In this paper we argue that enterprises do not always have to replace reliable legacy systems so that they can become AI-ready. Mainframes, AS/400 systems and COBOL applications still matter, but lots are siloed and have no modern APIs. The path of preference is modernization in place: building a controlled integration and intelligence layer around existing systems rather than going for costly replacement.
AI is only as good as the data it can access and the context it can work in. Where the legacy system offers a viable API, it is generally preferable to keep it and add AI on top. A system that has no reasonable programmatic access to its data may be a candidate for replacement.
The layered approach is comprised of integration, automation, decisioning and experience. Access to legacy data and actions can be achieved via APIs, event connections or RPA. Automating repetitive work, but people in charge. Predictive models rank options, score risk and recommend actions. Intelligence is embedded into the tools employees already use.
Four patterns are described: an API abstraction layer to protect old code; an AI sidecar to isolate prompts, retrieval, guardrails, and validation; a wrapper to allow AI agents to use RPA to retrieve legacy data and structure old outputs; and an orchestration layer to coordinate systems, models, context, and validation.
It’s all about governance. Allowing AI direct access to payroll, customer, financial or production systems is dangerous. Agents should employ narrow, auditable tools with limited permissions. The intention is to map processes, create guardrails, start with smaller, high-value sub-processes, create a governed layer of integration and then scale with role-based access and audit trails.
The document talks about faster queries, savings, faster movement from AI pilot to production and lower exception handling time. It concludes that legacy systems can stay the operational backbone, with a thin, controlled layer of intelligence connecting old infrastructure to modern AI.
