Ninety-three percent of technology leaders believe AI will eventually run their businesses autonomously. Seven percent have made it work. That gap — between belief and proof, between deployment and measurement — is the real story of the current AI cycle.
A vendor-commissioned survey conducted by Wakefield Research for Teradata polled 1,000 VP-level leaders across six markets in early 2026. The findings are stark: while 93% believe AI will run core business functions autonomously, only 7% have reached the operationalizing stage. The ambition to deploy is outpacing the ability to verify results. Sixty-three percent have seen no more than small or emerging positive returns. Forty percent report that more than 40% of their AI pilots never reach production. The barriers are foundational: 77% say 20% or less of their enterprise data is ready for agents, while 60% report decision paralysis on infrastructure.
This is a governance problem, not a technical one. Companies are deploying agents without the data infrastructure to measure them. As Louis Landry, CTO at Teradata, puts it:
Individual productivity gains — faster code, better drafts, quicker research — are real benefits, but they don’t show up on the P&L in a way that justifies significant infrastructure investment.
Josh Fecteau, Chief Data and AI Officer & CIO at Teradata, goes further:
The goal of contextualizing your entire data estate is likely the wrong goalpost, and chasing it is part of why organizations stall.
The disclosure gap makes it harder to learn from what is working. A Brief Script audit of 150 qualifying AI agent deployments from 119 organizations found that only 22 disclose any cost or financial impact. One hundred eighteen describe explicit human control; 57 describe bounded autonomous action. When even the largest technology providers fail to share the numbers, the problem becomes systemic. Microsoft’s own internal HR deployment — three agents handling Learning, Workforce Insights, and People — includes a section that begins “Speaking about ROI” and carries zero figures. A vendor grading its own homework without showing the scores.
The human cost is hiding in the review queue. A Harvard working paper by Chen and Stratton found that AI coding agents boosted code volume by 30% — adding 1,494 lines per worker-month — but issues resolved stayed flat. Review times jumped 49%. Revision requests nearly doubled. Fourteen percent more workers were pulled into the code review process. The bottleneck is not technical capability. It is the review infrastructure required to manage the output.
Workers are bearing the cost of the measurement gap. They are adapting to new tools without proven workflows, often doing more work to manage AI output rather than seeing their own workloads decrease. This “review tax” is a direct consequence of deploying agents before the governance and measurement frameworks are in place.
The people carrying this burden are the ones least likely to see the measurement data. They are adapting to AI without knowing if the adaptation is worth it. Until companies can show the numbers, the deployment curve is just a bet — and the workers are the ones covering it.