code segment and a natural language prompt characterizing a database operation; performing the database operation using at least the first machine learning model; and returning a response to the query as a structured result.

Translation: The system accepts plain English questions or code, runs the request through its model, and gives back a structured answer.

How the encoder labels entries and trains the query model

The patent describes a multi-stage pipeline for making relational databases queryable in plain English.

Stage one: encoding. Every text entry in a database table is fed into an encoder model (an AI that converts text into numerical vectors, called embeddings, that represent meaning rather than just words). Two entries with similar meanings end up with similar vectors, even if the wording differs entirely.

Stage two: labeling. The system clusters or categorizes those embeddings to assign a label to each database entry. Think of it like automatically sorting an unmarked pile of customer notes into buckets: complaints, compliments, shipping issues, returns. No human has to hand-label the data.

Stage three: training. The labeled entries are written into a new training table inside a copy of the database. A first machine learning model is trained on this labeled data so it understands the structure and meaning of what's stored.

Stage four: querying. When a user submits either a plain-English prompt or a formal query-language command, the trained model interprets the request, performs the appropriate database operation, and returns a structured result. The claim covers both natural-language prompts and traditional query syntax, so the system can serve technical and non-technical users from the same interface.

What this means for workers who can't write SQL

For most organizations, getting answers out of a database still requires a developer or data analyst in the loop. That bottleneck slows decisions and keeps useful information out of the hands of the people who actually need it. A system like the one IBM describes could let a sales manager, a nurse, or a supply-chain planner ask real questions in plain English and get reliable answers without waiting on IT.

IBM's track record in enterprise AI patents suggests the company is targeting large organizations that run structured databases but employ far more non-technical staff than developers. Whether this particular approach proves reliable enough for production use is a harder question, but the direction is clear: make the database answer to you, not the other way around.

IBM's 39th filing we've tracked in Language AI since May adds to a run that includes one tracing AI text origins and one tailoring prompts to each user.

Source: US 2026/0300325 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 · 4 sheets

4 drawing sheets from US 2026/0300325 A1 · click any drawing to enlarge

Steps for training a database search model., drawing sheet 1 from US 2026/0300325 A1
Sheet 1 Steps for training a database search model.
Sample customer transaction database table., drawing sheet 2 from US 2026/0300325 A1
Sheet 2 Sample customer transaction database table.
System for training and querying a database., drawing sheet 3 from US 2026/0300325 A1
Sheet 3 System for training and querying a database.
Computer system hardware components., drawing sheet 4 from US 2026/0300325 A1
Sheet 4 Computer system hardware components.
Patent filing page

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.