AI in Verification & Validation: What 28 Engineers Told Us in 2026

9 min read

Summarize this survey with:

Between 19 October and 12 November 2026, 28 V&V, QA and systems engineers told us how they use AI at work, where it fails them and what they want to automate next. 18 of them work on safety-critical software: aerospace, defence, space, automotive, energy and medical devices.

The short answer: AI is already everywhere around the work, in emails, summaries and log analysis. It is still rare in the evidence itself. 71% of respondents want AI to work between the requirement and the test next: clarifying requirements, keeping traceability, finding missing cases. Below are the numbers, what the respondents said in their own words, and the practices that answer each problem.

Key findings

engineers answered (V&V, QA, systems, research)
28
work on safety-critical software (18/28)
64%
want AI between the requirement and the test next (20/28)
71%
engineers who do not use AI at all work on safety-critical software
4 of 4
  1. AI is used around the work more than in it. The top uses are emails and meeting notes (57%), summarizing specifications (50%) and analyzing logs or defects (50%). Only 14% use it for system modeling.
  2. Safety-critical teams use AI much less for evidence. Only 2 of 18 safety-critical engineers use AI for requirements-to-test traceability, against 7 of 10 in IT and finance.
  3. The first blocker is not quality, it is access. 21% say they are not allowed to use AI on their real documents, or cannot share them.
  4. When AI does run, the main complaint is generic output: test cases that look right but miss edge cases, timing constraints and safety requirements.
  5. The next wish list is clear: clarify vague requirements (32%), keep traceability up to date (29%), derive test cases from requirements (25%) and find the missing cases (21%).

Who answered

Respondents were invited one by one on LinkedIn, based on their role. Most work in V&V, IVVQ or QA, and close to two thirds in regulated industries.

Industry

  • Aerospace, defence and space36% (10/28)
  • Automotive21% (6/28)
  • Software, IT and telecom18% (5/28)
  • Finance, banking and insurance18% (5/28)
  • Energy (nuclear)4% (1/28)
  • Medical devices4% (1/28)
Single choice. Aerospace groups aviation, defence and space. Safety-critical = aerospace, automotive, energy and medical (18/28).

Primary role

  • V&V / IVVQ engineer or lead39% (11/28)
  • QA / test engineer32% (9/28)
  • Systems engineer21% (6/28)
  • Researcher (MBSE, V&V)7% (2/28)
One respondent is both systems and V&V engineer; counted as V&V.

Tools are the usual mix. Jira is named by 19 of 28 respondents, IBM DOORS by 10 and Excel by 9. Then Python, Xray, Confluence and Word, and the unit-test tools of safety-critical software: RTRT, LDRA, VectorCAST, Tessy and Cantata. Requirements, tests and evidence live in several tools at once, which is exactly where traceability breaks.

What engineers use AI for today

Tasks where respondents already use AI

  • Emails, meeting notes, translation57% (16/28)
  • Understanding or summarizing specifications50% (14/28)
  • Analyzing logs, failures or defects50% (14/28)
  • Designing test cases43% (12/28)
  • Writing reports, reviews or qualification documents39% (11/28)
  • Writing or reviewing requirements36% (10/28)
  • Writing or reviewing code36% (10/28)
  • Writing test procedures32% (9/28)
  • Writing automated test scripts32% (9/28)
  • Traceability (requirements to tests)32% (9/28)
  • Architecture or system modeling (SysML, Capella)14% (4/28)
  • Do not use AI at work14% (4/28)
Multiple choice, share of all 28 respondents.

What works today is simple. Most people cite clear prompts and reusable templates for each type of task. Several use Claude, ChatGPT, Gemini or Copilot as a reading assistant: to learn a system without reading the whole specification, or to summarize a meeting. The more advanced users build agents, connect tools through MCP servers or use retrieval (RAG) over their specifications. One engineer uses a cheaper chat model to write the prompts for a coding agent, to save budget for the deep work.

This matches the wider market. In the World Quality Report 2025 (Capgemini, OpenText, Sogeti, more than 2,000 executives), 89% of organizations run generative AI pilots or deployments in quality engineering, but only about 15% have scaled it across the enterprise.

The safety-critical gap

Splitting the answers in two groups shows the biggest difference in the survey. Safety-critical engineers use AI for fewer tasks (3.5 on average, against 5.6 in IT and finance), and almost never for the evidence that an auditor will read.

Share of each group using AI for the task

Safety-critical (n=18)IT and finance (n=10)
  • Traceability (requirements to tests)Safety-critical: 11% (2/18)IT and finance: 70% (7/10)
  • Writing automated test scriptsSafety-critical: 17% (3/18)IT and finance: 60% (6/10)
  • Analyzing logs, failures or defectsSafety-critical: 39% (7/18)IT and finance: 70% (7/10)
  • Writing or reviewing requirementsSafety-critical: 44% (8/18)IT and finance: 20% (2/10)
  • Emails, meeting notes, translationSafety-critical: 61% (11/18)IT and finance: 50% (5/10)
Small groups: one person is 6 points in the first group and 10 points in the second. Read the gaps, not the decimals.

Two things stand out. First, safety-critical engineers do use AI on requirements (44%), more than IT teams. Reading and challenging a requirement is low risk: an engineer still decides. Second, they avoid AI where the output becomes certification evidence. A traceability link or a test that goes into a DO-178C or ISO 26262 file must be reproducible, reviewed and defensible years later.

That caution is rational. Under DO-178C, a tool whose output is not verified may need qualification under DO-330 (see our DO-178C guide). EASA's AI Concept Paper Issue 2 (2024) and ISO/PAS 8800 (2024, AI in road vehicles) now frame how AI must be assured. The practical rule for teams today: let AI propose, and keep a human review that is itself the verification.

Where AI fails, in their words

Main reason AI did not work (free text, coded by theme)

  • Not allowed, or data too confidential to share21% (6/28)
  • Generic output, missed edge cases, wrong assumptions18% (5/28)
  • Too complex: multi-step, multi-interface or large tasks18% (5/28)
  • Missing context: no documentation, little data14% (4/28)
  • Results not reproducible enough for safety evidence7% (2/28)
Open question. One answer can carry several themes. 2 respondents gave no usable answer.

“For security, I can't share the documents, so I have to give just examples of what I need to do.”

Systems engineer, energy (nuclear)

“AI sometimes generates test cases or code that look correct but do not fully match the actual API behavior or business requirements.”

QA / test engineer, software

“When the task is too elaborate, the created test cases are not efficient and require a lot of rework.”

QA / test engineer, finance

The edge-case problem is documented in research too. In a 2025 study of LLM-generated high-level tests (arXiv 2503.17998), completeness of edge and negative cases was the main weakness, for models and for human authors. In a study of system tests written by ChatGPT from requirement specifications (arXiv 2412.03693), about 72.5% of generated test cases were judged valid by the developers. Useful, but not a complete test set.

Confidentiality is the same story at a larger scale: in the World Quality Report 2025, data privacy risk (67%) and hallucination and reliability (60%) are the top two barriers to generative AI in quality engineering.

What engineers want to automate next

What respondents want AI to do next (free text, coded by theme)

  • Clarify vague or complex requirements, check testability32% (9/28)
  • Keep requirements-to-test traceability up to date29% (8/28)
  • Derive test cases from requirements or code changes25% (7/28)
  • Find missing cases, limit conditions and coverage gaps21% (6/28)
  • Write documents, V&V plans and summaries11% (3/28)
  • Analyze CI failures and fix pipelines7% (2/28)
20 of 28 respondents (71%) named at least one task between the requirement and the test.

“The key challenge is ensuring that AI-generated outputs are accurate, verifiable, and traceable to the original requirements.”

V&V engineer, aerospace

“Deriving test cases with relevant expected results and rationale from it, all while keeping traceability.”

V&V engineer, aerospace

The wish list is an order of operations. First make the requirement testable. Then derive the tests it needs. Then keep the links alive when either side changes, and show what is missing. That is the core of requirements-based V&V, and it is the part that today still lives in spreadsheets.

A practical playbook for each problem

Each problem the respondents raised has a known answer. None of them needs a new tool to start.

1. Your data cannot leave the company

Run an open-weight model on your own machine or server. A recent model on a single workstation GPU is often enough to read and challenge requirements. Our free Local LLM Finder tells you which model fits your hardware. For a team, ask for an on-premise or air-gapped deployment, and check where prompts are logged. Our security page lists the questions to ask any vendor.

Two open-source projects made this much easier in the autumn of 2026. llama.cpp v1.0.0 (13 November 2026) gives a stable API, a host-memory cache for mixture-of-experts models, FP8 KV caches, and a hardened llama-server (SSRF, CORS and DNS-rebinding fixes). That makes it a sound base for a small internal server. Strata runs a 125B mixture-of-experts model (Qwen3.8-Flash-Next) on one PC with a 12 GB GPU and 32 GB of RAM, behind a local OpenAI-compatible API. It is still a young v0.1 project for one user, not for a shared team server, so test it on non-critical data first.

2. AI test cases miss the edge cases

When you ask an AI for "test cases for this requirement", it writes the most likely behaviour. Edge cases are exactly where the text is silent, so the AI guesses there, with confidence. A better prompt helps a little. The real fix is to stop going straight from text to tests. Model-based testing (MBT) puts a small model of the expected behaviour in the middle, often a state machine with modes, events, thresholds and timings. The tests then come from the model. The ideas below work with or without AI.

  • Let AI read, let code decide. A language model is good at reading a spec and proposing a model. Deterministic code is better at producing the tests from it: the same model gives the same tests every time, and an auditor can reproduce them.
  • Every element of the model points to its source. Each state, threshold or timing cites the sentence of the spec it comes from. An element nobody can cite is a question for the author, not a test.
  • Ask the author concrete questions. "Is this requirement ambiguous?" gets a shrug. "The driver presses the button at exactly 30 km/h: does cruise control engage, yes or no?" gets an answer in ten seconds.
  • Check the rules before you test them. Formal tools such as SMT solvers can show that two rules can fire at the same time, or that no rule covers a case. DO-333, the formal methods supplement to DO-178C, is the frame for using this as certification credit.
  • Name the coverage. Every transition, the values just below, at and just above each threshold, the relevant combinations of a decision table, and each timing just inside and just outside its tolerance. A named criterion tells you when you are done.
  • Test your tests. Put small deliberate faults in the model (a wrong threshold, a missing transition, a late timer) and check that at least one test fails for each. This is mutation testing. It measures test strength better than a count of test cases.

Structured requirement styles such as EARS ("When <trigger>, the <system> shall <response>") make the model step much easier. So do the SysML or Simulink models that many teams already have.

3. Vague requirements break everything downstream

Review requirements before you generate any test. Look for five defects: inconsistencies, ambiguous words ("fast", "large"), conflicting constraints, unstated assumptions and missing edge cases. Give the model examples from your own project: in an industrial study presented at ICSME 2025, ten examples in the prompt raised ambiguity classification by about 20% over no examples. More on this in how to review a code change against its requirements.

4. Safety evidence must be reproducible

  • Pin the model version and set temperature to 0. Outputs can still vary, which is why the next step matters.
  • Store the prompt, the input, the output and a hash of each, with the date.
  • Treat every AI output as a proposal. The engineer's review is the verification, and it is logged.
  • Never let AI write directly into the baseline in DOORS, Polarion or your test tool.

5. Traceability goes stale

Let AI propose links and let an engineer approve them. Before you trust a tool, measure it on a part of your project where the links are already known: count the right links it finds and the wrong links it adds. Our TraceFuse benchmark shows how we measure this, and what an AI traceability matrix should contain. Then, after each change, re-check only the requirements and tests that changed.

6. CI failures take hours to read

Do not send the whole log to the model. Filter it to the error lines and their context first. ByteDance's LogSage (2025) does exactly this and reports 98% precision on a curated set of 367 CI/CD failures, and over 80% end-to-end precision in a year of production use. A small script that extracts failed steps and stack traces gets you most of the way.

7. Everyone keeps prompts to themselves

Put your team's prompts and templates in the same repository as your tests, and review them like code. To share tools between agents, use MCP servers: one server can give Claude, Cursor or Copilot the same access to your requirements and tests. See our MCP server and developer resources.

Methodology and limits

  • Online questionnaire (Google Forms), open from 19 October to 12 November 2026, 28 complete answers.
  • Respondents were contacted one by one on LinkedIn, based on their V&V, QA or systems role. They chose to answer: this is not a random sample.
  • Percentages are shares of all 28 respondents unless a chart says otherwise.
  • Open answers were read and grouped into themes by the KomAInu team. One answer can belong to several themes.
  • Answers are anonymous. We only quote people who allowed it, and we edited quotes for grammar only.
  • With 28 answers, read the results as signals from practitioners, not as market statistics.

Frequently asked questions

How do V&V engineers use AI in 2026?

In our survey of 28 V&V, QA and systems engineers, the most common uses are emails and meeting notes (57%), summarizing specifications (50%) and analyzing logs or defects (50%). Only 14% use AI for system modeling, and 14% do not use AI at work at all.

Can AI be used for safety-critical software testing?

Yes, as a proposal tool with human review. Safety-critical engineers in our survey use AI to read and challenge requirements (44%) but rarely for traceability evidence (2 of 18). Under DO-178C, an engineer's verification of the output is what keeps the tool from needing DO-330 qualification; check the approach with your certification authority.

Why do AI-generated test cases miss edge cases?

Because the model only sees the requirement text and fills gaps with typical behaviour. Ask it to list what the requirement leaves undefined first, then request normal-range, boundary and robustness tests separately, one requirement at a time.

What do test engineers want to automate next with AI?

71% of respondents named a task between the requirement and the test: clarifying vague requirements (32%), keeping traceability up to date (29%), deriving test cases from requirements (25%) and finding missing cases (21%).

How can I use AI if my documents are confidential?

Run an open-weight model locally or on a server inside your network, or use an on-premise or air-gapped deployment. 21% of our respondents could not use AI on real documents for this reason.

Try it on one of your requirements

Bring a requirement that gave your team trouble. In 30 minutes we show you the questions it leaves open, the test cases it needs and the ones you already have.

Or try KomAInu Individual for free