Ran Cherniavsky, CEO and co-founder of Unfold, says enterprise AI stalls because critical data remains trapped in legacy systems and proprietary applications without modern interfaces. Traditional integration connects one system to another field by field through vendor-exposed APIs, while AI needs everything a human can see plus its business meaning. Unfold works through interfaces businesses already use and completes onboarding in up to 14 days, a fact that matters because 45% of core systems still run on mainframes or on-prem client/server.
How Unfold extracts data without APIs
Unfold accesses closed systems the same way employees do, through the UI, reports, exports, files and command line, with or without an API. The company spent years integrating systems manually, a process that took months per system, and now uses AI to become a subject-matter expert on a system in days. It maps where each piece of data lives and how it moves, then builds a deterministic read-only collector that extracts data through the system own permissions. The approach covers mainframes like AS/400, on-prem ERPs and modern SaaS that exposes only part of its data.
Raw extraction is paired with context reconstruction, because tables in Oracle ERP JD Edwards carry numbers instead of names and status fields hold codes few remember. Unfold learns each system like a trained employee, linking what a user sees on screen to where it lives in the data, including object relations, business rules, exceptions and actual usage. The result is a catalog of every data object with business meaning and its full journey through the system. Once systems are understood, Unfold maps them into one canonical enterprise schema, so KUNNR in SAP and Account in Salesforce arrive as the same customer.
The timing reflects pressure from boards to deploy AI against core operational data that APIs do not expose. According to Gartner cited in the interview, 45% of core systems still run on mainframes or on-prem client/server, while 34% of organizations leave core applications as they are because modernization is too risky. Cherniavsky argues AI adoption suffers most, while security suffers as a symptom because teams cannot secure what they cannot see. Companies pay specialized consultants year after year to map fields by hand, analysts stitch capped reports in Excel, and AI pilots fail when models see only the API slice.
What this means for enterprise AI projects
For companies adopting AI agents, the change is access to data that never reached analytics platforms and the ability to keep it where it lives. Unfold feeds Splunk, Snowflake and Databricks rather than replacing them, or delivers data straight to agents through connectors to Claude, Gemini and others. At a global retailer, point-of-sale data from AS/400 systems in more than ten countries feeds a Gemini Enterprise fraud-detection agent, a control one regulator required before sales could continue. Small firms gain answers across systems without long integration projects, while large firms can connect plants, stores and ERPs under one schema.
The limits center on governance, determinism and verification of context mapping. Unfold keeps data in customer source systems, attaches lineage showing where each field came from and how it was transformed, and uses deterministic collectors that return the same data for the same question. That design matters before letting an agent act, yet buyers still need to check permission handling, coverage of every interface and accuracy of business-rule mapping. The interview does not by itself prove performance across all system types, so questions about audit, exception handling and maintenance of the catalog belong in vendor evaluation.
The marker to watch is whether Unfold cases move from pilots to production controls in regulated operations. Examples to follow include the retailer fraud-detection rollout across ten countries, the top-10 US insurer replacing weeks of manual claims-field mapping and ETL work, and the dealership group asking in plain English about warranty money owed. If such deployments sustain deterministic answers across SAP, Salesforce, JD Edwards and AS/400 sources, locked-system data will become standard AI input.
