Every team is being told to use more AI, and the bill is climbing — tokens, seats, models. Whether it pays off isn't a yes or no; it's a discipline. This is the evidence, drawn from working systems I built and independently audited, and what it means for how we adopt AI safely and cost-effectively.
The cheapest model out-reviewed the dearest. Default to a cheap model; pay the premium only where a wrong answer is genuinely expensive. The 143× spread is a cost dial every team controls.
It optimises the number you give it, not the goal you meant. Two of my systems had quietly learned to game their own quality scores. Spend only becomes value when you verify.
Verification built into the process — the AI checking the AI (six reviewers, then three skeptics), and self-heal to a green test suite — is what makes the output trustworthy. It is maker–checker–sign-off, applied to AI.
The scarce skill is no longer building; it's judging. Engineers, QA, BAs, risk and domain experts all stay essential — as the people who direct the work and know when not to trust it.
Route by stakes — cheap models by default, the frontier only for regulatory-critical work. Set per-team token budgets and track cost per successful outcome, not cost per call.
Never trust a green check. Prove it ran, slice the distribution (averages hide tail risk), and check the checkers. Make failure loud.
Run a 30-day pilot on one repetitive task per team; log every failure. In a month you have real fluency and a defensible answer to "is it paying off?"
Costs in AUD ≈ USD list price × 1.42 (MiMo US$0.36 · Gemini US$0.66 · Fable US$51.41).