AI in QA: Making suggestions, but humans make the final decision

How do we use AI in QA? Recently, AI has been making an increasingly prominent entrance into the testing field. During presentations, meetings, and roadmap discussions, the same question almost always comes up. As manual QA professionals with more than ten years of experience, we’ve learned one thing clearly: AI can provide support, but the final decision must remain with humans.

In this article, we’ll explore how AI can support manual testing while keeping human judgment at the center of quality decisions.

The role of AI in QA

AI can truly add value to day-to-day QA work. It can generate ideas faster than we can, suggest tests, improve test descriptions, and offer an alternative perspective on requirements. This can save a significant amount of time, especially in the early stages of analysis.

For example, when evaluating a new user registration flow, an AI model can propose fifteen possible test scenarios within seconds. Think of variations such as testing empty fields, an invalid email address format, or special characters. A manual QA engineer can then quickly determine which scenarios are actually relevant to the business context. In some cases, for example, the focus isn’t on irrelevant edge cases, but rather on issues such as the timing of an email confirmation link.

What AI doesn’t really understand, however, are things like:

• The business context behind a feature
• Technical limitations and hidden dependencies
• The dynamics of stand-ups and the pressure of deadlines
• Internal discussions within the team
• Users’ actual expectations and habits
• Lessons learned from previous bugs and incidents

AI can suggest twenty technically valid tests and still miss that one scenario that turns out to be crucial, such as a payment timeout issue that has already occurred in a previous release.

In our experience, the best decisions actually stem from these kinds of “invisible” details that no tool can fully capture. AI can provide a starting point and offer direction, but it’s still up to people to determine how to proceed.

Decision-making cannot be outsourced

It is essential that someone consciously takes responsibility for quality-related decisions. QA isn’t just about giving a final “OK” at the end of a sprint, but about countless small decisions that are made continuously, such as:

• What is critical and what is acceptable
• What needs to be tested thoroughly and what requires only basic coverage
• Which points need to be clarified first with the product owner or developers before testing begins
• Which risks are consciously accepted when time is limited

AI can make suggestions, but it cannot take responsibility for the consequences of those choices. A QA professional can. That is why decision-making should not be entirely delegated to a tool, no matter how advanced it is.

Working with AI

When we use AI in QA work, we treat it as a very fast colleague—not as an authority. We let AI make suggestions first, but we don’t give it the final say.

In practice, this means that we:

• Critically evaluate every suggestion
• Validate suggestions within the context of the project
• Adjust, remove, or rewrite anything that doesn’t fit
• Retain ultimate responsibility ourselves, not with the model

Suppose AI makes testing suggestions for a mobile payment feature and proposes testing “performance with 1,000 concurrent transactions.” A manual tester familiar with the application’s actual scale could replace this with “checking transaction times under poor network conditions (3G),” because that is the actual risk users face in practice.

Over the years, professional experience helps you distinguish between valuable insights and what ultimately turns out to be nothing more than “well-packaged noise.” AI lacks that instinct. You can only develop it by working on real projects, challenging releases, and unpredictable situations.

Why Boundaries Are Important

Imagine relying entirely on AI-generated test scenarios for a booking system that handles events with limited capacity. The model validates the successful booking flow, checks required fields, confirms the payment status, and even verifies the confirmation email. Everything seems fully covered. The release goes live, and the team is confident in its quality.

But the AI didn’t account for the complexity of real-world user behavior: two people trying to reserve the last available spot at the same time. For a fraction of a second, both users see the spot as available.

In production, this can cause various problems. The system might accept both bookings, display incorrect availability, or—conversely—block both users. The result is user frustration, after which the incident quickly spreads through support channels and social media.

When something goes wrong, who is responsible?

When a problem reaches the production environment, no one asks what suggestions the AI made. The questions that are asked then concern what QA checked, how risks were assessed, and who decided the product was ready for release.

Over the years, we’ve learned that:

• A well-formatted document does not automatically mean that a product is stable
• A long list of test scenarios does not guarantee that the right risks have been covered
• No tool, no matter how advanced, can replace human judgment and a sense of responsibility

Who determines what is truly important?

The line between “AI makes suggestions” and “QA makes decisions” is important for one simple reason: responsibility for the product and its users.

That’s why QA professionals choose to use AI as a support tool, not as a replacement for human judgment.

“AI proposes, QA decides” isn’t just a marketing slogan—it’s a way of working. It allows us to benefit from speed and scalability without sacrificing critical thinking. The tool generates more ideas; we determine which of them actually contribute to the quality of the product and the protection of users.

The most important skill for a QA professional today is therefore no longer writing the perfect prompt. It’s about the ability to look at the vast number of AI-generated suggestions and say: “This is noise, this is redundant, but this is exactly the scenario that’s going to save Friday’s release.”

Perhaps that is ultimately the real question behind everything AI can propose: who decides what’s truly important?

“So, who actually decided what you’re reading right now?”