Paris and San Francisco, December 2nd 2025 – Today, Basalt announces a $5M fundraising round to restore trust in AI agents. While launching an AI agent is easy, scaling it with consistent, production-grade performance remains a major challenge and most AI implementations fail in large companies today. Basalt is using this new funding to accelerate its mission to become the platform companies rely on to reach 99% quality in their AI applications. The company is already proving the demand, working with clients such as Swan, HealthHero and backed by leading investors including Entourage, Peak, Alpha Star, Kima Ventures, and Hexa.
AI is facing a reliability crisis
AI agents are rapidly spreading across industries, promising major productivity gains. But in large companies, most deployments still fail, because of a lack of reliability. When an agent delivers inconsistent or mediocre results, trust disappears, and organizations hesitate to scale it across critical operations.
Why is it hard to create a reliable AI Agent ?
AI agents today struggle with reliability for two main reasons. First, achieving high-quality performance requires continuous iteration on prompts, a workflow that differs significantly from traditional software development. Second, the tools available to assess AI quality are largely built for engineers, limiting the input needed to define and validate “good” outcomes. Basalt solves this with the first collaborative AI engineering platform, designed to help companies reach 99% quality on their AI apps.
“Anyone can ship an AI prototype, but getting from 80% quality to true production grade remains painfully hard. The last 20% requires constant iteration on prompts and learning from edge cases – much like a child learning to walk, taking a step, stumbling, understanding why, and adjusting. AI systems need that same repeated exposure and correction to become dependable.” Guillaume, cofounder of Basalt.
Introducing a new approach to reliable AI
Basalt is building the platform teams use to bring their AI agents from early prototype to production-grade quality. Getting it reliable requires fast iteration across the whole team, not just engineers.
“Prompts are the new building blocks of AI agents, the way code is the building block of software,” says François de Fitte, cofounder of Basalt with Guillaume. “The difference is that prompts don’t require you to speak JavaScript, just English. That’s why reliability can and should be owned by everyone, from PMs to operators and domain experts…
Three core capabilities for reaching production grade
Basalt focuses on the three essentials every AI team needs to deliver trustworthy agents:
Experiment – Try new prompts or chain them together, compare LLMs and validate improvements before anything goes live. Move fast without breaking your user experience.
Evaluate – Run structured tests across hundreds or thousands of scenarios, score outputs automatically and catch errors instantly. Bring rigor to what used to be vibe checking.
Monitor – Understand how your agent behaves in the real world. Surface hallucinations, regressions or unexpected behaviour as soon as they pop up. Add these errors are new scenarios to test in your next Experiments.
Companies like Swan and Healthhero enlist their product managers to review errors directly in Basalt, test new prompt versions or different LLMs, and push updates straight to production.
Fundraising
To accelerate this mission, Basalt announces a 5M dollar round led by Entourage (Aïkido, Epiminds, Conveo) and Peak (Workwize, Catawiki, Studocu), including Hexa, Alpha Star and Kima Ventures.
Investor quotes
PJ, Entourage: “AI agents are becoming the new software workforce, and Basalt is building the system that keeps them accountable and reliable. This is a once-per-decade opportunity, and Basalt has the potential to be a clear category leader.”
Tea, Peak: “Basalt is doing for AI agents what Datadog did for cloud infrastructure. Every company will need this reliability layer, and Basalt’s collaborative approach is way ahead of the market.”
Thibaud, Hexa: “In a world where AI systems make more and more decisions, trust becomes the ultimate currency. Basalt is the company that turns trust into something measurable, traceable, and scalable. Guillaume and François understand this problem at a depth we rarely see.”
About Basalt
Basalt was founded by repeat entrepreneurs Guillaume Marquis (Virtual Brain, Blockpulse) and Francois de Fitte (Popchef). The company is headquartered in San Francisco, with its tech team based in Paris. The money will allow the company to hire more tech talent in France and the US.
As Guillaume was working on Virtual Brain, his previous startup, he noticed that his team kept copying and pasting prompts from Notion or Google Sheets into the codebase whenever they wanted to test even a tiny change. Basalt began as a simple Prompt Registry to speed this up and quickly try new LLMs as they came out. But soon enough, both founders realised the real pain wasn’t tweaking prompts, it was iterating based on what actually happened with real customers. They saw that AI engineering isn’t like traditional software development at all. It requires a continuous loop between observation and experimentation, without which quality stalls.
“I was mindblown when I saw that Excel Sheets were used by large companies to iterate on prompting, and no one had any idea of how well AI was working in Production. Some companies would say ‘we have no feedback, so it seems to be working well’. This is missing the point: AI isn’t binary, some AI outputs are not fundamentally wrong, they’re just mediocre and don’t work well enough to be considered useful.”