It is not enough for companies to use AI. It matters to use it strategically.

Many AI deployments appear successful on short-term financial metrics, yet degrade service quality that organizations are not measuring. Klarna's AI customer service initiative initially improved profitability, but the CEO later resumed hiring human agents due to service quality concerns. Air Canada was held liable (Moffatt v. Air Canada) after a customer service chatbot provided incorrect information in an edge-case scenario, prompting the company to scale back customer-facing automation. A recent field experiment on Alibaba's Taobao platform found that AI reduced handling time but also reduced customer satisfaction, and critically, human intervention did not recover perceived service quality unless introduced promptly, early in the interaction.

Gartner predicts that half of companies that cut customer service staff due to AI will rehire by 2027. The proliferation of such cases has led many organizations to conclude that the key AI challenge is no longer adoption, but oversight and quality evaluation.

The common risks of AI deployment

A common risk in new AI deployments is that short-term KPIs miss risks with long-horizon effects: declining customer satisfaction, mishandling of novel cases, and employee rubber-stamping of low-quality outputs build up over time. While deployment reversals most visibly hit customer-facing operations, aggregate surveys, such as MIT Project NANDA and Qualtrics XM Institute's 2026 Consumer Experience Trends Report, report consistently disappointing AI adoption results across the board. According to Qualtrics, $3 trillion in global revenue is at risk from poor customer experience, and 70% of dissatisfied customers leave no feedback. According to MIT's Project NANDA, only 5% of enterprise AI pilots successfully reach production and deliver measurable, sustained productivity or P&L impact.

Why does this keep happening?

Research in forecasting, cognitive science, and machine learning points to a consistent structural explanation. AI systems are trained on historical data and perform best when future situations resemble past ones. Humans gain comparative advantage under novel or changing conditions, for instance, as documented in Tetlock's work on superforecasters.

This reflects a deeper difference in how humans and machines learn: human learning is oriented toward future performance, not accurate recall of the past. The differences between human and algorithmic reasoning are well understood through decades of cognitive science research, and more recently across work comparing humans and LLM-based agents.

These insights can be distilled into a practical framework for strategically integrating human and AI judgment.

Cognitive Routing

Cognitive Routing asks four questions about any workflow under consideration for automation.

  1. What is the scope? Parts of a workflow that are routine and historically stable are the clearest candidates for automation. A useful diagnostic: where does your team currently spend time gathering or aggregating existing data? Tasks that can be executed reliably based on past examples are strong candidates for AI support, and automating them can meaningfully support human decision-makers.
  2. What is the oversight policy? Before a workflow is automated, organizations need mechanisms to detect novel or unusual cases and escalate them to a human. An escalation pathway needs to precede a responsible deployment.
  3. What are the risks? Detecting faulty AI outputs can be tricky because errors are often superficially plausible. Employees who lack technical knowledge, or who are managing high volume work, may revert to rubber-stamping AI-generated outputs. Managing this risk requires maintaining employee skills to prevent degradation, and setting realistic expectations about output volume and review load.
  4. How will quality be monitored? A new AI deployment must be monitored and evaluated using operational KPIs relevant to the job, not vendor benchmarks. Vendor metrics are optimized to reflect what the product does well. Your metrics should reflect what your organization produces.

AI systems will continue improving. They will more readily integrate recent information, handle larger contexts, and respond with increasing accuracy across tasks. But the managerial challenge of deciding where to draw the boundary between automation and human judgment will remain.