AI tools are embedded across supply chain workflows, but human review still carries most of the burden. New Connext Global survey data shows widespread rework, uneven accuracy and rising customer risk, raising questions about how much productivity these systems actually deliver.
Adoption Runs Ahead of Confidence In Autonomous AI
Recent Connext Global research highlights a clear gap between widespread AI use and trust in fully automated decisions. When respondents defined what reliable AI means in practice, only 17% endorsed systems that operate without human involvement. A decisive majority favored a hybrid model, with 70% describing reliability as AI coupled with some level of human review, split evenly between light-touch checks and dedicated oversight.
That dependence on human intervention shows up in how often AI needs attention. Twenty-eight percent said tools require involvement almost every time, while 54% reported that issues arise at least sometimes. Only 4% felt systems can usually run with minimal supervision, a signal that current deployments function as assisted automation rather than autonomous operations.
Follow-up work has become a standard step in the process. Just 4% of respondents said they rarely perform additional tasks after AI runs. Most described a consistent pattern of editing, fixing and formal review. Forty-two percent cited correction and clean-up as their most common post-AI activity, and 34% pointed to review and approval as a regular requirement.
Perceived accuracy remains uneven. Only 37% reported that AI is right without fixes most of the time. Nearly two in three placed performance in the sometimes or less category, including 45% who said it is accurate only sometimes, 16% who said rarely and 2% who said almost never. For organizations embedding AI into planning, procurement analytics or customer communication, this level of variability turns oversight from a temporary safeguard into a permanent operating feature.
Time savings are also less predictable than adoption narratives suggest. Among those who needed to correct AI output, 46% said fixes take about the same time as doing the task manually and 11% said they take longer. In effect, 57% indicated that once rework is required, any speed advantage can disappear. That reality forces closer scrutiny of where AI sits in the process stack, which activities can tolerate rework, and how to calculate ROI beyond headline productivity claims.
Context Gaps, Downstream Risk and Customer Exposure
The survey identifies missing context as the dominant source of AI failure. Forty-two percent of respondents said systems left out important details or misunderstood key elements. In complex supply networks, where decisions depend on product constraints, local regulations, service commitments and partner capabilities, context gaps can distort demand signals, allocation rules or transport plans.
Rework is another recurring theme. Thirty-two percent reported that AI regularly creates extra work to fix or redo tasks, highlighting how incomplete data or weak prompts can amplify effort instead of reducing it. A further 31% pointed to cases where AI responses sounded authoritative but were wrong. That pattern is especially risky in exception-driven environments, where teams often rely on automated recommendations to manage volume.
Customer-facing consequences are already evident. About 19% of respondents said AI made at least one customer situation worse, shifting the issue from internal efficiency to external experience. AI-generated content and analytics increasingly shape service updates, order responses, pricing decisions and risk notifications. When outputs miss nuance or context, the impact reaches customers, partners and revenue streams.
Personal exposure to negative outcomes is widespread. Sixty percent said they have been directly involved in an instance where AI harmed results. Within that group, 18% reported customer frustration or complaints and 11% linked AI use to lost revenue or churn. These responses show why governance has moved beyond a technical concern toward an operating requirement that spans data quality, workflow design and customer protection.
Industry reports on digital operations echo this pattern: value emerges fastest where organizations define explicit guardrails, escalation paths and task-specific controls, rather than relying on generic policies. Teams that treat AI output as draft work that always needs checking are avoiding some of the sharpest failures, but they are also carrying a hidden layer of labor that traditional business cases rarely quantify.
Oversight Will Define Where AI Actually Pays Off
The next phase of AI adoption will be shaped less by model capability and more by how precisely oversight is designed into workflows. Teams that treat verification as measurable work, tracking rework rates, correction time and customer exposure, will have a clearer view of where AI delivers net value and where it adds friction. That visibility allows organizations to narrow deployment to tasks where iteration is acceptable and tighten controls where errors carry cost or reputational risk. Over time, this discipline will influence vendor selection, process design and even how performance is measured across planning, procurement and logistics.
