Will AI replace testers? What QA work is left

AI will not replace testers, but it is taking over much of the test drafting and repetitive scripting in their week. The work growing in its place is harder to hand off, such as deciding what correct means, testing systems that answer differently every time and owning the risk of a release.

I have worked in testing for 18 years, starting as a working student at a German IT company in 2008. Since then I have been a QA team lead, a freelance tester on uTest and a test manager for Audi through Applause. I have also worked as a business consultant on Volkswagen's online car configurator, and as a business analyst and senior test analyst in Australian life insurance. I also build AI systems at OIC Architects, and I use AI in my contract testing work. This is my honest view, for testers who are worried and for the people who hire them.

What is AI already doing in testing?

It drafts. Give a model a user story with acceptance criteria and it returns positive, negative and edge-case scenarios in moments. It writes Playwright scripts, suggests locators, generates test data, summarises logs and turns a rough note and a screenshot into a tidy defect report.

I see the change in my own work. On one contract I built a natural-language SQL engine with GitHub Copilot to search for test data across five or six tables. Writing the query stopped being the slow part. The bottleneck moved to defining the data question clearly, which is a requirements skill, and one good testers already have.

At uTest I reported more than 900 defects across over 250 test cycles, at a 90% approval rate. A model can now draft a clean defect report from a screenshot and a log. Deciding which of those defects the client will care about is still the tester's call.

What do the numbers say about AI in QA?

Adoption is wide and shallow. The World Quality Report 2025 surveyed more than 2,000 senior executives in 22 countries. It found that 89% of organisations were piloting or deploying generative AI in quality engineering (37% in production, 52% in pilot), but only 15% had scaled it across the enterprise. The average productivity gain was 19%, and a third saw minimal gains. 60% named hallucination and reliability concerns as a barrier, and half said they lacked AI and machine learning expertise.

Tal Levi-Joseph of OpenText put it well in the report's release. "AI amplifies capability, but it cannot substitute for it."

Jobs and Skills Australia reached a similar view for the whole workforce. Its August 2025 study found that generative AI is more likely to enhance jobs than replace them, and that adoption is still early and varies across industries.

Read together, the picture is plain enough. Most teams are piloting, few have scaled, and the top worry is whether the output can be trusted. Finding out whether output can be trusted is testing work.

Which QA work shrinks, and which grows?

Shrinking	Growing
Writing test cases line by line from a story	Deciding what a correct result is before anything is built
Hand-coding boilerplate scripts and locators	Designing tests for systems that answer differently on every run
Executing scripted manual steps someone else wrote	Building evaluation sets and pass rates for AI output
Formatting defect reports and status updates	Testing for prompt injection, made-up facts and wrong tool calls
Searching for test data by hand	Judging release risk and signing off with evidence

Good testers will recognise the right-hand column. Those tasks were always part of the job, and they used to get squeezed by the routine work on the left.

The National AI Centre's Guidance for AI Adoption shows where the demand comes from. "Test and monitor" is one of its six practices, and it expects acceptance criteria, testing before deployment and a documented sign-off from the accountable person based on test results. Somebody has to produce that evidence, and it is the same evidence a test manager has always produced for a release.

Why does AI make a tester's judgement worth more?

Because AI-built checks fail quietly. I built this website with AI assistants, and several checks kept passing while testing the wrong thing. One overflow check measured an element whose width could never fail. A fix for duplicated hero labels, 30 on screen instead of 15, turned the duplication into a visible flash on load. None of it showed up as a red test.

The cure was to break each check on purpose and confirm it failed before trusting it. That is plain testing discipline. A model will write a check in seconds. Deciding whether the check can fail at all is the tester's job.

The same suspicion applies to test content. On an architect exam preparation app I built, one existing question bank had 89.7% of its correct answers in the same position, so a candidate could score well without knowing the material. You do not find that by reading questions one at a time. You find it by asking how a score could be wrong, and then counting.

What should a tester learn next?

None of this needs a paid course. I would take it in roughly this order:

Write acceptance criteria for AI output. Say what a good answer contains, which failures are unacceptable and when a person must approve.
Build a small evaluation set from real requests and run each case several times. Report a pass rate instead of a single result.
Learn enough Python or JavaScript to write deterministic checks, such as counting fields, comparing numbers and confirming which tool an agent called.
Learn when a model can grade output. Anthropic's guide to evaluations rates code-based grading as the fastest and most reliable, and says model-based grading should be tested for reliability before you scale it.
Read the OWASP Top 10 for LLM applications, starting with prompt injection, and practise planting hostile instructions in test data.
Let an agent drive a browser under supervision, for example with Playwright MCP in VS Code, so you learn where it helps and where it guesses.
Get better at the requirements conversation. A report from my content research system scored 80 out of 100 against its own brief, and every deduction became a permanent rule or validator. That only works when the brief is clear enough to score against.

A good first exercise for this week is to take an AI feature your team already ships, write down what a wrong answer looks like, and count how often it happens. Recruiters and hiring managers can read my CV.

Questions people ask

Will AI replace QA testers?

Not as a group. AI is taking over the routine parts of QA, such as drafting test cases, writing boilerplate scripts and formatting reports. The need for people who can judge whether software, including AI software, is fit to release is growing. In the World Quality Report 2025, 89% of organisations were piloting or deploying generative AI in quality engineering, only 15% had scaled it and half lacked AI expertise. Testers who learn evaluation and risk are the ones who close that gap.

Can AI replace manual testers?

It can replace manual execution of fixed scripts, and that part of the job will keep shrinking. It is much weaker at exploratory testing, where a tester notices something odd and follows it, and at deciding what matters to the business. Manual testers who shift towards exploratory testing, test design and checking AI output are in a far better position than those who only execute steps someone else wrote.

Will AI replace automation testers?

Automation testers lose the slowest part of their work first, which is writing and repairing scripts by hand. Agents can now generate Playwright code and suggest locators. The job becomes reviewing that code, deciding what deserves automating, keeping the suite trustworthy and testing the AI features in the product. Self-healing tools still guess at times, so someone has to check what they changed and why.

Is QA automation still a good career?

Yes, if you widen it. Script writing alone is the part AI does best, so a career built only on that is exposed. Automation skills combined with evaluation, code-based checks, prompt-injection testing and requirements work are scarce, because organisations are piloting AI faster than they can prove it works. Australia's AI adoption guidance lists testing and monitoring as one of six practices, which tells you where the work is heading.

How do I upgrade my testing career with AI?

Start on your current project. Use an AI assistant to draft test cases and scripts, and review everything it produces as if a new starter wrote it. Then build one small evaluation set for an AI feature, write code checks for the countable parts and plant a prompt injection in the test data. Keep notes on what failed and why. That record makes a strong interview story, because it shows you can find problems in AI output.

Sources

Share on LinkedIn

Will AI replace testers? What QA work is left, from 18 years in testing