AI agents dominate enterprise tech, running in 96 percent of companies. Yet, most still rely on shallow web searches. Pointing a language model at the open internet is the default shortcut. It’s fast, but it only scratches the surface.
Sixtyfour’s agents take a different route. They answer questions tested against hand-assembled, real-world data. “Every answer has to show its work,” says Saarth Shah, builder of their research agents. If an answer can’t be traced to a filing or record, it doesn’t ship.
“The models keep getting better, and they still only see what a person with a browser can see,” Shah adds. This limitation caps the agent’s understanding at open web knowledge, excluding licensed or proprietary data hidden below the surface.
Structured data analysis remains a bigger challenge. Most AI models excel at generating text or images but stumble on tables and numbers. Fundamental, an AI startup, built NEXUS—a large tabular model designed specifically to tackle this. Trained on billions of tables from proprietary and public sources, NEXUS can read and analyze complex datasets.
NEXUS gained traction fast. Amazon Web Services embedded it into SageMaker in June 2026. CEO Jeremy Fraenkel points out the diversity and sensitivity of most data. “There are very few similarities between, for instance, a biology dataset and a financial one.” Custom models like NEXUS make sense in this fragmented landscape.
Meanwhile, Feedzai, Mastercard, and Google launched their own large tabular models in 2026. Google’s TabFM arrived in June, joining the wave of foundational tabular LTMs focusing on proprietary and synthetic datasets.
OpenAI rolled out ChatGPT Work in July 2026, a productivity agent powered by GPT-5.6. It bundles ChatGPT, Codex, and the Atlas web browser into one app. The company insists users control what the agent can access and when it needs approval. Yet OpenAI plans to retire Atlas by August 9, 2026, signaling a shift away from browser-based tools.
Data infrastructure plays a crucial role behind the scenes. MongoDB’s document model supports flexible structure and simplicity. Aram Shatakhtsyan credits MongoDB for keeping data in one place while controlling how structured it gets. Tomer Weiss at Tavily highlights tracking data freshness and user authentication to govern agent responses.
StatSocial’s Digital Twins showcase the power of data-driven AI. Their PeopleGraph models roughly 150 million U.S. adults. Digital Twins average a 3.3 point mean absolute error when compared to real-world surveys, a tight margin for social data.
Boris van Breugel, an AI researcher, puts it bluntly: “Most people don’t necessarily like to do data analysis, and these systems will be able to do it for them.” The future isn’t just about talking AI. It’s about proving AI decisions with facts and real data. The agents are here, but only the ones that show their work will last.
Based on
- The Eval Stack: Proving the agents are right instead of claiming It — thenextweb.com
- Digital-native startups are ditching rigid databases for their agentic stacks | VentureBeat — venturebeat.com
- StatSocial Launches AI Focus Groups Built on Real Audiences, Not Invented Personas | Markets Insider — markets.businessinsider.com
- Large Tabular Models Bring AI Analysis to Spreadsheets – IEEE Spectrum — spectrum.ieee.org
- OpenAI Releases A New ChatGPT Tool For All Things Work Related — engadget.com
