Skip to content
AI Workshop

AI in Software Testing 2026: 89% Pilot It, 15% Scale It — Claude, Codex, Playwright Agents and What Actually Works

AI now writes a large share of new code, and somebody has to check it. That makes software testing the place where the AI productivity promise either holds or falls apart. The headline numbers look like a done deal: the World Quality Report 2025-26 by Capgemini, Sogeti and OpenText found that 89% of organisations pilot or deploy generative AI in quality engineering. But only 15% have scaled it across the enterprise, the average productivity gain is 19%, and in Applause's 2026 survey 29% of teams reported more or more severe defects even as AI use grew. This guide explains what AI testing really does in 2026, how agents built on Claude, OpenAI Codex, Mistral and Manus fit in, and how companies in Croatia, Germany and Austria should approach it.

89%
of organisations pilot or deploy GenAI in quality engineering (World Quality Report 2025-26)
15%
have reached enterprise-wide scale
19%
average productivity gain reported; a third saw minimal gains

Everybody uses it, depending on who asks

Surveys on AI in testing produce very different numbers, because they ask different people different questions. BrowserStack's February 2026 survey of over 250 technology leaders reports that 94% of teams use AI in testing. Applause reports 92% using AI in functional testing, up from 60% a year earlier. Katalon finds 76% of more than 1,500 QA professionals using AI-powered tools, and mabl reports 57% of QA teams, against 72% of development teams. At the other end, Leapwork's study of 302 testing professionals found that only 12.6% use AI across key test workflows, even though 88% call it a priority.

Reported AI Use in Software Testing, by Survey (% of respondents)

Sources: BrowserStack (Feb 2026, 250+ tech leaders), Applause State of Digital Quality 2026 (functional testing), World Quality Report 2025-26 (Capgemini, Sogeti, OpenText; 37% in production, 52% in pilots), Katalon (1,500+ QA professionals), mabl 2026 State of Quality Engineering, Leapwork AI in Testing study (302 respondents, 2026). Mostly vendor research with different definitions; read as directional.

The honest reading: almost every team has tried an AI feature, often the one switched on inside a tool they already use. Very few have changed how testing works end to end.

The maturity gap

Every survey that asks about depth tells the same story. In the World Quality Report, 89% pilot or deploy, but 15% run at enterprise scale. BrowserStack's 94% shrinks to 12% with full autonomy. Alt.QA found that 73% of engineering teams have some AI testing infrastructure, but only 18% trust their test coverage. Respondents name test case design and requirements refinement as the leading use cases, and PractiTest data puts test case creation (69.6%) and script maintenance (59.6%) at the top.

Using AI in Testing vs Running It at Scale (% of respondents)

Sources: World Quality Report 2025-26 (89% piloting or deploying vs 15% enterprise scale), BrowserStack 2026 (94% use vs 12% full autonomy), Alt.QA Feb 2026 survey of 500 teams (73% have AI testing infrastructure vs 18% trust their coverage), Leapwork 2026 (88% priority vs 12.6% use across key workflows).

What AI testing actually means in 2026

"AI testing" covers five quite different things. Mixing them up is the most common reason pilots disappoint.

ApproachWhat it doesTypical toolsMaturity
Test generationDrafts unit, API and end-to-end tests from code, requirements or user storiesClaude, GPT models, GitHub Copilot, Mistral's DevstralProduction-ready, needs review
Self-healing UI testsRepairs locators when the interface changes, so suites break less oftenCommercial platforms such as mabl, Tricentis, Leapwork, KatalonMature for web UIs
Agentic testing in the IDEA coding agent writes a test, runs it in a real browser, reads the failure and fixes the test or the codeClaude Code or OpenAI Codex with Playwright via MCPFast-moving, strong for developers
Autonomous explorationA general agent clicks through an app like a user and reports what looks brokenManus, browser agents from OpenAI and othersUseful for smoke tests, not deterministic
Testing AI itselfEvaluates chatbots and agents for accuracy, safety and regressionsEval suites, LLM-as-a-judge, red teamingEssential for any AI feature

Classification by AI Workshop based on vendor documentation and the World Quality Report 2025-26. Tool lists are examples, not recommendations.

Where Anthropic, OpenAI, Mistral and Manus fit

The biggest shift of the past year is that testing moved into the coding agent. Claude Code runs in the terminal, writes tests, executes the test suite and iterates until it passes. Connected to Playwright through the Model Context Protocol, it can open the application in a real browser, click through a flow and turn what it saw into a repeatable test. OpenAI Codex follows the same pattern in cloud sandboxes and runs the project's tests before proposing a change. For teams that cannot send source code to a US cloud, Mistral's Devstral models are open-weight and can run on your own servers, which matters for banks and the public sector. Manus and similar general agents can explore a web app the way a user would, which is good for finding obvious breakage, but their runs are not repeatable enough to replace a regression suite.

The practical pattern that works: let AI write and maintain the tests, but run them as ordinary deterministic tests in your CI pipeline. An AI model should not be the judge at release time. We compare the coding agents themselves in our guide to AI coding agents.

The trap to avoid: when the same AI writes the code and the tests, the tests tend to confirm what the code does, including its bugs. Coverage goes up, confidence should not. Keep at least the acceptance criteria human-written, review generated assertions and track escaped defects, not test counts. Our article on the security and quality of AI-generated code shows why this matters.

What holds companies back

The barriers are less about the technology and more about data, integration and trust. In the World Quality Report, data privacy risks (67%), integration complexity (64%) and skill gaps (50%) top the list. In Leapwork's study, 54% name concerns about accuracy and quality. Meanwhile Applause's data shows the risk of moving too fast: 29% of teams reported more or more severe functional defects while AI use rose.

Barriers to Scaling AI in Testing, and the Defect Warning Sign (% of respondents)

Sources: World Quality Report 2025-26 (data privacy, integration, skills), Leapwork 2026 (accuracy and quality concerns), Applause State of Digital Quality 2026 (teams reporting an increase in number or severity of functional defects). Different surveys, not directly comparable.

A market that is still small, by most measures

AI testing tools are a fraction of total testing spend. Estimates for 2026 range from about $0.75 billion for AI testing tools (The Business Research Company) and $1.21 billion for AI-enabled testing (Fortune Business Insights, forecasting $4.64 billion by 2034) to $11.99 billion when Mordor Intelligence counts AI-powered testing and QA services broadly. For comparison, the overall automation testing market is put at about $40.6 billion in 2026. Most of the money still goes to people and conventional tooling.

AI Testing Market Size: Three Estimates for 2026 (USD bn)

Sources: The Business Research Company (AI-enabled testing tools, $0.75B in 2026, 29.1% CAGR), Fortune Business Insights (AI-enabled testing, $1.21B in 2026 to $4.64B in 2034, 18.3% CAGR), Mordor Intelligence (AI-powered software testing and QA, $11.99B in 2026). Scope definitions differ widely.

The EU angle: test data, DORA and the AI Act

  • GDPR and test data. Copying production data into prompts or test environments is processing personal data. Use masked or synthetic test data, choose EU-hosted or self-hosted models for sensitive code and data, and sign a data processing agreement with every AI vendor.
  • DORA for financial services. Since January 2025 banks, insurers and their ICT providers in the EU must run a digital operational resilience testing programme. AI-generated tests are welcome, but you must be able to show what was tested, by whom and with which result. Keep test evidence traceable.
  • EU AI Act. High-risk AI systems must be tested against defined metrics as part of their risk management (Article 9) and documented for accuracy and robustness (Article 15), with obligations applying from August 2026. If you build AI features, testing them is a legal requirement, not a nice-to-have. See our EU AI Act compliance guide and our guide to evaluating AI agents.

Our 30-day pilot plan: pick one product with a stable CI pipeline. Week 1: measure today's baseline (time to write a test, flaky test rate, escaped defects). Week 2: use Claude Code or Codex with Playwright to generate tests for the five most important user flows and review every assertion. Week 3: add self-healing or agent-based maintenance to the UI suite. Week 4: compare against the baseline and decide. Use synthetic data throughout and keep the release gate deterministic.

What this means for companies in Croatia and DACH

  • Your developers already use AI to write tests. Make it official, give them approved tools and set review rules, rather than letting each team improvise.
  • Start where testing hurts most. Brittle UI tests and missing coverage for core flows show results within weeks.
  • Measure outcomes, not activity. Escaped defects, flaky test rate and time to release matter. Number of generated tests does not.
  • Plan for regulated evidence. Banks under DORA and anyone building high-risk AI under the AI Act need traceable test records from the start.
  • AI-built apps need more testing, not less. If business teams build tools with vibe coding and AI app builders, add automated tests before those tools touch customer data.

The bottom line

AI has made writing tests cheap. It has not made deciding what "correct" means any easier, and that is still the job of people who understand the business. The companies that gain the reported 19% and more are those that combine AI-generated tests and agent-driven exploration with deterministic pipelines, human-owned acceptance criteria and honest metrics. The ones that let the AI grade its own homework end up among the 29% with more defects.

Want AI to speed up testing without lowering quality?

AI Workshop sets up AI-assisted testing with Claude Code, Playwright and your CI pipeline, builds evaluation suites for your AI features and prepares the evidence DORA and the EU AI Act require. Our Claude Certified Architects keep code and test data in the EU.

Talk to an AI consultant

Boris Agatić

Founder of AI Workshop and Claude Certified Architect. Over 10 years of building software, guiding companies from the first strategy call to production AI agents.

LinkedIn (opens in a new tab)