A new paper about building AI-assisted software just corrected its own headline number by half.
A six-person student team spent one academic term building a full conversational onboarding assistant: RAG-based code chat, guided tours, dependency graphs, and technical-debt analysis, using AI coding assistance throughout. Researchers tracked three cost layers: real AI subscription spend, self-reported human hours, and a counterfactual estimate of what the same tool would cost to build without AI. Their first analysis found AI-assisted development ran 19.4 times cheaper than that human-only counterfactual. A follow-up check found two errors, pricing AI usage per-token despite a flat-rate subscription and using labor rates from the wrong region for the counterfactual, and correcting both dropped the ratio to roughly 9.9x.
The real finding here isn't the multiplier, it's that a factor-of-two error sat undetected in a published cost claim until someone went back and audited it. Both mistakes are the boring kind, a mismatched pricing model and a mismatched regional wage assumption, and neither would show up just by scanning the final ratio.
The 10x-developer stats floating around AI coding tools were probably built the same way, so treat 9.9x as a first draft, not a verdict.