AI's return on investment is a property of the delivery system

DORA's 2026 report on the ROI of AI-assisted development does something vendor decks avoid: it puts numbers on both sides of the ledger. In its worked model of a 500-engineer organization, roughly $11.6 million of first-year value stands against $8.4 million of cost, and one line runs negative. A change failure rate drifting from five percent to six takes back about $344,000 by itself. The report calls this an instability tax.

The framing we find most useful is the report's own: AI is an amplifier. It raises the volume on whatever the delivery system already does, including the parts nobody staffed. Faster coding pushes more change into the same review queue, the same environments, the same release process, and the delivery slowdown we wrote about is what amplification sounds like when the queue was already the constraint.

So the diagnostic question is not "what would the assistant save." It is "what does our pipeline do when change volume rises thirty percent." Your own data answers that one: review latency, queue depth and failure rate, by team, in trend. Price the tool second. The system decides the return, and the system is measurable now, before the invoice arrives.