Generative AI can create test ideas in seconds. That speed is useful and dangerous for the same reason: a long list looks like coverage.
Good testing is not measured by the number of cases produced. It depends on understanding product risk, choosing revealing conditions, observing behaviour and deciding whether evidence is sufficient.
Give AI the risk context
Do not prompt “write tests for login”. Provide user goal, architecture, recent changes, known failure modes, security boundaries and acceptance criteria. Ask the model to identify assumptions and missing information before generating cases.
The Certified Agile Testing Foundation Professional provides the testing foundation needed to use AI as an accelerator rather than a substitute for test thinking.
Use AI for divergence
GenAI is useful for expanding possibilities: unusual sequences, data combinations, boundary cases, accessibility scenarios or alternative user behaviours.
The tester then converges. Remove duplicates, prioritise by risk and challenge cases that are plausible but irrelevant.
Keep expected behaviour grounded in evidence
Do not let a model invent the oracle. Expected results should trace to requirements, product rules, standards or validated behaviour.
When requirements are ambiguous, the ambiguity is a product question—not an invitation for AI to choose a convenient answer.
Protect sensitive test information
Source code, production data, logs and vulnerability details may be confidential. Use approved tools and sanitised or synthetic data where possible.
General AI literacy such as the Certified AI Literacy Professional can help teams understand data, verification and responsible-use boundaries.
Review generated automation like code
AI-generated test scripts can contain brittle selectors, weak assertions, hidden dependencies or insecure patterns. Review, run and maintain them under normal engineering controls.
Fast generation does not reduce ownership of the test suite.
Close the feedback loop
Record which generated ideas found meaningful defects, which were noise and what context improved the result. Build a small library of proven prompt patterns and evaluation examples rather than starting from scratch each sprint.
The Certified Software Quality Analyst supports the broader discipline of turning test activity into credible quality evidence.
A sprint workflow
At refinement, use AI to challenge assumptions and identify risk questions. During test design, expand cases. Before execution, human reviewers prioritise and verify oracles. During automation, review generated code. At retrospective, capture where AI helped or created rework.
Project leaders can connect this workflow to delivery governance through the Certified Software Project Manager pathway.
Use AI as a reviewer of the test model
Ask the model to challenge the team’s coverage: Which assumptions are untested? Which user groups or failure states are missing? What changes in this sprint create regression risk? Then require the tester to accept, reject or amend each suggestion.
This creates useful cognitive diversity without pretending that the model knows the product better than the team.
Measure net testing value
Track time saved, useful test ideas accepted, defects found from AI-assisted exploration and rework caused by incorrect generated tests. If generation creates large review queues, the team may be optimising output volume rather than testing effectiveness.
Review the evidence after several sprints and keep only the uses that improve detection, speed or clarity without weakening quality.
Final takeaway
Generative AI should widen a tester’s search space, not decide what quality means. Give it context, use it to generate possibilities, keep expected behaviour grounded in evidence and retain human risk judgement.
The goal is not AI-generated testing. It is better testing with AI.
Responses