Artificial intelligence tools can now generate source code and automate tests. In development teams, this raises an obvious question: will the role of QA engineers — the specialists who systematically check software for errors and vulnerabilities — become dispensable in the medium term? A current guide on Golem.de takes up precisely this concern and reaches an opposing assessment: of all things, the AI boom could make the work of testing specialists more important.
For companies that have websites, portals or specialised applications operated, this is not an academic debate. It determines how budgets are distributed in projects and what assurances an agency or an internal team can still give at all.
What automation can achieve — and where it ends
Automated tests are nothing new in software development. What is new is how quickly test cases can now be generated from existing code or from requirements descriptions. This lowers the entry barrier: projects that were previously not tested at all for reasons of time are now able to achieve test coverage in the first place.
At the same time, the problem is merely shifted. A generated test initially checks what the code does anyway. Whether that is also what was intended from a business perspective is not established by this. This distinction between technical correctness and business correctness is the core of quality assurance — and it cannot be derived from the source code alone, because the business intent is not fully contained there.
More code means more need for review
If AI assistance increases the amount of code produced, the amount of code that nobody has read in detail also grows. This creates additional need for review in several places:
- Has the change had unintended side effects in adjacent modules?
- Are edge cases covered that were not mentioned in the task description?
- Are access rights, data protection requirements and accessibility still correct?
- Does the solution remain maintainable, or is technical debt being created that will become expensive later?
None of these points can be answered by the mere number of green test runs. High test coverage indicates how many lines of code were executed during testing — not whether the right questions were asked.
The role shifts from executing to assessing
For a long time, testing work consisted to a considerable extent of repetition: the same click paths after every release, the same forms, the same browsers. Precisely this portion is automatable, and AI tools accelerate it further. What remains is more demanding: assessing risks, setting priorities, deriving test cases from business processes, interpreting error patterns.
These activities are less manual and more analytical. They require someone to understand the application's domain — the ordering process, the funding logic, the editorial workflows in the content management system. Those who know these interrelationships recognise deviations that a generated test would never classify as a deviation because it lacks the benchmark for comparison.
New testing tasks through AI functions in the product
There is also a second development: AI is not only in the development process but increasingly in the product itself — as a search function, assistance system or text generator. Such functions do not behave deterministically, so they do not necessarily deliver the same output for the same input. Classic tests following the pattern "expected result equals actual result" are only of limited use here.
In these cases, quality assurance must work with ranges, samples and assessment criteria instead of fixed target values. This is methodologically new and requires people who can assess which answer is still acceptable and which is not. A role that is more likely to be enhanced than dismantled.
What this means for project planning
From our perspective, this gives rise to some practical consequences for clients:
- Do not automatically cut the testing budget. The fact that tests can be created faster does not mean that less testing effort is needed. The effort shifts into conception and assessment.
- Record acceptance criteria in writing. The more clearly it is described what the software should achieve from a business perspective, the better both automated and manual checks can be aligned with it.
- Involve the business departments. Those who use the process daily recognise deviations faster than any tool. Structured test phases with real users remain valuable.
- Build up regression tests. Automated checks that run on every change primarily safeguard existing functions — especially when code is created faster than before.
Not an either-or
The juxtaposition of "AI or test team" is misleading. A division of labour is more realistic: tools take over repetition, breadth of coverage and initial drafts of test cases. People decide what needs to be tested at all, where the risk lies and whether a result is business-viable.
Those who regard quality assurance as a pure cost block will be tempted to replace it with automation. Those who understand it as protection against expensive production errors will tend to expand it — only with a different focus than before.
Classification for web and software projects
For ongoing projects, this means: speed gains from AI assistance should partly flow back into testing quality instead of being fully recorded as time savings.
All that shifts is when errors come to light — from testing into live operation. When selecting service providers, it therefore becomes more relevant how structured their quality assurance work is and how traceably they document acceptance.