How AI Has Transformed QA Automation at NEOZO
At NEOZO, AI has cut the time it takes to build end-to-end test automation from weeks to hours. The speed comes from AI agents that explore applications on their own. The quality still comes from our QA engineers, who review every test before it runs.
Before AI: every test written by hand
Until recently, every end-to-end test was handwritten. QA engineers had to locate each element on a page manually, choose a stable locator for it and write it into the automation script.
For user flows with many interactions and complex logic, this meant weeks of work per suite. It also meant constant maintenance: whenever the UI changed, locators broke and had to be tracked down again. Test automation often lagged behind development, and releases waited for it.

How it works: Playwright MCP and AI agents
Today, our QA engineers no longer hunt for locators by hand. We pair the Playwright MCP server with AI coding agents such as Claude Code. The Model Context Protocol (MCP) gives the agent controlled access to a real browser through Playwright.
The agent opens the running web application, navigates through user journeys and discovers interactive elements as it goes. It reads the page structure through the accessibility tree, which tends to produce more stable, role-based locators than brittle CSS or XPath selectors.
A practical advantage: the agent needs no prior knowledge of the codebase and little setup. From its exploration, it can generate a working draft of the automation script, including Page Object Model (POM) classes and locators. Work that used to take weeks can now produce a solid first draft within a day.
Our workflow: AI speed, engineering rigour
Our process combines the speed of AI with the judgement of experienced QA engineers in three steps.
1. Automated discovery. Playwright MCP explores the live interface and maps its elements. On request, the agent writes initial test scripts and POM definitions as a draft.
2. Human review and refinement. Our QA team reviews and refines every generated test. The AI handles the repetitive clicking and boilerplate; our engineers make sure the assertions validate real business logic, not just the happy path.
3. Spec-based bug hunting. Beyond generating tests, we give the AI product requirement documents. It compares the actual system behaviour with the requirements and points out gaps and missing edge cases, which our team then verifies.
Why humans stay in the loop
We treat AI-generated tests as drafts, never as finished results. An agent can record what an application does, but it cannot judge on its own what the application should do. A generated assertion may simply confirm a bug as expected behaviour.
That is why a senior QA engineer reviews every draft before it enters the test suite. The agents also run against dedicated test environments with test data, not against production systems. This keeps the speed gain without compromising quality or data protection.
Data confidentiality: keeping sensitive information in scope
Giving an agent access to a live application comes with a real risk to data confidentiality. The agent processes everything it encounters on its way through the application: page content, form data, requirement documents. Without safeguards, sensitive information can end up with third-party model providers.
Which safeguards fit depends on the application and the client’s requirements. We work with three:
- Choose where the model runs. A locally hosted model keeps inference within company-controlled infrastructure, provided the agent and its supporting tools are also configured to avoid external data transfers. For cloud models, we review the provider’s data retention, training and access policies before any application content is shared.
- Limit where the agent can go. Dedicated test accounts, restricted permissions and browser origin allowlists keep exploration within scope. Allowlists are guardrails, not a security boundary; stronger restrictions require network controls covering the browser, the agent and its supporting tools.
- Minimise the data exposed. We use mocked or sanitised test data, remove unnecessary confidential details from requirement documents and keep credentials out of prompts and generated scripts. Screenshots, logs and test reports get appropriate access controls and retention periods.
Together, these measures give the agent enough access to build useful tests while limiting the information and systems it can reach.
The result: more coverage, faster releases
By integrating AI into our QA lifecycle, we have reduced test creation from weeks to hours. What used to fill a sprint now fits into a review session.
A QA engineer can cover more of a release in the same week, and deployments no longer wait for a test suite that is still being written by hand. The time saved goes where human expertise matters most: exploratory testing, edge-case analysis and close collaboration with development teams.
The outcome is higher-quality software, delivered faster, with experienced engineers still accountable for every test that runs.
Let’s bring AI into your QA
Want the same speed in your own testing? We can help in two ways:
- We take it on for you. Our QA team builds and maintains AI-assisted end-to-end test automation for your applications, from discovery to a reviewed, production-ready test suite.
- We help you set it up. We advise your team on introducing Playwright MCP and AI agents: tool selection, workflow design, review processes and the guardrails that keep quality and data protection in place.
Get in touch and let’s talk about where AI can speed up your quality assurance.