Reading Time: 6 minutes
This blog further develops the themes in my previous post on patchy AI adoption and it summarises my recent presentation at our AI Market, join us for the next one!
It is late 2025 and you already solved the “not enough GitHub Copilot licenses to go around” problem. You even sent your teams on high-quality training to make the most of these latest tools. Yet your business stakeholders are still waiting for the fireworks…and they just got back from a lavish conference where they saw how AI makes development 2x or 5x or 10x faster overnight…sound familiar?
Economics of AI in the SDLC
After training hundreds of individuals, including plenty of software professionals, over the past few years, it is dawning on me that AI in the SDLC has very little to do with the technology itself. Sure, the models behind some 2023-era products led to laughable suggestions and spawned many fair “auto-complete on dodgy steroids” comparisons.
Since then, there’s been an explosion of both interfaces (GitHub Copilot, Cursor, Claude Code, Codex…) and underlying models (including various flavours of “open”). And so now our work comes down to some simple-looking questions:
- WHAT do we choose to build?
- HOW should we build it?
- WHO does what along the way?
- WHEN will we pay off our tech debt, old and new?
And those questions, my friends, are all about economics. In preparation for what follows, let me ask you: are you running a scalable software factory, or an artisanal software workshop?
The productivity paradox is not new
It feels like he could have said it yesterday, but In 1987 economist Robert Solow famously quipped “you can see the computer age everywhere but in the productivity statistics”. Recent research shows that meaningful productivity gains from new tech require complementary investments 5-10x larger than the tech itself. And it takes 5-15 years before it becomes obvious that it all paid off…
When we ran experiments at DevOn comparing high-performance teams both without and with AI assistance, we found the latter could deliver almost twice as fast. But the biggest gain was not in code completion…any experienced engineer will tell you that typing has never been the constraint. The greatest unlock was in unblocking different members of a team to work simultaneously, i.e. to reinforce ways of working with positive feedback loops.
Three things are happening all at once
If this problem is sounding familiar, you’ll have already seen some version of the current market solution:
- Measure the most obvious outputs (% of code generated by AI)
- Which in turn overestimate the value of precisely what AI is best at (endlessly generating code)
- Which in turn feeds the hype (and doom) cycle, further boosting stakeholder expectations while frustrating your people
Meanwhile, you’re staring at the bills for $20/month/developer tools and wondering where the universally claimed “10-20% productivity gain” is going to come from.
The awkward truth, as both economic research and our experiment showed, is that most teams’ ways of working fritter away those gains in coordination overheads. And if we use AI for the wrong things (“more code, more features!”) we could end up increasing the human overheads that slow teams down.
Where to begin unpicking this
First clarify your context, are you:
- Standing safely behind moats (regulation, market concentration, special IP), in which case you have the time to work incrementally for 5-7 years towards 10-30% gains in the value delivered from software development
- Facing disruption, because an AI-native startup could build (at least the front end of) your product in a few months or years, in which case you have no time to waste
- Ready to lead because you have the engineering culture to adapt, and the resources to make those 5-10x larger complementary investments in changing everything else
Then pick your approach to match. At DevOn we’ve helped a 300-person software organisation get unstuck in the space of 4 months. Our internal experiment described above took our smoothest-delivering team out of their usual flow, because we knew they would adapt the fastest, and you cannot help the laggards until the leaders show what is possible. And now we are building agents that put DevOn’s almost two decades of experience “on the shoulder” of every new colleague, especially those newly joined, so they can get up to speed faster on how we deliver.
We’ve been here before, and will be again!
The Agile Manifesto exhorted us to value “individuals and interactions over processes and tools”. Almost a quarter-century later, AI is but the latest test of whether we really mean it. Just like agile, adopting AI in your SDLC only delivers value when you use it to change how you deliver software…it’s not about which copilot(s) to buy. Will you lead that transformation, or wait for what economists politely call “creative destruction”?
Wherever you are on this journey, history tells us you must push on, and DevOn is here to help. Let’s talk! Get in touch at ai.devon.nl today.
