How Sphinx built an enterprise AI accuracy layer

Accuracy is one of the biggest problems plaguing enterprises as they seek to embed AI within their operations. The problem isn't only on the systems themselves, but rather the data they work from. 

It's the exact problem that Sphinx.ai wants to solve. Sphinx, a seed-stage startup backed by Lightspeed and Bessemer Venture Partners, enforces data accuracy across any AI tool an enterprise uses, giving companies the freedom to leverage whatever model they want without worrying about answers that drift from reality. 

"While AI is quite good at reasoning over things that are in the public domain that it's been trained on, every company has its own idiosyncratic way of laying out information," Rohan Kodialam, CEO and co-founder of Sphinx, told The Deep View. "It's not a question of intelligence, it's really a question of not knowing that tribal knowledge of how your organization works that limits AI's effectiveness."

Last month, the company unveiled Sphinx 2.0, the latest product in its lineup aimed at improving AI accuracy.

• This system essentially creates an institutional knowledge layer through which every AI query is filtered for accuracy. It sits between your AI tools and your data, continuously monitoring and refining outputs to better reflect your business context. 

• It's a stark change from Sphinx's first product, which used an agent to adjust and fix organizational data to prepare it for AI. Now, rather than making your data AI-ready, said Kodialam, Sphinx gets "your AI ready for your data." 

• The system also detects discrepancies and automatically surfaces them for human review by learning how an enterprise measures and tracks specific data. "We will find those instances where what we're seeing is in disagreement with what we already believe to be true," Kodialam said. 

Sphinx 2.0 solves a major pain point that enterprises are grappling with amid AI adoption: the inability to trust or rely on the outputs of these models. With too little data and organizational context, these models tend to hallucinate when queried about business-specific topics, making things up due to gaps in their training data. However, enterprise data is a moving target, with data stockpiles growing at a constant rate, while continually migrating it safely into AI systems is costly. 

Additionally, if enterprises simply connect all their data to AI systems in real time, they lose the observability and governance layers that make it safe. 

"We think there's like a happy medium of not spending a huge amount of money getting your data ready, but preparing your AI to operate in your real ecosystem," Kodialam said.

See how it works