Yesterday the argument was about where the human still touches the output. Today the signals move the question one layer back: can the firm see what the agent did at all. The LaunchDarkly survey gives the number that matters. Of 306 practitioners running agents in production, 68% keep runs to ten steps or fewer before a human intervenes and 74% rely primarily on human evaluation (LaunchDarkly). Governance registered at 3%. That is not neglect. It is teams spending their attention on reliability, because reliability is what breaks first and governance has not yet broken anything they can see.
Gartner's forecast is the bill for that choice arriving early. More than 40% of agentic projects canceled by end of 2027, on cost overruns, unclear value, and weak risk controls (Pixelmatters). Read those three together and they collapse into one. Unclear value is a measurement failure. A firm that automates a decision without instrumenting the decision surface cannot say what the agent produced, so it cannot defend the spend when finance asks. The ROI Impact score fell two points in the latest 2026-W46 index for the same structural reason: organizations are scaling ahead of the instrumentation that would justify the scale.
What moved since yesterday is that the vendors started shipping the instrumentation spec. OpenAI's Codex sandbox documentation is the first public specification of where the non-human action boundary sits, filesystem and network limits enforced at the runtime layer. Every command Codex runs carries the credential context of the calling identity. The frontier risk program does the same at the model layer, turning the brand promise of responsible deployment into citable architecture claims. These are gifts to the vendor risk team. Any gap between the published control and the observed behavior is now a finding you can write down.
So the move for a readiness audience is narrow and specific. Pull the sandbox and risk specs into your vendor framework and map them against your own privileged-access controls before a single Codex run touches a repo. Then look at Sumitomo Life, where executive use pulled broader adoption ahead of an IT mandate (OpenAI). Executive-first diffusion in regulated finance means the firm's judgment is already running on AI-assisted rails while the identity control surface lags. That is the 40% cancellation cohort forming in advance, visible only to whoever is measuring.
Watch item: the MIT figure Relevance AI is citing, that 95% of enterprise generative AI pilots show no measurable P&L impact. If the original study surfaces with methodology attached, it sets the baseline against which the Gartner cancellations should be read. Track whether the number holds up to its sourcing.
Index Reference · Applied AI Index 2026-W46, the latest published issue (Sunday, November 15, 2026)
WatchRelevance AI cites MIT research finding that 95% of enterprise generative AI pilots show no measurable P&L impact, positioning structured agent workflows as the path to outcomes. The MIT figure, if replicable, sets a baseline for how the industry should interpret the Gartner cancellation forecast. Worth tracking whether the original MIT study surfaces with methodology attached.