DORA puts a number on the tuition (external link)
Google Cloud's DORA team published an ROI framework for AI-assisted development in May, and Matt Saunders' write-up is the fastest way into it.
The headline model: a 500-person engineering organisation at $176,000 average salary sees roughly 39% first-year ROI and an eight-month payback — about $11.6m of value against $8.4m of investment. Over three years Google's own data shows an average 727% return.
Treat those figures as what they are, which is a calculator with assumptions you get to set, published by a company that sells the thing being evaluated. Run the conservative scenario. The report to its credit tells you to.
Two things in there are worth more than the ROI number.
The first is the J-curve. Organisations reliably get worse before they get better — learning curve, review overhead, process churn — and DORA calls that dip "the tuition cost of transformation". I have watched three teams go through this and every one of them interpreted the dip as a tooling problem and responded by changing tools, which resets the curve to zero. If you take one thing from the report, take this: the trough is the normal shape, and paying the tuition twice is a choice.
The second is the spread. 35–40% productivity gains on simple tasks; 10% or less on complex legacy code. Nearly all the money in most organisations is in the second category. Any business case built on the first number is a business case for work you were not doing.
The framing underneath all of it — "AI magnifies the strengths of high-performing organisations and the dysfunctions of struggling ones" — is not new, and DORA has been saying versions of it since the DevOps reports. It keeps being true, which is the problem: it means the tool cannot be the intervention, and buying the tool is much easier than being the kind of organisation where it pays.