How AI Is Changing Software Development — What It Means for a Tester

I recently watched a webinar about how AI is changing software development, with engineers from Atlassian, DX, Dropbox, Vercel, and Lovable. Here's a short, simple recap, because this affects every junior's future job, testers included.

Everyone is already using AI

This isn't a trend anymore. Almost 100% of the Dropbox team uses AI tools every week. At DX, about 95% of developers actively use them — and even the remaining 5% still ship code that AI helped write.

Code gets written faster, but that doesn't double the business value

At Dropbox, about 60% of the code that reaches production was written with AI, and the number of pull requests (a PR is when a developer proposes a code change for review) nearly doubled. But the number that shows how much genuinely new, useful stuff shipped only grew by about 1%. So code gets written faster, but coming up with useful things and following through hasn't sped up much.

The slow part moved — straight into testing

Writing code used to be the slowest step. Now code gets written quickly, while review, testing, deployment, coordination, and waiting for builds have become the visible slowdowns instead. That's good news if you're a tester: this is exactly where people who can check work carefully and quickly are needed most right now.

Trust is growing slower than the code itself

Dropbox noticed something odd: the code became easier to understand and work with later (a metric called maintainability went up), but confidence that a change won't break anything actually went down. The reason is simple — AI produces more code and bigger pull requests, and a big chunk of code is harder and riskier to review carefully than a small one.

Context is basically a spec — but for AI

If a task is explained vaguely, the AI will either guess, do the wrong thing, or two AI agents will end up with conflicting instructions that don't add up. The worse the task is explained (the "context"), the more rework happens and the more tokens get wasted (tokens are the units AI usage gets billed in, similar to minutes on a phone plan).

A human is still in charge

AI can write code and even do some research on its own, but deciding what actually matters, judging quality, and understanding why something is being built at all is still a human's job. Someone still has to look at the plan, look at the result, and say "good, ship it" or "no, redo this."

How companies actually check if AI is helping

• Adoption: how many people actually use AI every week
• How developers feel: surveys and self-reported hours saved (about 6 hours per engineer per week at both Dropbox and DX)
• Delivery speed: how many pull requests get made, how often releases ship, how long it takes to go from an idea to finished code
• Quality: confidence in changes, how effective reviews are, test coverage, how many bugs make it to production
• Business value: how much genuinely new and useful stuff shipped, not just "how much code got written"
• Money: how many tokens get spent, how much a single pull request costs — some companies track this as strictly as cloud spending

The bottom line is simple: strong companies don't measure AI by asking "how much code did it write." They look at the whole system — speed, quality, cost, and whether the actual end user got any real benefit. For a tester, that means one thing: quality checking is becoming more important, not less.