Analysis
AI-assisted software development: how AI in requirements, code and testing helps IT projects succeed — and where it does not
AI-assisted software development raises the odds of a successful IT project only when it strengthens a good process. In field experiments at three companies, developers with a coding assistant completed 26.08% more tasks on average, yet experienced developers working in mature projects took 19% longer with AI. The DORA report puts it plainly: AI does not fix a team, it amplifies what is already there.135
AI in software development is usually discussed in terms of speed: how many lines an assistant writes and how much sooner the team finishes. For a company commissioning an IT project, that is the wrong question. A project succeeds not because the code was written quickly, but because the system does what the business needed, runs reliably and can be maintained.
IT projects are rarely lost on the code itself. They are lost on vague requirements, testing pushed to the end and knowledge that lives in one person’s head. That is exactly where AI helps — provided a person still reviews every change and the process has automated tests. A team without that discipline simply produces bugs faster with AI.
— Dariusz Wtykło, owner and CEO of Primessio
Why IT projects fail
Flyvbjerg and Budzier studied 1,471 IT projects: the average cost overrun was 27%7. The tail is more dangerous than the average — one project in six overran its budget by 200% on average and its schedule by almost 70%7. Projects like that are rarely sunk by a single line of code. They are sunk by misunderstood requirements, risks spotted too late and dependencies on systems nobody documented.
That tells you where AI can genuinely help. If the assistant only speeds up typing code while requirements, tests and system knowledge stay the same, the project simply reaches the same problems sooner.
What the trials say about AI for coding
The strongest evidence comes from experiments in which some developers got a coding assistant and others did not. Results differ because each study measured different tasks and different people:
| Study | Who and what | Result |
|---|---|---|
| Peng et al., 2023 | a single programming task under experimental conditions | task completed 55.8% faster2 |
| Cui, Demirer et al., 2025 | 4,867 developers at three companies, everyday work | 26.08% more completed tasks on average1 |
| METR, 2025 | 16 experienced developers, 246 tasks in their own mature projects | 19% longer with AI; they had expected a 24% speed-up3 |
| DORA, 2024 | survey of technology teams | as AI adoption rose, delivery throughput −1.5%, stability −7.2%4 |
The conclusion is not “AI speeds up programming”. It is more useful than that: AI helps most where a developer does not yet know the code or the domain, and least where an expert works in a system they know by heart. Less experienced developers reached for the assistant more often and gained more1. A year later DORA already found a positive relationship between AI adoption and throughput, while the relationship with delivery stability stayed negative5.
Where AI helps across the project lifecycle
Requirements and acceptance criteria
A language model turns workshop notes, emails and existing documents into a requirements list, flags contradictions and asks about edge cases. A vague “the system should send reminders” becomes criteria you can test. Which requirement matters is decided by the client and the team — AI shortens the road to that conversation.
Code and code review
The assistant writes repetitive parts, proposes changes and supports review by pointing to places without error handling or out of step with the rest of the system. The person who approves a change owns it.
AI in software testing
AI speeds up writing unit tests and end-to-end scenarios and finds cases nobody thought of. Tests do double duty here: they check code written by people and are the safety net for code proposed by AI.
Documentation
In most projects documentation loses to deadlines. AI drafts it from code, change history and tickets, and the team corrects it. Less spectacular than generated code, but it directly reduces the risk that system knowledge leaves with one person.
Legacy systems
Legacy systems whose authors have long left the company are a frequent source of delay in integrations and migrations. AI reads such code, explains what it does, maps dependencies and proposes tests that pin down behaviour before a change. The explanation has to be verified, but it is a far better starting point than a blank page.
AI in project management
AI in project management prepares status reports, meeting minutes, risk registers and progress-versus-plan comparisons. The project manager spends less time retyping and more on the conversations and decisions no model makes for them.
Where AI does not help — or hurts
- Code security. About 40% of programs an assistant generated in high-risk scenarios were vulnerable8. Without review and scanners, AI also speeds up shipping security holes.
- Trust in the output. Developers themselves are cautious: more of them actively distrust the accuracy of AI tools (46%) than trust it (33%)6.
- Delivery stability. Faster changes without strong tests and small, frequent releases reduce stability4.
- Perceived speed-up. After the METR study, participants estimated AI had cut their time by 20%, although they had actually taken longer3. A team’s impression is not a measurement.
- Technology before the problem. AI projects most often fail when the team misunderstands the problem, lacks the necessary data or focuses on technology rather than the problem9.
Rules that keep a project under control
- A person who understands the consequences approves every change proposed by AI.
- Only code that passes automated tests and quality checks in the build reaches production.
- Releases are small and frequent — easier to check, easier to roll back.
- Approved AI tools and rules for client code and data (including GDPR) are agreed before the project starts.
- We measure the effect of AI on our own tasks (time, rework, defects after release) instead of borrowing it from surveys.
How we work at Primessio
AI models are part of our daily work on requirements, code, reviews, tests, documentation and existing systems. We describe the rules in the section AI in software development on our custom software and IT systems page: a person reviews every change, and end-to-end tests plus accessibility, SEO and structured-data checks run on every release.
Where to start
If your company is planning an IT project in which AI will be a tool for the team or part of the system itself, start with a process map and a single task whose outcome can be checked. The Primessio AI Implementation Framework walks through this step by step — a method developed by Primessio in project practice, from process map to maintenance. If you would rather take that path with a team that does it alongside you, see our AI automation and AI implementation services.
Frequently asked questions
Does AI-assisted coding make IT projects faster?
It depends on the task and the team. In three field experiments with 4,867 developers, those using a coding assistant completed 26.08% more tasks on average. In a study of experienced developers working in their own mature projects, work with AI took 19% longer, even though participants believed they had sped up. Measure the effect in your own project rather than assume it.13
Where does AI in software development help most?
Where the result can be checked quickly: requirements and acceptance criteria, tests, documentation and reading legacy code that nobody knows well any more. Less experienced developers adopted the assistant more often and gained more.1
How is AI used in software testing?
AI drafts unit tests and end-to-end scenarios and suggests edge cases the team missed. Tests then do double duty: they check code written by people and act as the safety net for code proposed by AI. AI-generated code is not safe by default — about 40% of programs an assistant produced in high-risk scenarios contained vulnerabilities.8
Will AI in project management replace the project manager?
No. AI produces good drafts — risk register, status report, meeting minutes, scope-versus-requirements checks — while decisions and accountability stay with people. RAND finds that AI projects most often fail when the team misunderstands the problem or puts technology ahead of it; no tool solves that for the project manager.9
How does Primessio use AI on client software?
AI is part of our daily work on requirements, code, tests and documentation, but a person reviews every change and only code that passes automated tests reaches production. The rules for using AI with your code and data, including the list of approved tools, are agreed before the project starts.
Sources
- MIT Economics (working paper; Management Science): The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers (Cui, Demirer, Jaffe, Musolff, Peng, Salz)
- arXiv: The Impact of AI on Developer Productivity: Evidence from GitHub Copilot (Peng, Kalliamvakou, Cihon, Demirer, arXiv:2302.06590)
- METR / arXiv: Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (METR, arXiv:2507.09089)
- Google Cloud (DORA): Announcing the 2024 DORA report
- Google Cloud (DORA): Announcing the 2025 DORA Report: State of AI-assisted Software Development
- Stack Overflow: Stack Overflow Developer Survey 2025 — AI
- Harvard Business Review / arXiv: Why Your IT Project Might Be Riskier Than You Think (Flyvbjerg, Budzier; Harvard Business Review 89(9), arXiv:1304.0265)
- IEEE S&P / arXiv: Asleep at the Keyboard? Assessing the Security of GitHub Copilot’s Code Contributions (Pearce et al., IEEE Symposium on Security and Privacy 2022)
- RAND Corporation: The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed