AI coding: what's actually happening

A bespoke suit from Savile Row is cut by hand, fitted through multiple sessions, and constructed to last fifteen years. It costs several thousand pounds and takes weeks. The tailor’s expertise is in every decision: the cut, the fabric, the construction.

An off-the-peg suit is designed by professionals but manufactured efficiently. It costs a few hundred pounds and you can walk out wearing it. By any objective measure of craftsmanship, it’s not as good. But for most professional contexts, it’s good enough. And the value proposition is completely different.

Something similar is happening in software development, and many executives are seeing it through one of two distorted lenses.

The first is breathless enthusiasm:

• “Build a startup in a weekend!”
• “Fire your developers!”
• “Anyone can code now!”

This narrative sells courses and generates clicks. It also sets expectations that lead to disappointment when reality turns out to be more nuanced.

The second is dismissive scepticism:

• “It’s overhyped.”
• “The quality isn’t there.”
• “Our team tried it and weren’t impressed.”

This feels like hard-won wisdom from people who’ve seen technology fads before, but that can make it even more dangerous. Banning AI is not a valid commercial strategy.

In case you’re wondering, you might want to read Why I am writing about this, which might also answer the question “Why should I listen to you anyway?”.

What’s actually happening

AI coding tools, particularly the agentic variety that can write, test, and iterate on code based on specifications, are producing significant productivity gains. In teams I’ve worked with, we’ve measured improvements of around 70% using relatively light tooling: code completion, prototyping assistance, automated test generation.

These aren’t impressionistic estimates. They’re measured in story points delivered per available resource, adjusted for quality by tracking rework. The methodology isn’t perfect, but the delta is large enough to be meaningful even accounting for measurement noise.

Going a step further, I’ve tried it myself. Using more capable agentic tools, I’m building systems in two weeks that would previously have taken me two months or more.

The variance matters. That range, from 70% team improvement to 4x individual acceleration, isn’t random. It depends heavily on the operator.

A scientific paper - a robust, pre-registered randomised control trial - published in December 2025 showed a 55% productivity gain for experienced AI coders, with no decrease in maintainability (a good measure of good quality), but a smaller 30% speed improvement when it was a less AI-skilled developer doing the same task.

The operator question

The developers seeing the largest gains aren’t the fastest typists or the ones who’ve memorised the most syntax. They’re the ones with breadth: product thinking, architecture, testing strategy, deployment, user experience. They can collapse multiple roles because they understand the full picture. They know what questions to ask and can recognise when an answer is wrong.

A developer whose primary skill is translating specifications into code sees smaller gains. The AI competes directly with that skill. Someone who can specify precisely, validate thoroughly, and course-correct effectively has the AI amplifying their judgment rather than replacing their typing. This is the bespoke tailor dynamic. The value isn’t in the sewing. It’s in the eye, the measurement, the ability to look at a client and know what will work. It’s experience. AI-assisted development still requires that judgment. It just accelerates everything that comes after.

For executives, this has implications for hiring and team composition. The profile of a high-leverage AI-assisted developer looks different from the profile you might have hired for just 18 months ago. Breadth matters more. Pure coding speed matters less. And remembering syntax or a particularly API, almost not at all.

The quality objection

The sceptical view often centres on quality. AI-generated code isn’t as good as hand-crafted code. This is true, in the same way that an off-the-peg suit isn’t as good as bespoke. But it misses two things.

First, the quality is high enough for most purposes. Not every system needs to be engineered for a fifteen-year lifespan. Most business software has a shorter useful life than people assume, because requirements change, markets shift, or something unexpected happens that makes replacement more sensible than repair. You might buy a suit engineered to last a decade, but if you spill red wine down it next year and the dry cleaner can’t save it, longevity was never the point. And let’s not forget, humans make mistakes too, especially when they’re tired, or under-pressure, or fed up, or they feel they’re not paid enough. Not so with AI.

Second, the quality gap is shrinking fast. The tools available today are noticeably better than six months ago. Objections formed from experience last year are already dated. Whatever limitations exist now (I’m writing this in January 2026) will look quaint by the end of this year.

The cost of dismissal

Companies treating this as a fad are making the same mistake, though bigger, as those who dismissed cloud computing in 2008 or mobile in 2010. The early versions were clunky. The mature versions changed everything. We’re in that transition now, and the pace is faster and much more impactful than those previous shifts.

The sceptical position feels safe. It isn’t. Every month of delay widens the gap between companies building this capability and companies waiting for more evidence. The evidence is already here. It’s just unevenly distributed.

What this means for you

If you’re a CEO, CFO, or board member, you don’t need to understand the technical details. You need to understand that the productivity gains are real and significant, that realising them requires skilled operators rather than just tools, and that the window for building advantage is open now.

In the next article, I’ll cover what happens when software development costs drop by 50-80% or more, including some assumptions about how software works that are so deeply embedded you may never have questioned them.

Next: AI coding: the new economics of building software

AI Coding Series Index