The AI Cost Curve Isn't What People Think

everyone talks about how cheap AI is getting. token costs drop every quarter, models get faster, inference gets commoditized.

the assumption is that this translates directly into cheaper software.

that assumption skips a few steps.

AI speeds up first drafts. but first drafts aren't the bottleneck in most real systems. the surrounding work expands as generation volume increases. validating output, triaging edge cases, integrating generated code with existing architecture. that work doesn't shrink. it grows.

the bottleneck shifts. it doesn't disappear.

AI delivers real value when the person operating it can define constraints clearly, spot subtle errors fast, and guide model tradeoffs. without that foundation, you accumulate plausible output that fails later in more expensive ways.

cutting experience to fund more generation is the wrong trade.

there's been pressure to remove mid-level engineers because AI can "do what they do." but those roles carry something models don't have: operational context. why the system is shaped the way it is. where the brittleness lives. what failed before and was quietly fixed.

removing that layer while increasing generated output volume creates a governance gap.

as AI embeds deeper into workflows, the overhead compounds. larger contexts, more retries, more control surfaces, more integration surface area. generation cost becomes one line item in a larger operating profile.

selection pressure does its work quietly. low-yield AI usage gets abandoned. high-yield workflows stick. the hype patterns fade and what remains is the subset that actually improved something measurable.

AI is a force multiplier for good systems thinking, not a substitute for it.