Andreessen Horowitz is placing a significant bet that AI reliability is not just a feature to bolt onto existing products, but an entirely separate market category worth building around. The venture firm led a $9 million seed round for Probably, a startup founded by Peter Elias that is building a verification layer for enterprise AI systems. The company's core thesis is straightforward: enterprises will not deploy AI into serious work,finance, healthcare, legal, or data science,until someone can prove the answer, not just generate it.

Why Are Enterprises Demanding Proof Over Promises?

The hallucination problem in large language models (LLMs), which are AI systems trained on vast amounts of text data, is not going away through better prompts or more sophisticated models. When an LLM produces an answer, it generates text that sounds confident and fluent, but the system has no built-in mechanism to verify whether that answer is actually correct. For a consumer chatbot, this is annoying. For a compliance officer, a chief financial officer, or a hospital administrator, it is a real blocker to adoption.

The market has already demonstrated this demand through real-world failures. A May 2026 research paper audited 111 million references across 2.5 million academic papers and found that 146,932 hallucinated citations appeared in 2025 alone, introducing false information into the scientific record at scale. More recently, consulting firm KPMG withdrew an agentic AI report after fact-checkers discovered false or misleading case studies involving organizations including UBS, the NHS, Swiss Federal Railways, and Transport for London. When a firm that sells trust for a living becomes a case study in AI hallucination, the market's appetite for verification becomes undeniable.

What Does Probably's Approach Actually Do?

Probably's first product is a data science tool that lets users query complex datasets and receive answers with citations and an audit trail showing exactly how the result was derived. The critical innovation is the second step: the system uses a deterministic validator, a rules-based layer that checks the LLM's answer before it reaches the user. The model does not get the final word. A verification system has to clear it first.

The company is targeting a 99.99% accuracy standard, the same level of reliability that enterprises expect from deterministic software like spreadsheet formulas, SQL queries, or validation rules. This is not a small ambition. Most enterprise AI pilots already prove that people like fast answers. The harder test is whether anyone inside a large company is willing to sign their name to those answers when money, regulation, or patient safety is involved.

How to Build Enterprise AI That Enterprises Will Actually Trust

• Provide Citations and Audit Trails: Every answer must come with documentation showing which data sources were used, what calculations were performed, and what assumptions were made, so the result can be verified after the fact.
• Implement Deterministic Validation: Add a rules-based verification layer that checks the AI system's output against known facts and constraints before the answer reaches the user, rather than relying solely on the model's confidence.
• Target Workflows Where Correctness Is Definable: Start with use cases where right and wrong answers can be clearly distinguished, such as data science queries that trace back to source tables, rather than judgment-heavy tasks where validation is ambiguous.

Probably is starting in the part of the market where this mapping is clearest: data science and analytics, where many answers can be traced back to source tables, calculations, and documented assumptions. The harder cases will come when the output is more judgment than computation. A validator can check whether a number came from the right dataset. It cannot always decide whether a business conclusion is sound.

Why Does a16z's Bet Matter Beyond One Startup?

Andreessen Horowitz's investment is significant not because of the dollar amount, but because it puts venture capital behind reliability as a standalone category, not a small feature buried inside every AI application. This is not investor theater. Security became its own market because companies could not simply trust every cloud product to protect itself. Observability became its own market because distributed software broke in ways people could not see from the outside. AI reliability has the same shape, with a stranger problem underneath it: the system can be wrong while sounding perfectly confident.

The open question is whether Probably can turn its first product into a broader platform before the model providers, cloud companies, or data warehouse vendors build enough reliability features themselves. That is the normal danger for infrastructure startups. The best wedge becomes a feature if the platform catches up. Still, the timing is good. Enterprises do not need another promise that the next model will hallucinate less. They need a way to know when today's model is wrong.

For founders building enterprise AI, the lesson is clear: the market is past being impressed by a fluent answer. Buyers now want citations, audit logs, validation steps, error rates, and someone accountable when the system fails. If a product cannot provide those things, it is asking a serious customer to take a leap it does not need to take.

" }