Dude, Where's My AI Productivity Gains? - The Sports Car Paradox

One of the curious aspects of the AI revolution is that despite quickly being embraced by programmers of all stripes we haven't seen an exponential growth in productivity in software development. We haven't seen hundreds of new quality games, nor a new crop of must-use SaaS tools. Bugs have not disappeared from our existing tools, nor have waves of new useful features arrived. To the casual layperson there isn't much difference between the software they're using today vs what was around in 2020. Every major company's web page had a chat window back then too, and those were just as responsive as today's AI-powered variants.

What gives?

I'm a developer myself and at this point AI is fully embedded into my process for both work and personal development–as I write this Qwen 3.8 is building out a tool for me via Hermes in the other window. When I clock into work next week, Codex will be the first app I fire up.

Despite this, at work I'm probably delivering a similar amount of code as I did before the AI bubble.

To explain why imagine the sports car paradox:

You're tired of getting to work late, so you buy a performance super car. 0-60 times are sub 3 seconds, top speed is a street-illegal 200 mph. Yet despite all that extra horsepower you soon find your commute time hasn't radically changed. The real problem was the traffic bottlenecks you consistently encounter day after day. Having a fast supercar doesn't mean you get around backups at the tunnel, it doesn't skip that long red light, or get you through rush hour gridlock any faster. It means that maybe you get to race a little faster on the open roads. Ultimately that's a negligible gain (and for the extra performance comes the possibility of running afoul of the law, or getting in more serious crashes, but I digress).

In software, particularly software that powers user interfaces, there are a lot of bottlenecks too. Even with much of the process standardized via tools like linters and code formatters, there are an infinite number of ways to satisfy even simple acceptance criteria. The difference between a bad interface and a good one can be subtle and take days of human interaction and observation to suss out. It's not as simple asking whether code compiles, how compliant it is with standards, or how fast it all loads. It's beyond that to questions of whether the tool facilitates users, conversions, sales.

Then there's organizational structure. No sane organization of any size would enable new features in production without some degree of testing and introspection.

In the mid 00's Agile swept through the industry, and while it was never fully or completely adopted the 2~4 week sprint cadence caught on as a way to sync the entire organization, not just developers, to the current state of a project. In the early days of AI adoption, that cadence has still been a valuable metric for assessing project status and whether a feature is still on target.

Even in a slimmed down Agile process, the various ceremonies that allow an organization to estimate remaining work, refine tasks, and communicate priorities throughout an organization take time. AI making development faster has no bearing on that. You're racing to hit a bottleneck that exists outside of the act of writing code.

Likewise the process of assessing not just whether code works, but if it is valuable to customers remains slow and tedious. Processes like A/B Testing can speed up decisioning for features or creatives, but entrusting an entire application to them is a fools errand.

In all the AI hype talk I've been privy to, no one's ever considered a new project management structure to account for AI. I don't know if there is one. Stakeholders will always want to see and interact with the tools being built on their behalf. They're as likely to trust AI with that as they are with their human assistants.

I don't know what an AI-enhanced project management system would look like in practice. I couldn't say whether such a thing would be well-advised. That agile cadence is really good at what it does and its failures are more due to organizational adoption than implementation. Still, without such a rethink, projects will continue to take the same amount of time to complete in a professional environment.