IBM is launching Bob, an AI development platform designed to accelerate software delivery without separating coding assistance from cost, security and governance. It works across the software-development lifecycle, coordinating specialised agents for architecture, modernisation, testing, documentation and deployment.
The product addresses a tension familiar to large engineering organisations. Code can now be generated in seconds, but legacy dependencies, compliance and quality assurance still require careful understanding. Speed without context can increase technical debt rather than remove it.
IBM says 60–80% of engineering budgets can be consumed by modernising older systems. Bob focuses heavily on that expensive work.
Legacy modernisation starts with dependency mapping
Mainframe and older .NET applications cannot be safely updated through isolated code suggestions. Their logic connects deeply with databases, jobs and undocumented business processes.
Bob maps architecture before refactoring and coordinates agents across related tasks. APIS IT used it on government systems spanning mainframe and .NET environments, producing architecture analysis and documentation ten times faster with reported 100% accuracy on legacy JCL and PL/I systems. It also migrated complex .NET services in hours rather than weeks.
Blue Pearl reduced a typical 30-day Java upgrade to three days, saving more than 160 engineering hours and reporting no defects after deployment.
These examples are promising but organisation-specific. Successful modernisation still depends on representative testing, complete context and expert review of behaviour that documentation may not capture.
Multi-model routing connects cost with task value
Bob chooses among Anthropic Claude, Mistral models, IBM Granite and specialised systems. Simple completion can go to a lighter model, while architectural reasoning uses a more capable and expensive option.
Routing considers accuracy, latency and cost, helping leaders avoid paying frontier-model rates for every task. Pass-through pricing provides visibility into usage and links AI expenditure more closely to production work.
The platform also normalises prompts, scans sensitive data, enforces policies and runs automated red-teaming. Human approval points can be configured by task, with lower-risk steps automated and consequential changes held for review.
BobShell records agent processes and code modifications from creation to deployment. That traceability is essential when AI-generated code enters regulated systems and teams need to show who approved a change.
Productivity claims depend on governed adoption
IBM began testing Bob with 100 developers in June 2025 and now reports more than 80,000 internal users. Surveyed employees cite an average 45% productivity improvement across feature work, remediation and modernisation.
The Maximo team reported 69% time savings in complex refactoring, while Instana reduced time on certain assignments by 70%, saving approximately ten hours per week.
These figures show potential, but organisations should measure defects, rework and maintainability alongside completion speed. Faster delivery is harmful if generated code bypasses review or creates architecture that teams cannot support.
Bob is available as software as a service with a 30-day trial and individual or enterprise tiers. Companies needing on-premises deployment for residency or compliance will have to wait for a planned version.
IBM’s approach treats the development assistant as an engineering platform rather than a conversational add-on. Persona-based modes, tool calls and approval gates embed AI across roles while preserving organisational standards.
The strongest business case is not the volume of generated code. It is shortening expensive modernisation while making each automated decision traceable and routing compute according to the work’s value.
If Bob maintains that control as adoption grows, enterprises can use AI to reduce technical debt without creating a new layer of unmanaged code, cost and security risk. Teams should compare every claimed time saving with post-release defects and future maintenance effort. That is the difference between coding faster and improving the complete software-delivery system.
