Who owns the bug when AI writes the code?

AI-assisted development is moving faster than many companies’ governance systems. Software engineering offers an early test of who remains accountable when machine-generated work enters production.

At Paisabazaar, AI now generates 86% of new code. The company has responded by increasing the pace at which it can build and release products, rather than reducing its engineering workforce. Yet the code does not move from an AI tool directly into production.

Developers review it, after which it passes through code review, automated testing, quality assurance, security checks and user acceptance testing. The process reveals something important about the next phase of software development. Generating code may be becoming easier, but taking responsibility for what that code does remains a human task.


Also read: Future of Knowledge Work: Why AI skills are becoming essential beyond the IT department

The distinction becomes harder to manage as the volume of machine-generated software grows. A developer may use an AI system to produce a function, modify an existing application, or generate test cases. The resulting code can look polished and perform as expected under the conditions it was tested for. A flaw may still sit inside it, waiting for a different input, a security breach, or a production environment that exposes an assumption the model made during generation.

Engineering teams have long dealt with imperfect code. What changes with AI is the distance between authorship and accountability.

A developer who writes a piece of software has some understanding of why it was written a certain way. AI-generated code can introduce logic that works without being fully understood by the person approving it. The risk becomes particularly relevant when junior engineers rely heavily on coding assistants.

Paisabazaar CTO Mukesh Sharma has warned that while AI can make a fresher’s output resemble that of a much more experienced engineer, excessive dependence on these tools can weaken the ability to understand code and the fundamentals behind it.

The concern is reflected in broader industry data. An Economic Times report on AI-assisted coding found that about 94% of developers surveyed in India use AI coding tools every day.

Nearly all technology leaders surveyed said AI-generated code undergoes peer review before production, while 92% flagged risks in deploying such code without human oversight, particularly around maintainability and security. Those numbers suggest companies already recognise the need for a human checkpoint. The harder question is whether the checkpoint is strong enough.

Governance becomes complicated when AI systems begin operating across larger parts of the development lifecycle. ET had reported concerns from technology and security leaders around inadequate visibility into AI workloads, data exposure and vulnerabilities, alongside the need for stronger review of AI-generated code and security architecture as AI agents connect with enterprise systems.

The answer cannot simply be another approval box. An engineer who approves code without understanding its logic may satisfy a process while leaving the underlying accountability problem unresolved. A useful governance system has to make responsibility visible: who reviewed the output, what was tested, what restrictions applied, and who had the authority to release it.

Software engineering may be one of the earliest places where organisations confront that question at scale. The same issue will eventually extend into other forms of knowledge work as AI moves from assisting employees to producing parts of the work itself.

For business leaders, the question is therefore becoming more specific. When a machine contributes materially to an outcome, how much human understanding must exist before someone can legitimately take responsibility for it?

The Future of Knowledge Work Summit 2026, taking place on 19 November 2026 in Mumbai, will examine the changing relationship between AI, work and organisational accountability. Software offers a particularly clear place to begin the discussion because the consequences of a bad decision can be traced directly into the systems businesses depend on. The larger challenge is deciding whether the structures surrounding those decisions are evolving quickly enough.

Register Now!