AI in Quality Assurance: Automate the Checking, Keep the Judgment

ai in quality assurance

AI is changing quality assurance fast, but "autonomous testing" frames it wrong. What AI does well in QA today, where human judgment still decides, and how to use both.

Key takeaways
The 30-second version of this article
AI is very good at accelerating checking (generating, running and maintaining verifications of expected behavior) and much weaker at testing in the full sense: deciding what matters, noticing what nobody specified and judging whether a result is right.
The oracle problem doesn't go away. An AI-generated test with a wrong expected result passes confidently and protects the bug, so review AI-written tests and verify the assertions.
People stay accountable for which risks matter, whether expected results are correct, which checks are worth keeping, what counts as a real defect and what ships. "The agent said so" isn't an answer a customer or auditor will accept.
Separate two jobs: testing with AI (making testing faster or broader) and testing AI (hallucination, inconsistent answers, bias, prompt injection and silent regressions when a model version changes).
Adopt AI in QA in steps: start where checks are cheap to verify, measure outcomes not output, track rework, keep a human gate on critical journeys and protect time for exploration.

The usual story about AI in quality assurance is a ladder: manual testing at the bottom, automation in the middle, and autonomous testing at the top, where machines take over. I think the ladder measures the wrong thing. The useful question isn't how much testing can run without people. It's who is accountable for each decision, and what evidence they need to make it well.

AI is very good at accelerating checking: generating, running and maintaining specific verifications of expected behavior. It's much weaker at testing in the full sense: deciding what matters, noticing what nobody specified and judging whether a result is actually right. Teams that get value from AI in QA put it to work on the first job and keep people firmly in the second. For background on the checking side, see what is automation testing.

What AI Does Well in QA Today

The practical gains are real, and most fall into a few groups:

  • Drafting test ideas and cases from requirements. A model can turn a user story into a first set of scenarios in seconds, including edge cases a tired human might skip.
  • Generating test data. Synthetic records that match a schema and cover boundaries, without copying production data.
  • Maintaining automation. Updating locators and steps when the UI changes, which removes some of the grind that makes UI suites expensive.
  • Triage and summarization. Grouping failures, suggesting likely root causes and drafting defect reports.
  • Learning from usage. Generating regression tests from how real users move through an application, so coverage follows actual traffic rather than guesses.

Each of these makes checking cheaper and faster. None of them removes the need to decide whether the checks are the right ones. When you compare automation testing tools, check how each one lets people review what the AI changes.

The Oracle Problem Doesn't Go Away

An oracle is how you recognize a problem: the source of truth that tells you whether a result is right. ISO/IEC TR 29119-11, the international technical report on testing AI-based systems, names the test oracle problem as a central challenge, because expected results for AI behavior are often hard to determine.

The same problem shows up when AI writes the tests. An AI-generated test case with a wrong expected result doesn't fail. It passes, confidently, and protects the bug. Plausible output isn't correct output. That's why I review AI-generated tests the way I'd review a capable new colleague's work: assume good effort, and verify the assertions.

Who Does What

Task

AI can take on

People stay accountable for

Test analysis

Drafting scenarios and edge cases from requirements

Deciding which risks matter to the business

Test design

Generating cases, data and boundary values

Confirming expected results are actually correct

Automation

Writing and healing scripts

Architecture, and which checks are worth keeping

Execution

Running suites at scale, flagging anomalies

Exploring what nobody thought to check

Triage

Clustering failures, suggesting causes

Deciding what's a real defect and what ships

Reporting

Summarizing results and trends

Telling stakeholders what the residual risk means

The right column isn't a temporary list waiting for better models. It's where accountability lives. Someone has to own a release decision, and "the agent said so" isn't an answer a customer or an auditor will accept.

A written test strategy is where a team records those decisions: which risks get covered, who reviews what, and what counts as done.

Two Different Jobs: Testing With AI and Testing AI

It helps to separate two jobs that often get blurred.

Testing with AI means using AI to make your testing faster or broader. Everything above applies.

Testing AI means evaluating features that are themselves built on models: chat assistants, recommendation engines, agents. Here the failure modes are new: hallucination, inconsistent answers to the same input, bias, prompt injection and silent regressions when a model version changes. Pass/fail assertions aren't enough; you need evaluation across many samples and human review of borderline output.

The standards community has caught up quickly, which gives teams a shared vocabulary. Alongside ISO/IEC TR 29119-11, ISTQB released its Certified Tester Testing with Generative AI (CT-GenAI) syllabus in July 2025. It covers prompt engineering, hallucinations, bias and data privacy risk, and the use of generative AI across test analysis, design, automation and reporting.

How to Adopt AI in QA Without Losing Trust

  1. Start where checks are cheap to verify. Test data and draft scenarios are low-risk places to learn how a model behaves.
  2. Measure outcomes, not output. Count risks covered and defects found, not tests generated. A thousand AI-written tests that never fail may be a thousand tests that check nothing.
  3. Track rework. If people spend as long fixing AI output as they would have spent writing it, the gain is an illusion.
  4. Keep a human gate on critical journeys. Payments, security and regulated workflows get human review of what is being checked.
  5. Protect time for exploration. The time AI saves on checking should go back into testing, not just more checking.

These measures are easier to track when generated tests, reviews, and results sit in one place, which is the job of test management.

Where Katalon Fits

Katalon True builds AI into each part of the test workflow: six agents, coordinated through the Katalon AI Assistant. Katalon's documentation describes the assistant as working "while keeping humans in control," and when it runs manual tests autonomously, it asks for human input where it needs guidance. Here is how the six agents line up with the table above, using the descriptions in that documentation:

Agent

What it does

Where it sits

Requirement Analyzer

Analyzes requirement completeness, detects ambiguity and supports test planning

Test analysis

Test Generation Agent

Generates the test cases needed to cover requirements and prioritized flows

Test design

Autonomous Test Runner

Executes natural-language or manual workflows and captures results and evidence

Execution

Bug Reporter

Converts failures into structured bug reports for issue tracking systems

Triage

Root Cause Analyzer

Investigates likely failure causes and recommends remediation steps

Triage

Report and Insight Generator

Summarizes testing outcomes and quality signals for stakeholders

Reporting

Every row of that table sits in the left column of the earlier one. The right column, deciding which risks matter and whether an expected result is correct, doesn't move. The demos below show what that looks like in the product.

See the Whole Flow, From Requirement to Report

This walkthrough takes one Jira requirement through Katalon True. A review agent scores the requirement, manual test cases are generated from your own prompt and files, the AI test runner executes them with screenshots and video, and Katalon Studio automates them. The flow then repeats autonomously from Claude Code through the Atlassian, Katalon True and Studio MCP servers, ending in an HTML report. Watch for the points where human review stays in the loop.

Nothing the AI Generates Is Saved Without Your Review

If an agent generates a test case, you should decide whether it gets kept. In Katalon, generated manual test cases arrive for you to keep or edit, and nothing the AI generates is saved automatically. Script changes from the Katalon Studio AI assistant appear as a side-by-side diff that you can mark as reviewed or undo. That's the review habit this post recommends, built into the workflow. The same video also shows a user journey map built from production telemetry that turns uncovered flows into test cases.

Check the AI's Failure Analysis Before You Trust It

The oracle problem applies to AI explanations of failures, too. Katalon adds AI analysis to every report automatically: failure category, root cause, impact and a recommended fix, with a confidence meter on the answer. A report review agent then condenses a run into key results, risks and red flags, and where to focus next, and it lists its sources so you can check each point. A confidence meter isn't a verdict. It's a prompt to check the sources.

Testing AI: Quality Gates for an AI Assistant

For the second job, testing AI itself, this demo runs a Katalon test that asks Katalon's own AI support assistant a question and scores the response against parameterized quality gates: relevance, accuracy, completeness, clarity, hallucination and context retention. The check runs inside a normal end-to-end flow, produces a machine-readable report and fails the test when the gates aren't met. Scores like these are a starting point. Borderline answers still need human review.

Where the Tools Stop

Katalon TrueTest generates regression tests from real user behavior. Because it follows traffic, rarely used but high-risk flows need to be selected deliberately. None of the agents decides which risks matter to your business or whether an expected result is right. That's a good example of the larger point: AI can scale your checking a long way, but deciding what deserves checking is still your call.

Frequently asked questions

Mush Honda

Mush is Chief Quality Architect at Katalon, with over 20 years of experience in Quality Engineering and Agile software delivery. In-depth experience of development / delivery methodologies with expertise in all areas of software quality engineering, test automation, DevOps, and test project management.

Connect on LinkedIn →
Get started

Test everything. Ship with confidence.

From generating test cases to spotting failures, Katalon's AI works across your entire pipeline, automatically.

The Next Evolution of Manual Testing: Introducing Katalon AI Test Runner

Introducing Katalon AI Test Runner. Watch it run manual test cases live, with evidence for every pass and fail.

Live webinar
9:00 to 10:00 AM EST · Live online
Alex MartinsVP of Strategy
Gia EspositoSenior Solutions Engineer
Save your seat