For most of my career, software quality rested on a familiar set of questions. Does the application work? Is it secure? Can it handle the expected traffic? Can another developer maintain it without pulling the system apart? Those questions still matter, but they’re no longer enough.
AI has changed how software is produced. Developers can now generate features, tests, integrations, and documentation in a fraction of the time once required. Coding agents can complete entire work items with limited human involvement. That speed has created an uncomfortable possibility. A team can produce more software, close more tickets, and release more frequently while the quality of its product declines.
I believe software quality now has to begin with a different question: Does the system produce the right result for the business and the people who depend on it?
Working Code Isn’t the Same as Quality Software
AI-generated code can look remarkably polished. It may follow naming conventions, include comments, pass automated tests, and survive a code review. None of those signals proves that it handles the right business rules or responds correctly when real-world conditions become messy.
Research published in March 2026 examined 304,362 verified AI-authored commits across 6,275 GitHub repositories. The researchers identified 484,606 issues introduced by AI-generated changes. Code smells accounted for 89.1 percent of them, and 24.2 percent of the tracked issues were still present in the latest version of the repository. More than 15 percent of commits produced by every AI assistant studied introduced at least one issue. The findings suggest that AI can create maintenance costs that remain long after the initial burst of productivity has passed.
A January 2026 study reached a related conclusion after examining AI-generated build code. Researchers found 364 maintainability and security issues, including missing error handling and hardcoded paths or URLs. More than 61 percent of the agent-generated pull requests were approved and merged with little human intervention. The code entered the system, but approval didn’t establish that the code was dependable.
This is where our traditional quality signals begin to weaken. A green test suite may show that the software behaved as expected under the conditions we anticipated. It can’t tell us whether our expectations were complete. A successful deployment tells us that the code reached production. It doesn’t tell us whether the software helped a customer finish the work they came to do.
Quality Has Become a Business Outcome
In the age of AI, I define software quality across five connected dimensions: correctness, security, maintainability, business performance, and trust. If one dimension fails, the product’s apparent success can conceal a growing liability.
Correctness means more than producing a technically valid response. The result has to reflect the company’s rules, the user’s circumstances, and the intent behind the request. Security now includes the behavior of models, agents, tools, data connections, and third-party services. Maintainability requires code that engineers can understand and change, even when no person wrote the first version.
Business performance asks whether the software completes valuable work. Trust measures whether customers and employees are willing to rely on it. A system that receives thousands of interactions but forces users to confirm every answer isn’t delivering quality. It’s transferring validation work to the customer.
A March 2026 synthesis of 24 empirical studies found that the quality of AI-generated code depends on prompt design, task specifications, developer expertise, the AI system, and the way people interact with it. Results varied across correctness, security, maintainability, and complexity. The researchers’ conclusion matters because it moves quality away from the model alone. Quality emerges from the full human and technical system surrounding the model.
Human review can’t be assumed either. A May 2026 analysis of AI-generated pull requests found that most received no review. When reviews occurred, AI agents performed much of the work. Human involvement frequently appeared as instructions to another agent instead of an independent evaluation of the change. A review count may therefore create the appearance of oversight without proving that a person examined the code.
This doesn’t mean companies should slow AI adoption. It means validation has to advance at the same pace as generation. We need tests tied to real requirements, review standards that identify who or what performed the review, production monitoring connected to customer outcomes, and clear ownership when the system fails.
I no longer think quality can be inspected into software at the end of development. It has to be defined before the AI begins, measured while the system operates, and challenged whenever the business conditions change.
AI can help us create more software than ever. The companies that win won’t be those that generate the most code. They’ll be the ones that can prove their software deserves to be trusted.
