What should I learn before starting a manual AI and automation testing role?
Asked by
MellowCedar42
I'm starting a new position as a tester on an AI and Automation team in about two weeks, and I'd like to prepare as thoroughly as possible. I have a Software Engineering degree and some experience with programming, databases, software development, Python, and Java, but I'm relatively new to professional software testing. The team builds AI and Power Platform solutions, including Copilot Studio agents. Testing is currently mostly manual because the team is still developing its testing strategy and learning the strengths and limitations of these tools. They plan to introduce automated testing later, so I want to understand both practical manual testing and the foundations of automation. What topics, skills, and resources would be most valuable to focus on before I start?
Answers
Focus on testing nondeterministic behavior. With an AI agent, getting the expected answer once doesn’t prove that the system is reliable. You’ll need to run representative prompts repeatedly, test variations in wording and context, and look for consistency, incorrect answers, unsafe responses, prompt injection, failures when information is missing, and inappropriate escalation. Think in terms of test datasets, evaluation criteria, pass-rate thresholds, and tracking results over multiple runs rather than a simple pass/fail check.
Reply to this answer
Learn the basics of software testing first: requirements analysis, test planning, equivalence partitioning, boundary-value analysis, exploratory testing, regression testing, defect reporting, and risk-based prioritization. For AI systems, add evaluation concepts such as accuracy, relevance, groundedness, hallucination rate, robustness, toxicity, bias, latency, and cost. Define what acceptable behavior means before testing, then use a varied set of normal, ambiguous, adversarial, and out-of-scope inputs. Even low-code agents benefit from structured test cases, repeatable test data, clear evidence, and documented limitations.
MellowCedar42: The solutions are built with Copilot Studio, so I’m assuming I won’t be testing traditional machine-learning models directly. Would these ideas still apply if most of the work initially involves manually entering prompts and checking accuracy, clarity, and hallucinations?
AmberWalrus18: Yes. You may not train or inspect the underlying model, but you still need to evaluate the complete agent as a system: instructions, grounding data, topics, connectors, permissions, fallback behavior, and generated responses. Learning common AI benchmarks can also help you understand how quality and safety are measured statistically, even if your team creates simpler internal evaluation suites.
Reply to this answer
Add your answer
Answers are reviewed before publication.
MellowCedar42: That makes sense. I’ve started looking at evaluating response quality, but I’m also trying to build a broader testing foundation so the role isn’t limited to checking individual prompts.