83% of Teams Use AI, but Only 24% Ship Its Code to Production
AI-generated code appears in minutes, but review, fixes, testing and release still follow. According to a SimpleOne survey, 83% of teams use AI, while only 24% ship AI-assisted changes to production. In one example, a 16-hour task takes 11.5 hours with AI (a saving of about 28%). If the agent misreads the requirements, it takes 15.5 hours. So measure the benefit by the accepted change as a whole, not by generation speed.
AI-processed from Habr AI; edited by Hamidun News
A neural network can write code in half an hour, but that doesn't mean the change reaches users any faster. According to a survey by SimpleOne, 83% of development teams already use AI, yet only 24% say they ship AI-assisted changes to production. The rest get stuck in review, fixes and testing. Take a task that used to take 16 hours: AI generated the code in 30 minutes, then developers spent four hours reviewing it and another five hours fixing it. If you are introducing AI in your team, run these numbers against your own processes: what matters is not how fast code appears, but how long it takes to reach release.
What the survey showed
SimpleOne's study had 71 respondents: 46% development managers, 24% developers, 15% analysts and 15% other roles. 71% of respondents work at companies with more than 1000 employees. The sample is small and the survey was run by a vendor with its own platform, so treat it as a rough guide rather than market statistics.
At 41% of respondents' companies, individual employees experiment with AI on their own initiative, and another 18% run team pilots. Those same 24% reported regular use, where the results go to production. Another 23% could not say what share of AI-generated code and configurations reaches production. Such changes may be shipping, but nobody tracks them separately, and then a manager has nothing to prove the tool's value with.
68% named the risk of code or data leakage as the main obstacle, and 23% named regulatory requirements. Output quality worries 48%. Almost half (48%) expect a speedup of at least 1.5x, and another 31% cannot estimate the effect.
How the gain disappears
The authors give a simple calculation. A typical task takes 16 hours. With AI: two hours to describe the task, half an hour to generate, four hours to review and five to fix. That totals 11.5 hours, a saving of 4.5 hours, or about 28%.
Now suppose the agent misunderstood the requirements. Fixes take nine hours instead of five, for a total of 15.5 hours. Generation still takes half an hour, but the gain has almost vanished. And the cost of the model comes on top of that.
Tests are a separate trap. If an agent writes both the implementation and the tests from the same description, it can make the same mistake in both places. For example, a support manager asks to see the share of tasks resolved on time. The agent counts all closed tickets, including cancelled and rejected ones, writes a test for that calculation and passes it. The screen shows a figure that does not match the support team's rules. That is why acceptance criteria should be agreed before generation: the process owner decides which statuses count as resolved, and an engineer checks the calculation against a reference set of records.
Case study: a dashboard in 17 hours and $20
At a SimpleGen webinar, a product owner showed how agents took a task from requirements to installation on a test environment. The goal was a dashboard for a support manager: filters by period, group and task type, ticket flow, processing speed, bottlenecks, and target values with highlighted violations. The multi-hour generation was run in advance.
- The agent asked the author 28 clarifying questions and found a mismatch with the test environment: the specification described three task tables, while the environment had ten child tables.
- A second agent reviewed the specification; this took 11 iterations and 5.5 hours.
- Planning took another hour and a half: 11 phases and 42 tasks with dependencies.
- The test environment returned an error twice because of a long Russian-language prompt, and a conflict from re-creating an automatically generated form had to be resolved by syncing the environment with the repository.
- Code review found a wrong default value, insufficient access rights to the report and an incorrect label.
- In the browser, the agent completed 25 of the 28 scenario steps in seven minutes; three steps required authorization, which it does not perform under internal rules. It found an empty block and invisible tooltips.
In total, about 17 hours went into the project, including approvals and pauses. The package was installed on a clean test environment with no conflicts, errors or warnings. Claude Code consumed 786 330 tokens, and everything fit within a $20 monthly subscription with the Opus 5 model. Through the API, output tokens alone would have cost $19.66, and input tokens would have added an amount that depends on context size.
What to do in practice
The gap between practice and wishes is most visible in testing: today 26% of participants get AI help with it, while 61% would trust agents with it. For bug analysis the figures are 28% and 54%, and for code review 26% and 46%. AI writes code at 60% of respondents' organizations, while 46% would hand it over to an agent. Only 23% use AI at every stage of the lifecycle.
Take your last AI-assisted change and write down the time spent on task definition, generation, review, fixes and release. The code may have appeared in half an hour, with days going to clarifications and checks. That is where the next agent belongs, once you have defined in advance which repositories, test environments and data it can access. Keep the limitations in mind too: the vendor itself presented the case, and the 17 hours were not compared with manual development.
- Nov 19, 2026The same Gemma 4 wrote a backend that passed 80% of checks in opencode, but only 41% in OpenHands
- Nov 19, 2026OpenAI's coding agents would cost $2400 a day per researcher at API prices
- Nov 19, 2026AWS AgentCore Agents Can Now Reach a Private Keycloak Behind a Corporate CA
- Nov 19, 2026Smysl.AI: An Agentic RAG Over Company Documents, Built in Kotlin Without Python
Key Questions Answered
What the survey showed?
SimpleOne's study had 71 respondents: 46% development managers, 24% developers, 15% analysts and 15% other roles. 71% of respondents work at companies with more than 1000 employees. The sample is small and the survey was run by a vendor with its own platform, so treat it as a rough guide rather than market statistics. At 41% of respondents' companies, individual employees experiment with AI on their own initiative, and another 18% run team pilots.
How the gain disappears?
The authors give a simple calculation. A typical task takes 16 hours. With AI: two hours to describe the task, half an hour to generate, four hours to review and five to fix. That totals 11.5 hours, a saving of 4.5 hours, or about 28%. Now suppose the agent misunderstood the requirements. Fixes take nine hours instead of five, for a total of 15.5 hours. Generation still takes half an hour, but the gain has almost vanished. And the cost of the model comes on top of that. Tests are a separate trap.
What are the key details on Case study: a dashboard in 17 hours and $20?
At a SimpleGen webinar, a product owner showed how agents took a task from requirements to installation on a test environment. The goal was a dashboard for a support manager: filters by period, group and task type, ticket flow, processing speed, bottlenecks, and target values with highlighted violations. The multi-hour generation was run in advance.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.