Software testing in 2025 is defined by two forces: AI and accessibility. AI serves as a productivity booster for testers, accelerating the generation of test cases, automated scripts, and test data, while requiring experienced humans to guide and validate results. Accessibility testing has grown due to regulatory pressure but is consistently underestimated in effort, since only part of it can be automated.
Key Takeaways
- AI tools accelerate testing work as a productivity booster, but they require experienced testers who know which problem they are solving, not a replacement for human judgment.
- Agentic AI will shift the tester’s role toward orchestration: the tester acts as the pilot directing AI agents, not as a co-pilot executing tasks handed down by automation.
- Companies routinely underestimate the effort required for comprehensive accessibility testing, because only a portion of checks can be automated and manual testing remains essential.
- The EU AI Act’s Article 4 demands AI literacy from all employees, making competent and risk-aware use of AI tools a baseline requirement across entire organizations, not just for specialists.
- Testing of AI systems is not simply traditional testing applied to a new object: non-deterministic behavior requires different quality assessment approaches, and the higher the risk domain, the more critical rigorous validation becomes.
AI in software testing sits between dystopia and utopia, and realism is winning
The dominant theme across testing conferences in 2025 was AI: testing with AI, testing of AI, and testing AI systems. None of these is going away. They will still shape the conversation in two or three years.
The shift this year was in attitude. The first wave, starting with the arrival of GPT a few years ago, was mostly about chances and possibilities. Then came two extremes. On one side, the prediction that AI brings the biggest disruption of all time and that humans will be replaced. On the other, a new realism from the tester’s seat.
That realism is healthy. AI is a good tool and a real help, but it gets you to roughly 80 or 90 percent, not all the way. You still need to be skeptical. You still need your skills. The Eurostar Conference in June caught the mood in three words: AI on trial.
What is the real benefit of AI for testers?
The benefit is speed and productivity, not the removal of human work. AI acts as a booster for what you already do, letting you reach your goals and run your activities in far less time than before. It also lets you take on tasks that were out of reach previously.
That booster only works with experienced people behind it. You need to know what you are doing and which problem you want to solve, then put powerful tools in your hands. Skills, upskilling, and a basic AI literacy are the precondition, not an afterthought.
There are thousands of imagined use cases: generate this, analyze that. The strategic goal is often missing. Doing something with AI for its own sake is not a use case. Surveys and statistics still show the real productivity gain is sometimes absent, and many companies stay reluctant and watchful for now.
Testing work is shifting from analysis toward generation
The strongest near-term value of AI in testing lies in generating artifacts, not in analytical work. A figure in the last World Quality Report points to a move away from analyzing data and finding patterns, toward producing the things testers need.
Those artifacts are familiar. Test ideas. Test cases with steps. Automated scripts. Test data. When this generation is done well, it becomes one of the bigger advantages AI offers a test team.
Accessibility testing grew because regulation forced it
Accessibility testing moved up the agenda in 2025, and the EU Accessibility Act is the reason. Without that regulatory pressure, accessibility would still be a nice-to-have for many decision makers who have to justify spending money on it. Regulation is the driver here, the same way the AI Act and security rules drive their fields.
Compared with AI, accessibility is a smaller and more contained topic. It is more relevant than in previous years, but you cannot weigh the two on the same scale.
Why companies underestimate accessibility testing
Accessibility testing is not a single tool or a browser add-on that scans your website and declares it done. Only some checks can be automated. The rest needs people running manual tests, and the two approaches complement each other.
That manual share is where the effort hides. Clients this year repeatedly underestimated what a thorough accessibility test costs to deliver. In some cases the test effort can exceed the original design and implementation of the website.
Accessibility belongs in the design, not in the final test
The cleanest way to handle accessibility is to build it in from the start, with good design and the right front-end technologies. Accessibility testing then becomes validation, confirming the product is sound, rather than a rescue operation.
The alternative path is common and expensive. A team never made up its mind about accessibility, starts testing only during UAT, and finds a long list of problems. Catching those issues before production is effective. It is not efficient.
How much rework is needed depends on what came before. If there was never a sustainable architecture, you may need different technologies or front-end frameworks and effectively start from scratch. Technology alone does not fix it. The whole organization needs the understanding that accessibility is built in, the same way security, usability, and performance testing demand built-in quality.
Some companies still have no awareness of accessibility at all, by choice or otherwise. One reason: the penalties are far smaller than under GDPR or the upcoming AI Act, with nothing tied to annual revenue. So the calculation for many becomes “let’s see what happens”, and accessibility gets treated as a nice-to-have. It is not.
The tester’s role is moving toward orchestration
Agentic AI will push the tester’s role further away from manual task execution. Instead of doing every task by hand, you direct one or more agents that carry out parts of the work for you.
The mental model matters here. You are the pilot using a co-pilot, not the co-pilot serving a pilot. Your job becomes orchestration: deciding what needs doing, steering the agents, and judging the result.
I’m the pilot, I use a co-pilot, I’m not the co-pilot and using a pilot. — Florian Fieber
AI literacy is a legal expectation, not just a skill
Article 4 of the EU AI Act requires AI literacy across a company’s staff. That means enough skill, real experience, the right tools, and knowing how to work with them, not only for specialists but for everyone.
For 2026, this turns into a concrete task for every company. AI is a commodity in daily work. The question is no longer whether you use it, but whether you can use it competently, knowing its risks, its limits, and its boundaries.
Testing of AI is still a niche, but a high-stakes one
Split the field into testing with AI and testing of AI. The far larger impact today comes from AI as testware, supporting testers in their work. Testing of AI systems remains comparatively small.
For anyone shipping AI components in their product, it is no niche at all. The test object is a non-deterministic system that behaves differently from traditional software. Your existing methods, processes, and tools still apply, but you have to work out how to assess quality and validate behavior that will not repeat identically.
Risk rises with usage. The AI Act sets boundaries for systems where the stakes are high: chat assistants in health insurance, self-driving cars, and similar. In those domains, testing and quality assurance carry real weight.
How to plan your testing skills for 2026
There is no single channel for upskilling. Conferences, YouTube videos, on-site trainings, webinars, and meetups each suit different people, and a mix works best. Conferences give you exchange with people from the industry and inspiration you will not get from a screen alone.
On the certification side, the ISTQB Certified Tester portfolio offers two relevant paths. One syllabus has covered testing of AI systems for about three years. A newer one, available for roughly three months, covers testing with generative AI. Both feed directly into the AI literacy the AI Act now expects.
There is a quiet method hidden in this variety. When you listen to five or ten speakers on the same topic, you start to see the pattern they share. That repetition is the common ground of a field, and spotting it is how you learn to judge what actually matters.
Use this rough split when you build a learning plan:
| Learning channel | What it gives you |
|---|---|
| Conferences | Exchange with practitioners, thought leaders, fresh inspiration |
| Trainings and certification (ISTQB testing of AI, testing with Gen AI) | Structured foundation, AI literacy for your discipline |
| Videos, webinars, meetups | Flexible top-ups, repeated exposure that reveals the common ground |
