722 AI math papers: what to verify before trusting the claims

OpenAI published 722 mathematical papers.

According to the company, they solve 90 of the 500 best-known open problems.

A researcher at a competing lab called it the most significant moment in mathematics in more than a century.

For an executive buying AI for business results, this news suggests something else: a model's value is determined by how quickly and reliably its answer can be verified.

In a typical company, an AI project has no formal verifier.

The model writes code, product descriptions, customer replies, and contract summaries, but a human checks them, making that person's time the bottleneck.

If verifying an answer takes as long as doing the work yourself, there is no gain.

The gain is measured by TTU — time to use: how quickly the tool produces a usable result.

Mathematicians got a verifier from mathematics itself.

Businesses have to build it: code tests, data schemas and rules, source verification for texts, and limits on agent actions.

RAG anchors the model's answer to company documents.

The answer includes a link to a source excerpt, reducing verification to a comparison of two texts.

This is usually much faster than checking everything from scratch. ###

Rule-based verification In product catalogs and integrations, schemas, taxonomies, and business rules serve as the verifier.

The model fills in a record, and the validator rejects any value outside the approved reference list.

A human reviews only disputed cases. ### Access and action control

When an agent accesses internal systems through MCP, every operation passes through a gateway: what can be read, what can be written, and which data must not leave the environment. In KT.Team, this role is performed by the LLM & Security Gateway.

It logs calls and filters sensitive data, so the agent's work can be audited. ###

Test-first development In AI-native development, the model generates code, while tests and CI define the quality threshold.

The broader the coverage, the more changes reach production without manual rechecking.

The simple “ask — receive” interface relies on a complex verification layer, which engineers have to build.

If some of the papers are confirmed, this will show that the combination of “generation plus automated verification”

handles problems that people could not solve for decades.

If only a small share is confirmed, the conclusion remains the same on a smaller scale: generated volume is worthless until the verified volume is measured.

For budget planning, three questions are useful for any AI project: -

How is the model's result verified, and what does it cost per result? -

What share of answers passes verification without human involvement? -

Which business metric changed during the quarter?

A project without answers to these questions remains theater, and theater eventually faces budget cuts.