More than 90 percent of executives report no effect of AI use on employment over the past three years, and 89 percent report no impact on labor productivity, according to a National Bureau of Economic Research study published in May. Research teams at the Federal Reserve Bank of Atlanta, the Bank of England, the Deutsche Bundesbank, and Macquarie University fielded identical survey questions between November 2025 and January 2026, yielding responses from nearly 6,000 CEOs, CFOs, and senior finance managers. The finding is not an outlier. A recent survey of more than 4,500 business leaders by consultants at PwC found that more than half reported seeing neither increased revenue nor decreased costs, per The Register. The pattern is consistent across multiple surveys and geographies. Organizations deployed AI. Executives used AI. Ninety percent said it changed nothing measurable.
I run AI agents in production at coenconstruction.com, estimate.pro, and valhalla-k9.com. The invoice validation agent at the construction ERP processes 400 invoices per month and flags 11 percent for human review. The estimating agent at estimate.pro generates scope statements in three minutes that previously required 90 minutes of manual drafting. The SMS scheduling agent at Valhalla sends appointment reminders within a two-hour window six days before the session. All three agents produce measurable time savings because I measured the baseline before deployment and tracked the delta after. The seventy-four percent who deployed AI but cannot prove it worked are not lying. They are measuring deployment instead of integration. The NBER data suggests most of the 89 percent fall into that category.
In the US, UK, Germany, and Australia, roughly 70 percent of firms have adopted AI, but its effects so far on employment and productivity remain small. The gap between 70 percent adoption and 89 percent reporting no impact is the difference between having access to a tool and redesigning a process to depend on it. The invoice agent at coenconstruction.com is integrated, not deployed. The old workflow no longer exists. Bookkeepers do not scan every invoice manually. They review only the flagged subset. If the agent stops working, invoices pile up in the approval queue faster than the bookkeeper can clear them manually. That dependency forces a fix. If the agent were optional and bookkeepers still reviewed every invoice, a broken agent would be invisible. The productivity gain would also be invisible.
The distinction between deployment and integration explains why executives predict sizable effects over the next 3 years, predicting that AI will boost productivity at their firms by an average of 1.4%, according to the NBER working paper. The forecast is not irrational. It reflects the assumption that organizations will eventually redesign workflows to make AI load-bearing. The three-year lag is consistent with the time required to identify bottlenecks, redesign processes, retrain staff, and measure outcomes. The problem is that most organizations skipped the first step. They deployed AI into existing workflows without identifying which tasks were bottlenecks and which tasks were already fast. The result is AI being used for tasks that save seconds, not hours.
The estimating agent at estimate.pro produces a 30 percent time savings because scope generation was a bottleneck. The agent replaced 90 minutes of manual work with three minutes of generation plus 20 minutes of editing. The 70-minute savings is measurable and repeatable. If I had deployed the agent to a task that already took five minutes, the maximum possible savings would be five minutes, and the overhead of reviewing the agent's output might eliminate the gain entirely. The NBER findings suggest most deployments look like the second case. AI is applied to tasks that were not bottlenecks, or applied to workflows where the agent's output still requires the same manual review that existed before.
The adoption problem is not convincing employees to use AI. It is redesigning processes so the AI is load-bearing and its absence would break the workflow. The invoice agent meets that test. The estimating agent meets that test. The SMS agent meets that test. Each agent replaced a step in a workflow that cannot proceed without it. The agents that got defunded after six months were likely the ones that never became load-bearing. The budget holder could not point to a process that would fail without the agent, so the renewal was an easy cut. The 89 percent who reported no productivity impact are running agents that are optional, not load-bearing.
The survey asked 6,000 executives about three years of AI use. Eighty-nine percent said it changed nothing measurable. The problem is not the models. It is three years spent deploying tools into workflows that were never redesigned to require them.
The next three years will determine whether the 1.4 percent productivity forecast materializes or whether the deployment-without-integration pattern continues. The organizations that hit the target will be the ones that spent 2026 identifying bottlenecks and redesigning workflows to assume AI is present, not the ones that spent it adding AI to workflows designed for humans. The NBER survey is a snapshot of what happens when deployment precedes redesign. The lesson is not that AI does not work. It is that deployment is not the same as integration, and only integration produces measurable results. The agents are still running. The question is whether anyone redesigned the process to depend on them.