Every few months, someone declares Selenium dead. Yet roughly 75% of organizations still run it in some form in 2025. What is ending is not a tool. It is a way of working: hand-written, locator-dependent scripts that a team repairs by hand every time a screen changes.
I have spent 15+ years building software in financial services, much of it leading quality engineering for credit and mortgage platforms. In that world, a test suite does more than catch bugs. It is evidence that a regulated decision behaved the way the rules say it should. That shapes how I read the current wave of AI testing tools.
AI is taking over real parts of testing that used to be manual. But the organizations getting value from it are not the ones automating the most. They are the ones deciding, deliberately, what AI is allowed to change without a human looking.
Key Takeaways
- AI is re-scoping Selenium and Playwright, not replacing them: execution stays, while resilience moves to AI layers around the engine.
- Self-healing fixes broken locators, not broken requirements, and in regulated software those are very different failures.
- Adoption is running ahead of operations: about 89% of organizations pilot generative AI in quality engineering, but only about 37% run it in production.
- “AI drafts, humans confirm” is the operating model that closes that gap.
- The quality engineer is moving from script author to quality owner, and leaders need to hire and measure for that shift.
Table of contents
Selenium Isn’t Dying. It’s Being Re-Scoped.
The legacy test stack is a straight line. A requirement becomes a Cucumber feature file, then step definitions, then page objects, then Selenium WebDriver calls, then a pass/fail report. Every layer is deterministic, and every layer breaks the same way. Rename one button and the failure ripples all the way down to a red build.
The modern stack keeps the execution engine. Playwright adoption grew about 235% in 2025, and most enterprise teams now run both: Selenium for legacy regression, Playwright for new work. What changes is everything wrapped around the engine:
- An AI test-selection layer decides which tests each commit actually needs.
- DOM, network, and API observers watch the run in real time.
- A self-healing locator resolver absorbs UI changes before they become failures.
- An ML failure-prediction layer feeds a live risk dashboard instead of a static report.
Execution has also moved off local machines onto managed cloud browser grids, so parallel cross-browser runs are a pipeline setting rather than a hardware purchase. Cloud solved scale. It did not solve resilience. That is the job of the AI layers.
The economics explain the urgency. Test maintenance typically consumes 30 – 40% of a QA budget. Industry data attributes roughly 28% of real-world test failures to broken selectors and about 30% to timing issues, the two classes self-healing targets directly. Reported maintenance reductions of 30 – 40% are common. Vendor claims of 60% – 90% should be treated as directional until your own pilot confirms them.
Self-Healing Can’t Heal a Changed Requirement
In credit risk and mortgage software, the most expensive test failures I have seen were not caused by UI changes. They were caused by rule changes: a new disclosure timing requirement, a revised income calculation, an updated adverse-action reason code. The screen looks the same. The correct answer is different.
This is where self-healing becomes a risk. A locator healer sees a missing element, finds the closest match, and turns the test green. If the element moved because the business rule moved, the healer has hidden the exact change the test existed to catch. A red build costs an hour. A green build that validates the wrong rule can cost a compliance finding.
The distinction leaders need to draw is between structural and semantic failures. Structural failures (selectors, waits, environment flakiness) are safe candidates for automated healing. Semantic failures (a changed calculation, eligibility rule, or disclosure) must stop the line and go to a person. In practice that means four rules:
- Trace every critical test to the requirement or regulation it proves, so a rule change flags every test that depends on it.
- Keep business assertions separate from page interaction. The Page Object Model still earns its place here.
- Log every heal, and review heals on regulated flows before they merge.
- Tier flows by risk: auto-heal on marketing pages, human review on underwriting, pricing, and disclosures.
Lessons From a Platform Transformation
From 2019 to 2021, I led quality engineering for a platform transformation: 50 – 60 SDETs and quality engineers across the US and India. The test framework we built became the standard across more than 100 projects in the business unit. Three lessons from that work still shape how I evaluate AI testing.
First, requirement churn, not UI churn, drove maintenance cost. 50%: share of test maintenance traced to rule or requirement changes versus UI changes. Separating rule assertions from page interaction let us update a calculation once instead of in dozens of scripts.
Second, speed came from selection and parallelism, not from writing tests faster. Regression cycles went from 1 week: before to 1 day: after, automated coverage of critical borrower journeys rose from 35% to 95%, and we were able to ship faster.
Third, false positives destroy trust faster than missed defects. Before the redesign, 70% of failures were flaky or environmental. Engineers had learned to rerun until green, which meant real failures were rerun too. Bringing that to <5% did more for release confidence than any new test we added.
Those lessons predate generative AI, but they map directly onto it.
Selenium AI Drafts, Humans Confirm
The World Quality Report 2024 – 25 finds that about 70% of organizations are piloting or deploying generative AI in quality engineering, but only about 37% run it in production. The gap is not model quality. It is trust.
The organizations closing that gap share one operating model. AI produces candidate tests, healed locators, and test priorities. A human reviewer or validation gate confirms intent before anything ships. In practice:
- Treat AI-authored tests like code: reviewed, versioned, and owned by a named engineer.
- Require an explanation for every heal and every deprioritized test. In regulated industries, that traceability is becoming an expectation, not a best practice.
- Measure defect escape rate and heal acceptance rate, not test count.
This model changes who you hire. Generative AI is now cited by 63% of respondents as a top in-demand skill for quality engineers, and QA engineering roles reportedly grew 17% over two years, ahead of 9% for developers. The function is not shrinking. It is moving from script author to quality owner: people who decide what deserves testing, judge where AI coverage can be trusted, and translate quality into business risk.
Read Next
A few related pieces worth your time:
- Your Enterprise Has a Knowledge Debt Problem. AI Is Sending the Bill
- The Rise of Neurotechnology in Consumer Wellness Devices
- The Hidden Cost of Delaying Legacy System Modernization in 2026
The Selenium Decision in Front of Engineering Leaders
Gartner’s widely cited projection has enterprise use of AI testing tools rising from about 15% in 2023 to about 80% by 2027. Adoption, in other words, is not the decision. It is already happening.
The real decision is where to draw the line between what AI may change on its own and what needs a human. Draw it by risk, not convenience. Teams that get this right will ship faster, spend less on maintenance, and still be able to explain every green build to an auditor. Teams that do not will have very fast pipelines and very little evidence.
