Axiomat, a Y Combinator startup that Y Combinator's directory lists with three employees, sells a data layer that checks AI agents' answers about company data against a logical model of the business. The company has not disclosed any funding.
Its pitch is that mathematics, rather than another model's opinion, should decide whether an agent's query is correct. The claim is narrower than it sounds, and the company's own materials say as much.
What Axiomat Says It Does
According to its website, Axiomat converts a customer's documentation and database or API schemas into formulas that the customer then approves. Answers are checked against those formulas. Y Combinator's directory says the company formally verifies agents' queries before people rely on the results.
The about page states that the workflow requires human approval of business rules, and that the system answers only when those rules determine exactly one answer. That is a tighter promise than "eliminating" agent errors, a phrase that appears in the company's positioning. I could not find an independently reproducible benchmark supporting universal error elimination, so the phrase should be read as marketing, not as a demonstrated result.
The website's examples include incorrect revenue calculations and customer counts that wrongly include trial accounts. Axiomat presents them as illustrations, not as incidents at named customers. It offers demos and consultations, and I did not locate a separately dated launch announcement.
Team and Backing
Y Combinator's directory, as accessed on November 5, 2026, lists Axiomat as a Fall 2026 company in San Francisco with three employees. Directory headcounts are observational and may change. The listing attributes team experience to AWS and to research at EPFL, the University of Toronto and Georgia Tech, the last described at the team level and not tied to any one person.
The website names Jack Naimer and Ethan Kharitonov as co-founders, along with Elad Perlman. I could not verify who holds the CEO or CTO role.
Y Combinator's description says Kharitonov was an Applied Scientist at AWS working on formal verification of AI responses, though dates and his individual contributions were not verified. Naimer's self-published résumé lists engineering science studies at the University of Toronto from 2021 to 2025, followed by master's work in EPFL's MLBio Lab. EPFL's directory links him to its Machine Learning for Biomedicine laboratory. Whether he is still enrolled needs confirmation.
The website's privacy notice, last updated October 25, 2026, names the operator as "Axiomat (Signiture Inc.)." A Signiture LinkedIn page describes mathematically proving queries correct and lists the same founders. Beyond Y Combinator, no investor, financing amount, round date or valuation turned up.
The Problem It Targets
The demand case rests partly on a Nylas survey released October 20, 2026. Of 1,030 U.S. professionals, 54% reported an agent error in the preceding 90 days, 31% a workflow that had to be reversed, and 25% an update to the wrong record. Only 53% said they could produce a full audit log.
Some caveats apply. Nylas sponsored the survey, and the errors it counts include actions and communications, not only data queries. The 54% reflects respondents with at least one incident, so it is not a per-action failure rate. Nylas also reported that incident prevalence climbed from 45% among organizations running one to five agents to 70% among those running more than 20. That is an association. It does not show that adding agents causes more errors.
Broader adoption data points in a similar direction. McKinsey's August 25, 2026 survey of 1,719 respondents found 44% reporting enterprise-wide scaling of AI, up from 38% a year earlier, yet only 6% cleared its high-performer bar of at least 5% of earnings before interest and taxes attributable to AI. Gartner, on October 2, 2026, predicted that 40% of agentic AI projects will be canceled by the end of 2027, citing costs, unclear business value and inadequate risk controls. That is a forecast, not a measured cancellation rate.
Who Axiomat Is Up Against
The nearest large-platform comparison is AWS. Its Bedrock Guardrails feature, Automated Reasoning checks, converts policy documents into formal rules and tests generated responses against them. AWS added automated policy-refinement workflows on June 23, 2026.
AWS's documentation is candid about limits. A VALID result covers only statements captured by the policy's variables, and the natural-language translation step affects accuracy. Complex policies add latency and can return TOO_COMPLEX.
Snowflake and Databricks are competing for trusted natural-language access to enterprise data. Snowflake's Cortex Analyst relies on semantic definitions and verified pairs of questions and SQL. A "verified query" there is a curated reference query, not automatically a formal proof for each new one. Databricks' Genie Agents draw on curated knowledge stores, example SQL and trusted parameterized queries, which it defines as logic verified by an agent's author.
Further afield, Microsoft released an open-source framework called Interwhen in February 2026, offering test-time monitors and generated verifiers. Harmonic's Aristotle and Axiom Math work on adjacent proof infrastructure. Axios reported on May 26, 2026 that Axiom announced $200 million at a $1.6 billion valuation in March 2026. That financing has nothing to do with Axiomat, whose name is easy to confuse with it.
Formal verification of SQL also predates the agent wave. Axiomat did not invent it.
What a Proof Does Not Cover
Harmonic's documentation notes that proofs hold for specified properties under explicitly listed assumptions. A proof can show that an answer complies with a specification without showing that the specification captures what the user meant. That puts Axiomat's rule-approval and interpretation-review steps at the center of any assurance it offers.
AWS's policy-refinement release treats iterative policy tuning as a source of customer friction. Axiomat's workflow, which turns documentation into rules, aims at something similar, though how the two compare in effectiveness is unverified.
Looking Ahead
The open commercial question is whether per-query formal guarantees justify the added setup, restrictions, latency and cost, given that Snowflake and Databricks already offer semantic definitions and evaluation tools. That is my inference from the available material, not a sourced market forecast.
Policy attention is growing. The National Institute of Standards and Technology launched its AI Agent Standards Initiative on February 17, 2026, naming trust and access to internal data as adoption obstacles. It is a standards effort, not a mandate for formal proofs.
For now, the better test of Axiomat may be its own rule. A system that answers only when the approved rules determine exactly one answer will sometimes decline to answer, and buyers will decide how often they can live with that.
