Accelerating the software delivery lifecycle with generative AI: Lessons from IBM's Enterprise Payment Services

Three months ago, my team’s regression test backlog and execution time were a persistent drag on every sprint. Today, Bob has reduced our testing effort by more than 70%. That outcome was achieved by deploying IBM Bob™, IBM’s AI development assistant natively embedded in the testing workflow, as a full member of our testing team. It proves that AI-augmented software delivery is no longer a future concept.

Software engineering teams are facing growing pressure from multiple directions simultaneously. Regulatory bodies are tightening compliance requirements. Security auditors are demanding greater visibility into vulnerability management and remediation timelines. Meanwhile, clients are accelerating their functional demands, expecting faster time-to-market without impact to quality or stability.

The IBM Payment Center® Enterprise Payment Services (EPS) team operates a highly regulated, high-value payments platform, and we live with those pressures every day. Payments platforms sit at the intersection of every one of them: financial regulation, operational resilience mandates, security scrutiny and client competitive demands all converge in a single codebase.

The question I set out to answer this year is straightforward: how do we do more, faster, with the same people, without sacrificing quality or stability?

The answer is generative AI, specifically IBM Bob, which we onboarded in early 2026 as a full member of the EPS engineering team. We have achieved significant progress in just a few months. Those results mirror a broader transformation happening across the industry.

Generative AI is no longer a pilot experiment confined to autocomplete suggestions. It is being woven into every major discipline of the Software Development Lifecycle (SDLC), from requirements gathering to production incident response. The results are compelling enough that organizations not yet investing will experience a real competitive disadvantage.

Our experience maps cleanly onto the full arc of software delivery. Let me walk through each phase.

Requirements: Turning noise into structure
 

Requirements gathering has always been the phase with the most ambiguity and the greatest potential for downstream costs. Poor requirements are the root cause of a disproportionate share of project rework, missed scope and defect injection. A defect found in requirements costs a fraction of one found in production.

Generative AI fundamentally changes the economics of this phase. In the EPS context, Bob handles:

• Rapid user story and documentation generation—transforming meeting notes and stakeholder conversations directly into structured, Jira-ready user stories and acceptance criteria.
• Identification of missing requirements—analyzing existing stories against business context and flagging gaps that human business analysts might not catch until late in the sprint.
• Translation of unstructured inputs—converting transcripts, emails and informal notes into formal requirements artifacts.

We have observed 30–40% productivity improvements in the requirements phase when using Bob to assist business analysts, primarily through the elimination of manual transcription, reformatting and cross-referencing work. These team-measured estimates are based on tracked business analyst effort before and after Bob’s addition to the team.

The key enabler is Bob’s contextual awareness. We loaded it with the platform’s full documentation set, regulatory standards, coding standards and non-functional requirements. This grounding ensures that generated requirements reflect the architectural and regulatory realities of a payments platform, rather than being merely syntactically correct.

Engineering and development: The biggest opportunity
 

Development is where the largest share of engineering hours is spent and correspondingly it is where generative AI offers the most transformational upside. IBM’s projections for Bob in the engineering phase range from 30–50% efficiency improvements—a range that reflects the breadth of engineering tasks and the depth of context required for each.

In our implementation, we are pursuing engineering acceleration across three vectors. In each, Bob operates in semi-autonomous mode, performing the work while a subject-matter expert reviews and validates the output before acceptance. This pairing is not a limitation; it is the correct architectural pattern for AI at this stage of maturity.

1. Bob-assisted code reviews: Bob was given access to the EPS source repository and performs code reviews alongside human SME reviewers. This dual-review approach ensures adherence to regulatory requirements, security and privacy standards and functional and non-functional architecture requirements—catching defects during development rather than in testing or production. GitHub’s research on Copilot found that AI-assisted code review caught 15–20% more security-relevant issues per pull request cycle and reduced time-to-resolution because the AI proposed corrective diffs alongside its findings.
2. Vulnerability management and upgrades: As language and library versions evolve and regulatory requirements change, Bob evaluates the blast radius of a library upgrade across the codebase. It identifies incompatible call sites and recommends necessary modifications in minutes, work that previously required senior engineers to devote days on manual dependency tracing.
3. Code generation and modification: Bob provides recommendations for modifying existing code given clear requirements (functional and non-functional), codebase context and relevant regulatory constraints. An AI generating payment processing code without awareness of data residency rules, error-handling standards or regulatory obligations is a liability, not an asset. Loading Bob with that context upfront addresses this challenge directly.

Testing: Eliminating the bottleneck
 

Testing has historically been the pacing constraint in software delivery. Development teams produce code faster than QA teams can validate it. This asymmetry is the primary driver of sprint carryover, release delays and regression testing backlog growth.

Our QA team’s results are the most mature and the most striking. Three months after deploying Bob as a test automation generator for the EPS Wires payment platform, we’ve achieved:

• 70% reduction in test automation creation 
• 90% reduction in regression test backlog
• 80% reduction in regression execution cycle time

These results are not marginal improvements. They represent a structural transformation in how the team operates. QA engineers have shifted from writing boilerplate test automation to reviewing and refining AI-generated tests—a fundamentally higher-leverage activity. The CI/CD pipeline integration means regression failures surface within hours of a commit, not days, dramatically reducing the cost of a defective change.

This experience is consistent with broader industry observations. Research on organizations implementing generative AI for test generation has found regression cycle time reductions averaging 60%, with high-performing implementations, such as ours, achieving even greater gains. The consistent differentiator is context: AI test generators with access to the full codebase, existing test cases and business requirements produce dramatically better tests than systems operating on code snippets alone.

IBM Bob, loaded with EPS’s full GitHub repository, regulatory standards and functional documentation, exemplifies this contextually grounded approach.

Design and architecture: Proactive improvement
 

Generative AI is beginning to play a role in proactive design intelligence. It is moving beyond the reactive cycle of building and testing what has already been specified. Our roadmap includes enabling Bob to:

• Identify unrecognized performance bottlenecks and architectural inefficiencies that have accumulated over time.
• Evaluate and prioritize technical debt removal by correlating code structure with incident history and performance data.
• Generate design recommendations for new functional enhancements while considering existing architecture, regulatory constraints and non-functional requirements.

This approach is a more sophisticated use case than code modification or test automation and it requires a higher level of contextual trust in Bob’s recommendations. Keeping human architects in the loop to validate recommendations before acting on them is the right posture for this phase of AI’s maturity.

The value is significant: a system that has ingested the full codebase, its documented architecture and years of design decisions can surface insights that would otherwise require weeks of architectural research.

Level 3 support: Closing the loop
 

The most operationally impactful AI use case on our roadmap is incident resolution acceleration. L3 support for a high-value payments platform is a high-stakes, time-sensitive activity. When a production incident occurs, the time required to identify and resolve the root cause directly determines client impact.

Bob’s SDLC deployment roadmap includes analyzing L3 support incidents, correlating reported symptoms with code paths in the repository, identifying likely root causes and proposing corrective actions. The same context that makes Bob effective as a code reviewer—full knowledge of the codebase, regulatory requirements and error-handling patterns—makes it uniquely suited to this task.

A human L3 engineer augmented by Bob can scan the entire codebase in seconds for patterns that match an incident signature, accelerating incident resolution and minimizing client impact.

None of this productivity comes without risk. Generative AI systems can hallucinate. They can produce plausible-looking but incorrect outputs with apparent confidence. In a payments platform operating under regulatory scrutiny, an unreviewed AI-generated code change or an incorrectly specified test case can have serious downstream consequences.

Our risk management framework is built around the semi-autonomous model: Bob is deployed as a deeply integrated team member across every SDLC phase, always paired with a human reviewer who validates output before it is accepted. Business analysts review Bob-generated requirements. Code owners review Bob-identified bugs and proposed fixes. QA engineers review Bob-generated test cases. L3 engineers will review Bob’s incident analyses.

This approach is not a limitation. It is the correct architectural pattern for AI at this stage of maturity. The model trades a fraction of the theoretical efficiency ceiling for a substantial reduction in error risk and it does so in a way that allows trust to be built incrementally.

As Bob’s outputs prove reliable in each domain, the level of human scrutiny can be appropriately calibrated. My expectation is that this semi-autonomous configuration will remain the norm for several years, with Bob continuing to learn from its work experience and produce incrementally higher-quality outputs over time.

Across requirements, engineering and testing, IBM projects 25% to 70% efficiency improvement from Bob, depending on the phase and specific task. Even if outcomes settle at the conservative end of each range, the compounded effect across the team is equivalent to a 25–40% increase in effective team capacity.

To illustrate the math concretely, consider a 10-person team that achieves a blended 30% efficiency gain across all SDLC phases. The effective output is equivalent to 13 engineers working at the original pace, without adding headcount or increasing cost proportionally.

For an organization facing accelerating demand on a fixed budget that is not an incremental improvement. It is a capability multiplier—one that allows the same team to deliver more features, maintain higher quality, respond faster to incidents and absorb increasing regulatory obligations without proportional cost growth.

Our experience surfaces three questions that every engineering leader should be working through right now:

1. How do you contextualize AI for your environment? The EPS approach of front-loading Bob with the codebase, standards, regulatory requirements and documentation is the foundation for everything that follows. An AI assistant operating without this context is less valuable—and potentially riskier—than one that has been properly grounded.
2. Where do you start? Testing is the fastest path to measurable ROI, as our results demonstrate. The combination of clear success criteria, a relatively low risk threshold and high manual effort in the status quo makes it the ideal beachhead for any team beginning this journey.
3. How do you manage trust incrementally? The semi-autonomous model—AI with a human in the loop—is not a compromise. It is a responsible implementation pattern that allows organizations to capture value while building the institutional confidence required to expand AI’s role over time.

My team’s journey from a manual, backlog-constrained QA function to an AI-augmented, CI/CD-integrated quality organization in three months demonstrates that this transformation is achievable at pace. The broader lesson is clear: generative AI embedded in the SDLC is not a future pilot program. It is already delivering measurable results for teams that have made the investment. The time to build that foundation is now.