Enterprises are racing to adopt AI, but much of the data those tools need is still trapped inside legacy systems, proprietary software and applications that lack modern interfaces. Ran Cherniavsky, CEO and co-founder of Unfold, says that gap has become the central problem for enterprise AI, and his company is built to close it.
Why don't traditional integration tools work for AI?
Traditional integration tools connected systems, but they did so narrowly. They moved data one field at a time, through whatever the vendor chose to expose. Cherniavsky said that model was never designed for AI.
“AI needs access to everything a human being can see in a system, and an understanding of what it means,” he said. “And it needs it fast, without a project for every field. That’s a different problem.”
The costs show up well before an AI project launches. Companies hire specialized consultants to map a system’s fields by hand. Analysts pull capped reports and stitch them together in Excel. AI pilots stall because the model sees only the slice the API exposes, and then it gets the answers wrong.
“In effect, you’re renting access to your own data through someone else’s UI,” Cherniavsky said.
He also pointed to security. If no one can read what sits inside a system, no one can be confident it is secure. Cherniavsky said that risk is often invisible until it causes a problem.
Cherniavsky said AI adoption is where the pressure is strongest right now, and where locked data costs companies the most. Security and modernization face the same underlying problem, he argued. Citing Gartner, he noted that 45% of core systems still run on mainframes or on-premises client/server architecture, and that 34% of organizations now leave core applications alone because modernizing them is too risky. Solving data access and context for AI, he said, goes a long way toward solving the other two.
Unfold works the way an employee does. Rather than relying only on an API, it uses the user interface, reports, exports and files that an application already exposes. The company studies each system, maps where every piece of data lives and how it moves, and then builds a deterministic, read-only collector that retrieves the data through the system’s own permissions. Cherniavsky said the approach works on AS/400 mainframes, on-premises ERPs and modern SaaS products that expose only part of their data. Onboarding takes up to 14 days per system.
Why does enterprise data need context before AI can use it?
Access alone is not enough, Cherniavsky said. Raw data without context is noise. In Oracle’s JD Edwards ERP, for example, tables store numbers instead of names, and status fields hold codes that few people remember defining. Two systems may refer to the same customer by different names.
Unfold learns each system the way a trained employee would, identifying what each object means, how objects relate, which business rules and exceptions apply and how people actually use the software. The company builds a catalog of every data object, with its business meaning and its full path through the system. Cherniavsky said that knowledge often lived with subject-matter experts, implementation partners and vendors, and much of it left when those people did.
Once each system is understood, Unfold maps the data into one canonical enterprise schema. Cherniavsky offered an example: the same customer is identified as KUNNR in SAP and as an Account in Salesforce. The company keeps the data in the customer’s own source systems and tracks lineage, so it knows where each field came from and how it was transformed.
Cherniavsky said Unfold does not compete with data platforms such as Splunk, Snowflake and Databricks. “We don’t replace data platforms; we feed them,” he said. The company can deliver data to a customer’s data lake, warehouse or AI agents, including through a connector to Claude and Gemini.
Cherniavsky described several customer uses. A global retailer needs point-of-sale data from AS/400 systems in more than ten countries to feed its fraud detection, which now runs on Gemini Enterprise. One country’s regulator ordered the company to stop selling until that control was in place. A top-10 U.S. insurer had contractors map claims-system fields to its database by hand, followed by engineers building the ETL process. Cherniavsky said that work took weeks and that Unfold completes it automatically in days.
A multistore auto dealership group had its parts, service and warranty data locked behind reports in its dealership management system. Now, its owner can ask in plain English how old the warranty money owed to the business is.
Cherniavsky said customers measure value in three ways: time, cost and outcomes that were not possible before. Time means days instead of months. Cost includes the consultants, professional services and engineering that are no longer needed. The outcomes include answers that span multiple systems.
Gartner says that through 2030, the biggest compounding challenge for agentic AI will be data access and control. The research firm also treats AI context platforms as a separate market, growing from $28 billion in 2026 to $78 billion by 2030, and notes that those platforms are limited by how much they can extract from systems of record.
How should companies start giving AI access to locked data?
Cherniavsky’s advice to companies is to start with a business outcome, not a data project. Pick one AI initiative blocked by a system that cannot be read, solve that problem, then expand system by system. Companies should not wait for modernization projects to finish first, he said, and they should treat the knowledge inside their systems as an asset before the people who understand those systems leave.
Cet article a été rédigé avec l'assistance de l'IA.
News Factory APP - actualités agentiques pour booster votre SEO et AEO.