AI Still Can't Code (But We Keep Pretending It Can)

Here's a statistic that should make every CTO pause: 90% of software developers now use AI tools daily, spending a median of two hours working with these systems. That's according to Google's 2025 DORA report. Impressive adoption rate, right?

Now here's the reality check: even the top-performing AI model - Claude 3.5 Sonnet from Anthropic (full disclosure: I'm rather fond of it) - completes only 26.2% of individual engineering tasks and 44.9% of management tasks.

Think about that for a moment. We're investing two hours per day, per developer, for a tool that successfully completes barely a quarter of individual tasks. This is what passes for revolutionary productivity improvement in 2025.

The Emperor's New Productivity Tools

I've been watching this AI coding assistant phenomenon with growing skepticism. The marketing pitch is seductive: pair developers with AI, get 25-30% productivity boosts. Some claims even suggest we're approaching autonomous software engineering.

The reality is considerably messier. Yes, some companies report productivity improvements of 10-15% when developers use AI assistants. But here's the problem nobody wants to discuss: that saved time isn't being redirected toward higher-value work. It's just... disappearing.

This reminds me of previous technology "revolutions" that promised to free up developer time. Remember when we said automation would eliminate repetitive tasks and let developers focus on creative problem-solving? We automated the repetitive tasks, but the creative problem-solving remained perpetually just out of reach.

What AI Actually Does

Let's be honest about what we're dealing with. Current AI coding tools are sophisticated pattern-matching systems. They're remarkably good at:

• Generating boilerplate code
• Completing common programming patterns
• Translating natural language descriptions into code that looks plausible
• Producing variations on solutions they've seen in training data

What they're demonstrably poor at is genuine software engineering: understanding system architecture, making design trade-offs, debugging complex interactions, and most importantly - reasoning about what code should do versus what it does do.

The shift from "AI as assistant" to "AI as co-creator" sounds impressive until you realize that a co-creator who fails 75% of the time isn't really creating - they're suggesting, and you're debugging.

The Economics Don't Add Up

Here's where it gets interesting from a business perspective. Companies are reporting that pairing generative AI with "end-to-end process transformation" yields 25-30% productivity improvements. Notice the qualifier there? End-to-end process transformation.

That's not the AI delivering value - that's organizational change delivering value, with AI as an excuse to finally fix broken processes. I've seen this playbook before with blockchain, with cloud transformation, with agile adoption. The technology becomes a Trojan horse for necessary organizational improvements.

Meanwhile, DeepSeek in China just demonstrated they can train comparable models at 70% lower cost than Western equivalents. This suggests the moat around AI development isn't technology - it's capital deployment. That's a fundamentally different competitive landscape than the one being marketed.

So What Should We Actually Do?

I'm not advocating abandoning AI coding tools. I use them. They're useful for specific, bounded tasks. But we need to be ruthlessly honest about their limitations:

First, stop expecting AI to do software engineering. It can't. It can assist with coding, which is a subset of engineering. The design, architecture, requirements gathering, trade-off analysis, debugging, testing strategy, deployment planning - that's still on you.

Second, measure the actual value, not the perceived productivity. If developers spend two hours with AI tools but you can't demonstrate improved delivery, reliability, or quality, you're not gaining productivity - you're adding overhead.

Third, recognize that AI coding tools are most useful for junior developers doing routine tasks, and least useful for senior engineers solving novel problems. Adjust your expectations and training accordingly.

Finally, remember that the real intelligence still needs to come from humans. AI tools are pattern-matching systems with impressive presentation. They're not thinking, reasoning, or creating. They're suggesting based on statistical correlation.

I wrote a few years back: "Be smart, because AI really isn't." That's even more true today, despite exponentially more compute power and training data. These tools have gotten remarkably better at appearing intelligent. But appearance isn't capability.

The 26.2% task completion rate for the best available model should be a wake-up call. We've achieved impressive technological progress in making AI coding assistants that can almost help with a quarter of real engineering work. That's genuinely useful. But it's a very long way from autonomous software engineering, and pretending otherwise wastes time, money, and developer trust.

Maybe the real productivity improvement would come from spending less time trying to coax AI into solving problems and more time actually solving them ourselves.

All Posts