Version two is the real test of production-grade software.

A company recently told me they’d hired an AI development team to “deliver” what had previously been scoped as a three-year workflow project in a matter of months. They were excited about it, and honestly, who wouldn’t be? The assumption that AI will make development cheaper is a reasonable one when you can watch tens of thousands of lines of code being generated in the Anthropic “Usage” tab in a couple of hours.

I gently enquired about how the quality of the code was being measured and whether a data consolidation approach was being considered as part of the process, and there was a pregnant pause.

AI has made it genuinely easier to produce code quickly, and clients I speak with are drawing the obvious conclusion: surely you can deliver faster now and surely it must cost less? In many cases, the answer is yes. Code can be written fast, but the quality of that code is getting lost in that conversation, and the cost of that tradeoff tends to arrive at the worst possible time.

Speed Without Quality Is Debt

Speed without quality isn’t progress. It’s debt that compounds and comes due when the system is under load, when it’s in a client’s hands, or when a regulator is looking for some answers.

Nobody wants to be left behind when competitors are moving, and the pressure to adopt AI quickly is real. But the businesses genuinely winning aren’t simply the fastest. They’re the ones delivering well, at pace, without sacrificing the fundamentals.

This distinction matters more in AI-augmented development than it did before. When AI is writing significant portions of the code, the risk of accumulating structural problems beneath the surface increases. Code that passes a demo may fail completely when the architecture underneath doesn’t hold.

Testing Is What Makes Speed Sustainable

Testing harnesses can now run through the whole development process from day one, and not as a final check before delivery, and that shift changes everything.

A testing harness isn’t about slowing things down, but rather what makes it safe to go quickly. When every change risks breaking three other things, you can’t move fast without introducing risk. A solid testing foundation removes that constraint.

The difference between a system that scales and one that doesn’t is rarely visible in the first version. It shows up in version two, when you want changes, when the user base grows, when the edge cases start arriving. Version two is the real test of production-grade software. Organisations that invest in quality from day one find version two is manageable. Those that don’t find it’s a rewrite.

The Cost Problem No One Is Talking About Openly Enough

A separate conversation the same week raised something different, but related. A client asked how us to help them understand how were being charged for the use of their large language models. They genuinely didn’t know, and that surprised me less than how common it’s become. Organisations are committing to AI-powered workflows without a clear view of what costs will look like as usage scales. LLM pricing isn’t simple. It’s token-based, varies by model and context window, and architectures not designed with cost in mind can become surprisingly expensive in production.

The cost of running AI in production is absolutely something you can design for. The catch is that it has to be a design consideration from the beginning, not a problem you discover six months after launch.

Two Symptoms, One Root Cause

I believe that quality shortcuts and unplanned costs are different symptoms of the same underlying problem: companies are moving into AI-augmented development without a clear enough picture of what good needs to look like.

The focus seems to be on the headline statement of “faster, AI-powered, innovative”,  while the harder questions about architecture, testing, and cost get deferred or skipped entirely. I understand the pressure and we feel it ourselves. But the organisations making the most progress aren’t necessarily the most ambitious. They’re the ones going in with eyes wide open – a focus on how things work, not just that they work.