Loading...

Skip to main content

Aliz AI Survey 2026 Q3 Results

· 16 min read
Sandor Muranyi
Sandor Muranyi
Principal Architect at Aliz

Internal survey · engineering AI adoption · wave 1 vs wave 2

Two identical-sized samples, four months apart. 42 responses in March 2026, 42 in July 2026. Every question in both questionnaires is charted below. Wave 2 added five questions and several new answer options; those are marked in green and have no baseline.

How to read the charts: grey bars = March 2026 · dark‑green bars = July 2026 · bright‑green bars = answer options new in wave 2.

Mean maturity level: 3.6 → 5.4 (+1.8 levels).At level 5 or above: 31% → 74% (13 → 31 people).Use no agentic features: 24% → 2% (10 → 1 person).Review every line: 57% → 26% (24 → 11 people).

01 — Who answered.​

Not a matched panel. Both waves are anonymous and self-selected, so some movement below is composition rather than behaviour change. Wave 2 skews more senior, more European, and introduced a Lead / EM grade that absorbed the previous Lead category.a

GradeTrack
Grade distribution, March 2026 vs July 2026: Architect 15 → 16, Senior 16 → 13, Medior 4 → 5, Junior 3 → 2, Lead / EM 4 → 6.Track distribution, March 2026 vs July 2026: AppDev 15 → 13, Infra / GCP 10 → 8, Data 9 → 7, ML 4 → 1, AI 2 → 4, Lead / EM / other 0 → 6.
ExperienceRegion
Experience distribution, March 2026 vs July 2026: 10+ yrs 19 → 23, 6-10 yrs 15 → 10, 3-5 yrs 6 → 7, 0-2 yrs 2 → 2.Region distribution, March 2026 vs July 2026: EU 33 → 36, APAC 9 → 6.

Wave 1 tracks are not perfectly comparable: wave 2 renamed AppDev to "AppDev (frontend, backend)" and moved six people into a Lead / EM bucket that did not exist before.

02 — Maturity moved almost two full levels.​

This is the headline. In March the distribution peaked at level 2, cautious assistant. In July it peaks at level 6, multi-agent. Nobody reported level 7 in March; four do now, plus three orchestrators.

Maturity level distribution, March 2026 vs July 2026: L1 0 → 0, L2 15 → 1, L3 5 → 1, L4 9 → 9, L5 10 → 11, L6 2 → 13, L7 0 → 4, L8 1 → 3.

One respondent pushed back on this question directly, arguing the ladder implies IDE work is "low-level" and penalises people who cannot hold context across parallel agents. Worth reading before treating the level as a performance metric.

03 — Usage frequency was already saturated.​

There was little headroom here. Daily-or-more use went from 90% to 98%. The change is not that more people use AI, it is that the people already using it went much deeper.

How often during a work dayConfidence to explain and debug own AI code (1–5)
Usage frequency, March 2026 vs July 2026: Constantly 26 → 30, Few / day 12 → 11, Few / week 3 → 1, Never 1 → 0.Confidence to explain and debug own AI code (1–5), March 2026 vs July 2026: score 1 = 0 → 2, 2 = 0 → 2, 3 = 8 → 4, 4 = 13 → 19, 5 = 21 → 15.

Confidence mean slipped from 4.31 to 4.02. The two new 1-scores are people who answered "I don't use AI to generate code", so the underlying dip among code-writers is smaller than it looks, but it is not an increase.

04 — Where the growth actually happened.​

The eight-task matrix is the most granular behavioural data in the survey and it contains the sharpest finding: code completion is the only task that did not grow. Autocomplete is flat at roughly half the team while every generative and comprehension task climbed. The team is not using AI to type faster, it is using it to produce and understand whole units of work.

Use this regularly or every possible time​

Tasks used regularly or every possible time, March 2026 vs July 2026: Code completion 56% → 55%, New features 78% → 88%, Debugging 76% → 86%, Writing tests 78% → 83%, Code review / refactor 68% → 79%, Documentation 68% → 90%, Learning unfamiliar code 80% → 88%, Non-coding tasks 41% → 64%.

Use this every possible time (the intensity measure)​

Tasks used every possible time (intensity), March 2026 vs July 2026: Code completion 20% → 36%, New features 41% → 62%, Debugging 22% → 52%, Writing tests 44% → 74%, Code review / refactor 24% → 45%, Documentation 49% → 71%, Learning unfamiliar code 49% → 55%, Non-coding tasks 12% → 19%.

The intensity view is the more dramatic one. "Every possible time" for writing tests went 44% to 74%, documentation 49% to 71%, new features 41% to 62%. Documentation is now the single most AI-delegated activity in the company, which is worth holding next to the slop concern in section 08.

05 — Agentic features went mainstream.​

Skills went from a niche 31% to the most-used feature at 88%. Every capability in the original list grew, and the five capabilities added in wave 2 arrived already in double digits. Only one person now reports using no agentic features at all, down from ten.

Agentic feature usage, March 2026 vs July 2026: Skills 31% → 88%, Instruction files 69% → 86%, MCP servers 40% → 71%, Parallel agents 26% → 60%, Memory / context 33% → 57%, Hooks 17% → 45%, Git worktrees 10% → 40%, Adversarial review (new) 36%, Cloud / remote agents (new) 19%, Scheduled runs (new) 17%, Goal loops (new) 17%, Mobile access (new) 14%, Uses none of these 24% → 2%.

06 — Tooling consolidated onto one vendor.​

Claude Code CLI overtook GitHub Copilot and roughly doubled. Copilot fell. Cursor collapsed from eight users to one. Anthropic models are now named by every single respondent, up from 86%.

Tools actively used​

Tools actively used, March 2026 vs July 2026: Claude Code CLI 45% → 76%, GitHub Copilot (any) 55% → 40%, Gemini chat / CLI 71% → 57%, ChatGPT / OpenAI chat 36% → 29%, Claude.ai web 21% → 36%, Claude Desktop (new) 33%, Cloud agents (new) 19%, OpenAI Codex CLI 5% → 17%, Cursor 19% → 2%.

Model providers used​

Model providers used, March 2026 vs July 2026: Anthropic 86% → 100%, Gemini 69% → 60%, OpenAI 52% → 50%, Open source / other 21% → 17%.

Read this next to section 10. Consolidation onto a single paid vendor is happening at the same moment that cost and quota complaints appear as a new unprompted theme. That is a procurement conversation, not an enablement one.

07 — Governance improved. Review rigour did not.​

Project conventions genuinely matured: AI-aware documentation nearly tripled, AI-specific review gates more than tripled, and "no conventions at all" fell by a third. But human review thinned out at the same time. Line-by-line review halved, and a pattern that did not exist in March, delegating review to a second AI, now accounts for roughly a quarter of answers.

Conventions specified in projects​

Conventions specified in projects, March 2026 vs July 2026: AI instruction files in repo 48% → 52%, AI-aware documentation 17% → 45%, AI-specific review process 7% → 24%, Verbal team agreements 21% → 19%, No conventions at all 43% → 29%.

How carefully AI code is reviewed​

How carefully AI code is reviewed, March 2026 vs July 2026: Review every line 57% → 26%, Skim + spot-check edge cases 26% → 24%, Check it compiles / passes tests 2% → 12%, Depends on the project 12% → 10%, Delegate review to an AI 0% → 24%, Rarely review / YOLO 2% → 2%, Don't use AI for code (new) 5%.

"Delegate review to an AI" is reconstructed from free-text answers in both waves; it was not an offered option in either. Zero people described it in March, ten did in July.

08 — The worry profile inverted.​

Security anxiety nearly halved and logic-bug worry fell, which reads like familiarity. But the concerns that track comprehension went the other way. Loss of codebase understanding is now the single most-cited concern, and "too much slop", a brand new option, debuted at nearly half the team.

Concerns about AI-generated code, March 2026 vs July 2026: Loss of codebase understanding 62% → 74%, Juniors not building skills 74% → 69%, Subtle logic bugs 88% → 71%, Too much code to review 48% → 57%, Maintainability decay 62% → 50%, Security vulnerabilities 67% → 38%, No governance standards 21% → 24%, Proprietary data exposure 17% → 12%, Too much slop (new) 48%.

Held together: more code, generated faster, reviewed less carefully, by people increasingly worried they no longer understand the codebase. Every individual movement is defensible. The combination is the risk.

09 — Quality verdict barely moved.​

Despite the enormous jump in capability and adoption, the perceived effect on code quality is almost unchanged. "Mixed, faster but harder to maintain" remains the modal answer. "About the same" emptied out entirely, with those people splitting toward both poles.

Overall quality impact of AI on code seen​

Overall quality impact of AI on code seen, March 2026 vs July 2026: Mixed: faster but harder to maintain 48% → 52%, Clearly better 19% → 21%, Can't tell 10% → 14%, About the same 10% → 0%, Worse 7% → 5%, Other / write-in 7% → 7%.

"Everyone should increase AI usage" (1–5)​

Everyone should increase AI usage (1–5), March 2026 vs July 2026: score 1 = 0 → 0, 2 = 0 → 2, 3 = 14 → 7, 4 = 11 → 11, 5 = 17 → 22.

Appetite for the mandate rose slightly, from 4.07 to 4.26. One architect objected to the framing itself: "Where to increase your usage, if you already use AI wherever possible?"

10 — The bottleneck moved from skills to contracts.​

This is the most actionable shift in the dataset. Demand for workshops and recipes, the top two asks in March, collapsed. What replaced them is commercial: clear agreements with clients debuts as the joint third-highest ask, unclear client acceptance appears as a new barrier, and cost or quota limits surface as an unprompted write-in from five people.

What enablement would help most​

What enablement would help most, March 2026 vs July 2026: Internal workshops / demos 64% → 40%, Curated recipes & playbooks 64% → 33%, Dedicated AI exploration time 40% → 40%, Pair-programming with a power user 31% → 31%, Clear company guidelines 31% → 24%, More tokens / quota 29% → 24%, Access to better tools 26% → 26%, Clear client agreements (new) 31%, I don't need enablement 7% → 12%.

Biggest barrier to using AI more effectively​

Biggest barrier to using AI more effectively, March 2026 vs July 2026: Lack of time to learn 52% → 50%, Workflow doesn't suit AI 7% → 19%, Security / compliance 17% → 19%, Nothing, well enabled 19% → 19%, Output quality too low 19% → 12%, Don't trust AI output 17% → 5%, Unclear client acceptance (new) 14%, Cost / quota limits (new) 12%, Don't know where to start 7% → 10%.

Lack of time is the one constant across both waves, at roughly half the team. It is the only barrier that has not moved at all.

11 — New in wave 2: no baseline exists.​

Five questions were added. Everything in this section is a first reading with nothing to compare against, so treat these as the wave 3 baseline rather than as a trend.

Use an Aliz AI account: 88% (45% also use a client account).Client restricts the provider: 50% (43% no, 3 never discussed).Use the Aliz AI Framework: 50% (48% no, 1 asked what it is).Perceived value multiplier: 6.0 mean (median 6, but bimodal).
Whose AI account is usedClient provider restriction
Whose AI account is used (July 2026): Aliz account 88%, Client account 45%, Private account 19%.Client provider restriction (July 2026): Yes, restricted 50%, No restriction 43%, Never discussed 7%.

What people think of the Aliz AI Framework​

What people think of the Aliz AI Framework (July 2026): Like it, needs improvement 40%, Don't know / not using 33%, Still evaluating 19%, Don't like it at all 7%.

"Using AI we can generate this much more value" (1–10)​

Perceived value multiplier "Using AI we can generate this much more value" (1–10), July 2026 responses: 1 = 0, 2 = 8, 3 = 2, 4 = 2, 5 = 3, 6 = 7, 7 = 6, 8 = 7, 9 = 1, 10 = 5.

The value question is bimodal, not centred. Eight people answered 2, five answered 10. The mean of 6.0 describes almost nobody. There are two populations here with incompatible views of what AI is worth, and averaging them hides the disagreement.

Compiled from the two Google Forms response sheets. 42 responses per wave. Wave 1 fielded 13–23 March 2026, wave 2 fielded 20 July – 30 July 2026. Percentages are of all respondents in that wave; multi-select questions sum above 100%. Free-text answers were coded into the nearest offered option where unambiguous and left out of the counts where not.

aliz.ai