Using AI in Software Development: What It Speeds Up, and What It Doesn’t

Every software company now says it uses AI. Very few explain what that changes for you. So here is a plain account of where it helps, where it does not, and what it should mean for your budget.

The research is mixed, and that is useful to know. In a 2023 controlled experiment by researchers at GitHub and Microsoft Research, developers using an AI coding assistant finished a set task, building a simple web server in JavaScript, 55.8% faster than those without one.

In 2025 the research group METR ran a different test. Sixteen experienced developers worked on 246 real tasks in large codebases they already knew well. With AI tools allowed, they took 19% longer. They had expected to be faster, and believed afterwards that they had been. METR is clear that those results are now out of date and cover one narrow setting, but the lesson holds: AI is a strong helper on some jobs and a distraction on others, and people are not always the best judges of which is which.

• First drafts of routine code. Login screens, forms, data tables and the plumbing every application needs.
• Tests. Writing the checks that catch mistakes is time-consuming, and AI is good at drafting them.
• Prototypes. Turning an idea into something you can click through in days, so you can react to it early.
• Documentation and clean-up. Explaining what code does and tidying what is already there.
• Exploring options. Trying two approaches quickly before committing to one.

• Understanding your business. It does not know why your approval process has an odd step, or which customers are the difficult ones.
• Deciding what to build. Scope, priorities and trade-offs still need an experienced person and a conversation with you.
• Security and data handling. AI can write code that looks right and is not. Every line has to be reviewed.
• Awkward integrations. The other system's quirks, outages and undocumented behaviour are not in any training data.
• Accountability. When something breaks at 7 pm, you want a person who understands the whole system.

Be wary of anyone who promises AI will make your project half the price. It will not, because writing code is only part of the work. Understanding, planning, testing, security and support all remain.

What you should expect is that the build itself moves faster, so a first version can reach you in weeks and not months, and that more of the budget goes to the parts that need judgement. That is why we describe it as an accelerator. Our own line is that AI is the accelerator and experienced engineering is what keeps it honest.

• Who reviews the code that AI produces, and how?
• What testing happens before I see anything?
• Where is my data going, and which tools see it?
• Where will AI not be used on my project?
• Do I own all of the code at the end?

On the last one, with us you do: ownership passes to you on final payment, and we sign an NDA with every client. The answer to the rest should be specific, not a slide.

Sources: The Impact of AI on Developer Productivity: Evidence from GitHub Copilot (arXiv) | METR: Measuring the impact of early-2025 AI on experienced open-source developers