Probably Raises $9M to Build a More Reliable Kind of AI
Probably, a startup working on making large language models more accurate and reliable, has raised $9 million in funding to advance its approach to preventing hallucinations and factual errors in AI systems. The company, founded by Peter Elias, is pursuing a fundamentally different strategy than most AI labs: rather than building bigger models, it is engineering tighter harnesses around existing ones to catch errors before they reach users.
Series Seed Round Led by Andreessen Horowitz
The company announced today that it has secured a series seed round led by Andreessen Horowitz, with participation from existing investors. The funding will be used to expand the team, accelerate product development, and further refine Probably's proprietary technology for reducing hallucinations in generative AI. The round comes at a pivotal moment for the industry, as enterprises increasingly demand reliability guarantees that raw model capability alone cannot provide.
The "Data Science Mech Suit"
Probably's first product is a data science tool built to produce quick answers from complex datasets. Each result comes with a citation and an audit trail documenting how the answer was developed — an increasingly common practice among AI tools, but one that Probably has taken further with its proprietary harness system.
Elias describes this system as a "data science mech suit." The LLM's first-pass answers are checked against a deterministic validator that bounces back any results that don't match the underlying dataset. Crucially, the LLM has been trained specifically against this validator, and the entire system is optimized for both speed and accuracy. This creates a feedback loop where the model learns to produce outputs that satisfy the validator's constraints from the start, rather than relying on post-hoc correction.
"What we learned building this was that the better your harness engineering is, the weaker the model can be," Elias says. "If you can refine the context enough, the model does not have to work very hard to do the right thing. Basically, it's an exercise in reducing ambiguity."
This insight is central to Probably's approach: by investing heavily in the engineering of the system surrounding the model, the company can achieve high accuracy even with relatively modest underlying models.
Tackling the Hallucination Problem
The core challenge in modern AI is that large language models often generate factually incorrect information — a problem the industry commonly refers to as "hallucinations." These errors can range from minor inaccuracies to completely fabricated facts that appear convincingly detailed. As LLMs have grown more powerful, hallucinations have proven stubbornly difficult to avoid, with errors appearing even in the smartest models.
Probably aims to solve this problem by developing techniques that detect when an AI model is uncertain or lacks sufficient grounding in factual knowledge. The company's goal, as Elias puts it, is to "prevent hallucinations and simple factual errors from ever reaching the user, and achieve the kind of 99.99% accuracy that's common in deterministic systems but much more difficult to reach with AI."
"The goal is to get accuracy on par with deterministic systems," Elias says. "Users shouldn't have to second-guess every answer they get from an AI. When the model doesn't know something, it should be able to say so clearly and helpfully."
Bringing LLMs to that level of accuracy requires rethinking many of the basic assumptions of AI engineering, according to Elias. The approach involves not just better models, but better systems built around those models.
Running on Smaller Models, Local Hardware
One of the most significant implications of Probably's harness-first approach is what it means for deployment costs. Because the system does so much of the heavy lifting in error prevention, Probably's data science tool can run on significantly smaller AI models. Elias says the current version is running on a model that's "four classes weaker than the frontier models."
This has practical implications: smaller models can be run on local hardware — a desktop computer instead of a data center — which reduces a huge amount of the token costs associated with AI use. The cost savings are substantial, particularly for enterprises running these tools at scale throughout the day.
It's a welcome idea at a time when token costs are rising and many customers are reassessing their AI budgets. The approach also reduces dependency on cloud providers and gives organizations more control over their AI infrastructure.
The Business Case for Reliable AI
While many startups are focused on making AI models more capable, Probably is taking a different approach: improving the reliability of existing models. This strategy has resonated with enterprise customers who need trustworthy AI for customer-facing applications, legal research, medical decision support, and other high-stakes domains.
But the applicability extends beyond data science. Elias says the same engine can be extended to cover use cases like accounting or medical services — what he calls "any precision-sensitive use case." The underlying principle is universal: wherever accuracy matters more than raw capability, Probably's harness approach can add value.
The company plans to release its first commercial product later this year, with a focus on integration through APIs that can be easily incorporated into existing AI applications. This API-first strategy allows enterprises to adopt Probably's reliability layer without overhauling their existing AI infrastructure.
Why Now: The Accuracy Imperative
The timing of the funding round reflects growing industry awareness that raw model capability alone is insufficient for real-world deployment. Recent high-profile incidents involving AI hallucinations have demonstrated the financial and reputational risks of deploying unreliable systems. Enterprises that have deployed AI at scale are now facing the reality that accuracy gaps translate directly into business risk.
Investors are increasingly looking for solutions that address the "last mile" of AI deployment — making sure the outputs users actually see and act upon are accurate, safe, and trustworthy. Probably's approach addresses this need directly by focusing on the deployment layer rather than the model training layer.
"I think it's really interesting that the big AI labs have not even attempted to do this," Elias says. "They're incentivized not to, because they make money the more times you have to correct the model."
This observation highlights a structural tension in the AI industry: the companies building foundational models have different economic incentives than the enterprises trying to deploy them reliably.
Looking Ahead
With the new funding, Probably plans to expand its research team and accelerate development of what it calls "confidence-aware" AI systems. These systems wouldn't just generate answers, but would also indicate their level of certainty and provide sources when available. The audit trail functionality already built into the data science tool is a precursor to this broader vision.
The company is also exploring partnerships with other AI startups and established players who want to improve the reliability of their products without rebuilding foundational models from scratch. This partnership strategy could amplify Probably's impact across the industry, as its harness technology becomes a standard component of enterprise AI deployments.
For more information, visit the company's website or follow updates on their blog.
