A new system builds fake companies from nothing, and by its own tests, the fakes are now indistinguishable from real ones.
The generator described in a new arXiv paper takes an industry, a company size, a business model, a set of software tools, and a random seed, then invents a full company around them: employees, customers, sales deals, support tickets, call recordings, chats, and documents. Every piece points back to the same underlying entity graph, so a single fictional customer shows up consistently across a CRM, a support desk, and a call system, exported in the native formats of 66 different business products. A companion generator builds databases to match a list of business questions, planting the rows needed to answer each one along with deliberate near misses, then computing the correct answers automatically. Because there is no real company on either end, the team graded the output with its own scorecard: 28 statistical checks across five categories, plus a detector trained to sniff out synthetic records.
Across 23 generated companies, average realism scores rose from 60.3 to 99.1, and the detector's hit rate on synthetic records fell from 55.2 percent to zero, even on a seed it had never seen during development. That matters because testing CRM, support, and analytics software usually means either scrubbing real customer data or living with toy datasets that fall apart under scrutiny. A generator that produces large, internally consistent, privacy-free business data, complete with answer keys, is a genuinely useful plumbing fix for anyone benchmarking enterprise software.
The service is already live at console.era.eon.io, served over MCP and REST, and packaged as offline containers, which puts the burden of proof back on whoever claims their AI can spot synthetic data from the real thing.