The future of software testing depends on expert testers who understand both testing techniques and the business behind the software. AI handles repetitive, clearly defined tasks well, but fails at the edges: hallucinated requirements, unverifiable results, and unexpected human behavior. The skills that matter most are business knowledge, requirements engineering, prompt engineering, and communication.
Key Takeaways
- AI-generated output is correct often enough to look trustworthy, but the remaining faulty fraction, hallucinated references, invented requirements, fabricated test cases, causes real business damage if no human checks the result.
- A tester who only runs pre-written scripts without understanding the underlying business is directly replaceable by AI; domain knowledge is what makes a tester hard to automate away.
- Prompt engineering and requirements engineering are now core tester skills, because the fundamental problem of describing software behavior precisely and unambiguously has not changed in 30 years of the profession.
- Communication skills are a critical gap in tester education: presenting data honestly, managing stakeholder expectations, and writing clear prompts all depend on abilities that engineering curricula do not teach.
AI is here to stay, and it raises the value of expert testers
AI will not remove the need for testers. It shifts what testing has to be. Olivier Denoo, nearly 30 years in the field, sees AI as an opportunity for the testing community rather than a threat to it.
The reasoning is direct. AI is not a passing fad. There may be a bubble, and it may deflate at some point, but the technology itself will remain part of software work. Testers who learn to use it well will keep their place. Those who cannot will be exposed.
The exposure is specific. If your job is pressing the same button and running the same script that you did not even write, for twenty years, you are seriously at risk. That work is exactly what AI absorbs first.
Why AI-generated code creates more work for testers, not less
The rise of vibe coding produces software faster and controls its quality less. That gap is where expert testers become necessary.
When someone prompts an application into existence without being a developer, a set of questions goes unanswered. Does the code match the requirements? How does it evolve? Is it maintainable? Is it secure? Repositories can carry backdoors and Trojans, and a non-developer has no way to judge whether the generated code is sound.
Olivier expects the consequence to arrive within months to years, once enough AI-generated applications exist. Whether you put agents in charge of agents, LLMs checking LLMs, or a human in the loop, the defect volume rises. More defects mean more testers, and specifically more expert testers.
What “expert tester” actually means
An expert tester combines technique with business understanding. Both halves matter, and the second is the one most often missing.
The technical half is knowing how to test: applying techniques, knowing where to look, knowing how to approach a problem. The business half is understanding what sits behind the application and why it exists. Nobody writes software to show off a new platform or language. Code exists to serve a goal and help someone do a task.
Business understanding grows harder as systems connect to each other. Applications are intricately linked, and the business flows running across them become difficult to follow. You need people who understand how the whole process works end to end, not just how one component behaves.
Why “AI can test, so we can cut testers” is a dangerous conclusion
A manager who returns from a conference convinced that AI removes the need for testers has misread the technology. Acting on that belief leads toward a business catastrophe.
The trap is that AI looks smart. It always returns an answer, because it is built to return an answer. It agrees with you, tells you that you are right, and presents its output with confidence.
The trouble is that this answer looks smart at first sight, but probably is smart for 85, 90 percent. But what about the 10 or 15 percent remaining that are faulty? — Olivier Denoo
That remaining slice is where the damage hides. AI can hallucinate invented requirements, invented test cases, invented code. A common example is references: ask AI to cite its sources and it may produce a plausible author, a plausible book, a plausible co-author, with the whole thing made up. If you do not cross-check, if you do not know enough to spot the flaw, you are running a real risk. Keeping a human in the loop is what prevents the confident-but-wrong output from reaching production.
Where AI helps testers, and where it stops
AI handles the easy parts of the testing process well. It struggles at the boundaries of the requirement and the oracle, which is where human judgment stays essential.
AI can write and check requirements. It can sort and prioritize. It can generate test cases. The catch is that you still have to be sure the requirements and the test cases are correct in the first place.
The hard problems have not changed in thirty years. Where are the requirements? What is the software actually supposed to do? How precisely and unambiguously can you describe what you want, down to the fine detail? And on the other side, the oracle problem: how do you confirm that the result is right?
A tax declaration shows the limit. There are so many possible cases that checking whether the final amount is correct, the right sum owed or reimbursed, is genuinely difficult. That kind of verification is not an easy task for an AI.
The user is a real person, not a persona you modeled
Software is built for someone, and that someone has to be able to understand and use it. Testing that forgets the actual user misses defects that no test case describes.
The problem with personas is that they cover the people you already imagined. Olivier’s example is his mother, turning 80, who does not use a smartphone or apps. In Paris you now need an app to take the metro. So how does it work for an 80-year-old who simply wants to travel and ride the metro? The process may be well defined and helpful for many people, yet it quietly excludes whole categories of users.
UX belongs in the tester’s scope for this reason. The human in the loop is not only the operator inside your system. It is the person on the receiving end of the output, the one who has to make sense of what the software produces.
Edge cases live where the neat process meets the real world
A perfectly defined process still fails at the single step where a person does something the model did not account for. Olivier’s lost suitcase makes the point.
Airport luggage management is specified end to end. Every step and every case is handled. On a trip back from Thailand, his luggage was stuck in Geneva while he landed in Brussels. Swiss Airline confirmed it would follow on the next morning’s flight, and his own tracker showed it arriving in Brussels. Then Brussels Airlines emailed to say the luggage could not be found, even though he could name the room it was probably sitting in.
His suspicion: one person pulled the bag off the conveyor and forgot to scan the tag. One missed scan, and the item vanished from the system. This is the space testers have to think into. People will use, and sometimes misuse, the software, the car, the plane. Unexpected events happen at exactly the seams the specification treats as solid.
Which skills to build now for a tested future
The skills that keep testers relevant sit at both ends of the technical spectrum: sharper requirements work and stronger communication, with AI fluency assumed.
On the technical side, requirements engineering leads the list, alongside prompt engineering and Gen AI techniques, because AI is now part of the job. DevOps practice, pipelines, and automation are in demand. Manual testing stays, but it moves toward exploratory testing rather than repetitive scripted runs. Underpinning all of it is business understanding: knowing what the application is for.
If you want a structured path, ISTQB offers two AI-focused certifications. One covers using Gen AI for testing, recently updated to track a fast-moving market. The other, CT-AI, covers testing AI itself, including agentic architecture, and has also been updated. A DevOps syllabus is on the way, alongside the established tracks in test management and agile.
Communication is the skill schools skip
Communication is the tester skill Olivier rates highest, and it is the one IT education leaves out.
It starts with prompts. The old problem is the new problem: we still do not know how to describe exactly what we want. Fuzzy, unclear, ambiguous statements break down with an AI just as they broke down with a development team thirty years ago. Writing a precise prompt is a communication skill.
It does not end there. Even when AI generates the report, sorts the test cases, and builds a prioritization framework, you still have to carry results to business stakeholders and management. That means presenting data honestly, without distorting the outcome, and handling the psychology and the power games around it.
Engineers gravitate to technology because they like it. That pull is real, and it is also a blind spot. The soft skills, the ones set aside or forgotten in technical courses, are what let a tester turn correct findings into decisions other people act on.
Frequently Asked Questions
Which testers are most at risk of being replaced by AI?
Testers whose work consists of running pre-written scripts are the ones AI absorbs first. Someone who has pressed the same button for twenty years, executing scripts they did not even write themselves, is directly replaceable. Domain knowledge is the protection: a tester who understands the business behind the application is far harder to automate away.
Does AI-assisted coding reduce the amount of testing needed?
No. Vibe coding produces software faster and controls its quality less. When a non-developer prompts an application into existence, nobody answers whether the code matches the requirements, whether it stays maintainable, or whether it is secure; repositories can carry backdoors and Trojans. More generated code means more defects, and therefore more demand for expert testers.
Why does business knowledge matter as much as testing technique?
Software exists to serve a goal and help someone do a task, not to show off a platform or a language. A tester who does not understand that goal cannot judge what correct behaviour looks like. The difficulty grows as applications become intricately linked: business flows run across several systems, and someone has to follow the whole process end to end.
How trustworthy is AI-generated output in testing work?
Confident, and wrong often enough to matter. AI always returns an answer because it is built to return one, and perhaps 85 to 90 percent of it looks and is sound. The remaining 10 to 15 percent carries the damage: invented requirements, fabricated test cases, or plausible references with made-up authors and books. Someone who can spot the flaw has to cross-check.
Can AI confirm that a test result is correct?
Not reliably. Checking whether a result is right, the oracle problem, is one of the hard parts AI does not solve. A tax declaration shows the limit: the number of possible cases is so large that verifying the final amount owed or reimbursed is genuinely difficult. Generating and sorting test cases is easier for AI than confirming correctness.
Why are personas not enough when testing for real users?
Personas only cover the people you already imagined. Olivier Denoo’s example is his mother, turning 80, who uses neither a smartphone nor apps, while taking the metro in Paris requires one. A process can be well defined and helpful for many people and still quietly exclude entire categories of users. That is why UX belongs in the tester’s scope.
Why do fully specified processes still fail in practice?
Because a single person can step outside the model. Airport luggage handling is specified end to end, yet Olivier Denoo’s suitcase stayed in Geneva while he landed in Brussels, most likely because someone pulled the bag off the conveyor and forgot to scan the tag. One missed scan, and the item vanished from the system.
Is manual testing disappearing?
No, but it changes shape. Manual testing moves away from repetitive scripted runs towards exploratory testing, where human judgement contributes something a script cannot. Alongside it, requirements engineering, prompt engineering and Gen AI techniques are in demand, as are DevOps practice, pipelines and automation. All of it rests on understanding what the application is actually for.
