State of AI in Testing Survey 2026: What We Are Asking

What the 2023 Survey Found

Every number in the table below comes from the Future of Quality Assurance Survey, fielded by TestMu AI between July 21 and October 31, 2023 across 1,615 respondents in 70 countries. It is the baseline the 2026 survey is designed to be compared against.

Read the top three use cases together and a pattern falls out: in 2023, AI in testing was overwhelmingly an authoring and analysis assistant. Test data, test cases, and log summaries are all tasks where a human states the goal, a model drafts an answer, and the human keeps the pen. Nothing in the top three involves AI deciding what to run or fixing something without being asked. If your team is weighing where AI fits in your own QA process, the practical use cases are broken down in our guide to AI in quality assurance.

What Changed Between 2023 and 2026

The 2023 instrument measured assistive AI because that was the only kind in production. Three years on, four capabilities exist in real pipelines that the old questions cannot describe.

• Agentic authoring - Tests are generated from the requirement itself. A Jira ticket, a PRD, a screen recording, or a GitHub pull request diff becomes a scenario with proposed assertions, rather than a human re-interpreting the spec into steps.
• Self-healing in production suites - When a UI change would break a step, the step is re-anchored automatically and the change is surfaced for review. Maintenance moves from rewriting a selector to approving or rejecting a heal.
• Merge-gate validation - End-to-end suites are generated and run against a pull request, with results posted back into the review thread before a human finishes reading the diff.
• AI systems as the thing under test - Chatbots, voice agents, and autonomous browser agents are now shipped products that need their own quality strategy, which is a category the 2023 survey never asked about.

TestMu AI builds in this category, which is part of why the survey exists. KaneAI is a GenAI-native end-to-end testing agent: tests are authored in natural language or generated from PRDs, tickets, PDFs, recordings, and GitHub PRs; smart element detection and self-healing re-anchor steps as the application changes; and the resulting tests export to Selenium, Playwright, Cypress, and Appium so teams keep their existing codebase. Runs execute in parallel on HyperExecute across 3,000+ browser and OS combinations and 10,000+ real devices, and every plan is reviewable before it executes, with user input taking priority over agent decisions.

Building the tooling is not the same as knowing how the industry uses it. Vendors see their own customers; a survey sees the teams who chose something else, or chose nothing. For a wider view of where these capabilities currently stand, see our overview of AI in software testing.

Four Questions the 2023 Data Left Open

These are the specific gaps the 2026 survey is built to close. Each one is a number that either does not exist yet or is three years stale.

• Does reliability still lead the blockers? In 2023, 60.3% of organizations named reliability concerns as their primary obstacle, at a point when AI mostly drafted text a human then reviewed. Self-healing and agentic execution raise the cost of a wrong call considerably, so the same concern may now rank higher, or may have been displaced by governance and cost.
• Did the skills gap close or move? 54.4% of organizations reported a lack of skilled professionals in 2023. The scarce skill then was writing automation. The scarce skill now may be reviewing what an agent produced and knowing when to reject it, which is a different hiring problem with a different training answer.
• Who is deliberately staying out? Asked whether they currently use AI tools in their development process, 16.2% of the 33,662 respondents to the 2025 Stack Overflow Developer Survey said they do not and do not plan to. No equivalent figure exists for quality engineering specifically, and the reasons behind a QA team's refusal are likely to differ from a developer's.
• What is nobody willing to hand over? The 2023 survey measured what teams automate with AI. It never asked the inverse: which checks a team insists a human signs off on, and why. That boundary is where quality strategy actually gets decided, and it has never been measured at scale.

The fourth question matters most for the categories now emerging. We took a first pass at one of them in our field review of AI browser agents in 2026, which found the gap between demo and dependable to be considerably wider than vendor claims suggest. Survey data is how that observation gets tested against a few thousand other teams instead of one.

How to Take Part

The survey is open to anyone working on software quality, with no TestMu AI account required.

• Open the survey - The survey link sits in the highlighted callout above, directly under the 2023 findings table. There is no login step and no email gate before the first question.
• Answer as your team actually works - The value of the report depends on honest answers about current practice, not aspirational ones. A team that tried an AI tool and abandoned it is a more useful data point than a team that reports the tool it intends to buy.
• Finish in one sitting - Budget about 10 minutes. Nothing in the survey requires you to pull metrics from a dashboard first.
• Pass it to your team - One response describes one person's practice. Three or four from the same team describe an organization, which is the unit most of the analysis is cut by.
• Send questions to the conferences team - Email conferences@testmuai.com for clarifications about the survey itself or about how the results will be used.

Why Your Response Counts

The Future of Quality Assurance Survey report is cited because its sample was large enough to cut the data by company size, region, and role and still say something defensible about each slice. Sample size is what separates a report from an opinion.

The thin slices are the ones that need volume most, and they are predictable:

• Manual and exploratory testers - Under-represented in every AI adoption study, despite being the role whose day-to-day work agentic authoring changes most directly.
• Regulated industries - Finance, healthcare, and public sector teams operate under audit constraints that make the automate-or-not question genuinely different, and there are rarely enough responses to report those cuts separately.
• Mobile and device-heavy QA - Adoption patterns on real-device suites diverge from web automation, and blended figures hide that difference entirely.
• Small teams - Organizations with one or two people covering all of QA have a different constraint set than a 40-person quality org, and their answers usually get averaged away.

If you sit in any of those groups, your single response moves the published numbers more than a response from an over-sampled segment does.

Where the Results Go

The survey is open now and keeps collecting responses past TestMu Conference 2026, which runs August 19 to 21, 2026. The conference is free, virtual, and built around the same question the survey asks, with 80+ sessions and 100+ speakers on agentic engineering and quality.

TestMu AI publishes the final report in September 2026. It will be free and public, the way the 2023 report is, so the numbers are citable by anyone building a case internally rather than locked behind a form. Sessions across the three conference days will reference the emerging results, and our rundown of reasons to attend the Testμ Conference covers what else is on the agenda.

Responses are reported only in aggregate. No individual answer, name, or employer appears in the published report.

Add Your Data to the 2026 Report

Open the survey, block out 10 minutes, and answer for how your team works this quarter rather than how you expect it to work next year. Then send it to two colleagues whose practice differs from yours, because the cuts that need volume are the ones your own answer does not fill.

If the survey questions prompt you to look harder at where agents could take work off your team, the KaneAI getting started documentation walks through authoring a first test from natural language and running it on the TestMu AI grid. Answer from what you have actually run, and the September report will be worth reading.