Every enterprise wants AI to understand its business. Few realize how much work it takes to make that happen.
That realization became impossible for Rohan Kodialam to ignore during his years leading AI research at Citadel. Despite working alongside some of the industry’s most accomplished researchers, he found that their most valuable contribution was rarely inventing new algorithms. Instead, they were spending their time continuously tweaking prompts, updating instructions, rewriting documentation, and manually correcting AI systems that still made critical mistakes. The models were powerful, but they required constant supervision to remain accurate as businesses evolved. Kodialam came to see this “model babysitting” as one of the biggest bottlenecks preventing enterprise AI from delivering on its promise. Rather than building another AI assistant, he co-founded Sphinx AI to eliminate the need for that manual intervention altogether.
“The best marginal use of incredibly smart people wasn’t improving the models,” Kodialam says. “It was manually keeping those models working. That simply isn’t scalable.”
Today, Sphinx is tackling what Kodialam believes is the next major infrastructure challenge in enterprise AI: teaching AI systems the unique context that defines every business.
Unlike public knowledge, business context cannot be scraped from the internet or learned from training on books. It lives inside organizations through years of accumulated processes, terminology, historical decisions, and institutional knowledge. While large language models excel at general reasoning, they know nothing about how an individual company measures success, defines its metrics, or makes operational decisions. According to Kodialam, that gap explains why so many enterprise AI deployments continue to produce inconsistent answers, conflicting reports, and unreliable business insights.
“Your business exists because of knowledge that nobody else has,” he explains. “That’s your competitive advantage, and it’s exactly the information foundation models don’t come with.”
Most enterprise AI platforms assume companies will manually document that knowledge before agents can use it. Kodialam compares the process to filling an empty box with standard operating procedures, documentation, and business rules. The problem is that creating and maintaining those documents can consume months of work, and even then they quickly become outdated. He points to examples where organizations, including leading AI companies, have found themselves relying on teams of experts to continuously rewrite documentation just to keep AI systems accurate. Sphinx approaches the problem differently. Rather than replacing the AI agent itself, the company focuses on automatically building and maintaining the knowledge layer that sits beneath any agent, allowing organizations to keep business context current without dedicating teams to manual maintenance.
That approach also addresses another growing concern among enterprise buyers: flexibility. As new foundation models continue to emerge, many organizations worry about becoming locked into a single AI provider. Sphinx’s knowledge layer exists independently of whichever model a company chooses to deploy, allowing customers to adopt new technologies without recreating years of documentation or retraining their AI systems from scratch. In a market where model capabilities and inference costs change rapidly, Kodialam believes preserving organizational knowledge separately from the models themselves gives businesses a significant long-term advantage.
That vision recently attracted $9.5 million in seed funding from investors including Lightspeed Venture Partners and Bessemer Venture Partners. Much of the capital is being directed toward research, where Kodialam says Sphinx’s biggest challenge is not fundraising but attracting exceptional talent and continuing to improve the algorithms that learn from enterprise work product. The company recently entered production with enterprise customers and is already seeing organizations use the platform to improve analytics workflows on top of existing enterprise data environments such as Snowflake and Databricks. Sphinx remains intentionally lean, with a nine person team made up almost entirely of researchers and engineers with deep data science backgrounds. As Kodialam puts it, building products for data scientists requires people who understand data science firsthand.
Although Sphinx can be applied across industries, the company has seen particularly strong traction among financial services organizations and consumer brands with sophisticated analytics operations. What unites its customers is not the industry they operate in, but the pressure they face to make enterprise data accessible across their organizations without sacrificing accuracy, governance, or trust.
Increasingly, executives want every department to interact directly with enterprise data using AI. Marketing teams want forecasting. Finance teams want instant reporting. Product organizations want immediate analytics without waiting on specialized data scientists. Yet those ambitions often create a new bottleneck, placing enormous pressure on data organizations responsible for ensuring every answer remains accurate. Sphinx is designed to remove that burden by allowing AI systems to inherit organizational knowledge without requiring data scientists to manually document and maintain it themselves.
Enterprise adoption is also about governance. As organizations deploy AI across multiple teams, chief data officers increasingly need confidence that employees asking identical business questions will receive consistent answers. If executives receive conflicting reports or inaccurate metrics, responsibility ultimately falls on the data organization. Rather than allowing organizational knowledge to become fragmented across multiple AI systems, Sphinx provides a centralized knowledge layer that gives enterprises greater visibility into how AI arrives at its conclusions while maintaining consistency across the business.
That philosophy also shapes how Kodialam thinks about the future of data science itself. He describes the profession as evolving through distinct eras. Years ago, data scientists primarily built custom machine learning models. As mature libraries standardized much of that work, their focus shifted toward preparing and curating data. Now, he believes the next transition is already underway. As AI automates more technical implementation, the most valuable skill will become curating the context that allows intelligent systems to make sound decisions. Rather than disappearing, data scientists will become stewards of organizational knowledge, ensuring AI understands not just the data, but the business behind it.
“I think the role becomes knowledge curation,” Kodialam says. “Making sure agents have the context they need to solve problems that data scientists used to solve manually.”
Another distinguishing feature of Sphinx is its emphasis on interpretability. While many companies are specializing AI models through additional training or reinforcement learning, Kodialam argues those approaches often become opaque black boxes that business users cannot understand or control. Sphinx instead prioritizes making enterprise AI both auditable and transparent, allowing organizations to inspect how business knowledge is being represented and applied. For chief data officers, that transparency is more than a technical preference. As organizations rely on AI to inform increasingly important decisions, maintaining confidence in the numbers presented to executive leadership becomes a business imperative. Kodialam believes trust will come not only from better accuracy, but from giving humans visibility and control over how AI reaches its conclusions.
Despite the technical complexity of the product, Sphinx has remained founder led in its go-to-market strategy. The company currently has no sales team, preferring to work closely with early customers to refine the platform before expanding outbound efforts. With larger enterprise pilots underway and growing interest from organizations seeking to operationalize AI responsibly, Kodialam says the company’s focus remains on helping enterprises move beyond experimentation toward AI systems that genuinely understand how their businesses operate.
For Kodialam, the opportunity extends well beyond improving today’s AI assistants. He believes enterprises shouldn’t have to rebuild their institutional knowledge every time a new AI model arrives. If foundation models continue improving at today’s pace, the organizations that win won’t necessarily be those tied to a single provider. They’ll be the ones that own their knowledge independently of the models they use. As enterprises race to deploy AI across every function, Sphinx is betting that the future of enterprise AI won’t belong solely to the smartest model, but to the infrastructure that allows businesses to preserve their knowledge, remain adaptable, and confidently evolve alongside the next generation of AI.
