27% of AI Coding Work Wouldn't Have Happened Without Agents
Anthropic's 2026 report measures the part of agent value everyone misprices: not faster work, but work that was never viable before. The ROI question you're asking is the wrong one.
NeuroX AI · June 22, 2026
Anthropic's 2026 Agentic Coding Trends Report buries the most important number in the productivity section. Developers now use AI in roughly 60% of their work — but about 27% of that AI-assisted work is tasks that wouldn't have been done at all otherwise. Not done faster. Done, period, where before they were quietly dropped.
That reframes the ROI math most teams are running. The internal Anthropic finding is sharper still: engineers report a net decrease in time per task but a much larger net increase in output volume. Agents aren't compressing the same backlog — they're expanding what counts as a feasible backlog. The dashboards nobody had time for, the papercut fixes that always got deprioritized, the exploratory spike that wasn't cost-effective by hand — that's the 27%.
TELUS is the concrete version: teams shipped over 13,000 custom AI solutions and engineering code 30% faster, saving 500,000+ hours at roughly 40 minutes per interaction. The headline is the speed; the story is the 13,000 things that didn't exist before.
If you're measuring agents only on "how much faster is the work we already planned," you're pricing them like a faster compiler and missing most of the return. The win is the work you couldn't justify yesterday — which is exactly the work that has to survive production.