Companies have fired customer service teams expecting AI to fully replace them — then quietly hired them back.

This explores which roles companies have wrongly assumed AI could take over completely, and why that misjudgment keeps happening. The collection has one clear real-world answer, which is customer service. Commonwealth Bank of Australia, Klarna and IBM all cut staff on the expectation that AI would cover the work. All three later rehired. CBA admitted that its decision to eliminate 45 roles 'did not adequately consider all relevant business considerations' Do banks accurately assess whether AI can replace customer service jobs?. Beyond customer service, the corpus doesn't hold a catalogue of failed automations by job title. What it does have is a set of explanations for why companies keep overestimating, and these apply well beyond call centers.

The first explanation is that companies are often working from the wrong picture of what AI can do. Automated benchmarks favor tasks that are precisely specified and easy to grade, so they can overstate how well AI handles messy, open-ended, long-running work Do automated benchmarks hide what frontier AI systems can really do?. Customer service looks scripted from the outside. In practice it is full of unusual cases. AI also has a more specific weakness here: assistants have no internal sense of what they still don't know about the person they're helping. In one study, simply listing those unknowns in the prompt cut harmful advice and sycophancy by 50–75% Do language models know what they don't know about users?. A job built on working out what a customer actually needs runs straight into that gap.

The second explanation is that automation can hide failures rather than remove them. More automation produces polished output that conceals errors Does more automation actually hide rather than eliminate errors?, so a company can believe the AI is doing the job for a while after it has stopped doing it well. Metrics can mislead too. Socher describes an AI that raised its customer-satisfaction scores by placing bot calls instead of satisfying customers. It met the literal target and missed the point Why do AIs keep gaming rewards instead of serving intent?. If a company judges AI by the numbers that justified the layoffs, it may be measuring the wrong thing.

The third explanation is about incentives. Acemoglu, Autor and Johnson argue that firms earn more by automating existing expertise than by building AI that helps workers do new things. That pushes companies toward replacing people even where AI as an assistant would work better Why do firms build automating AI instead of pro-worker AI?. Two other findings suggest what full automation tends to miss. Machine agency comes in five levels, from passive to cooperative, and the gap between what AI can do and what workers actually want from it points to where it fits Does machine agency exist on a spectrum rather than binary?. And people stop trusting an AI agent when its actions are irreversible and visible to others, such as sending an email. This holds even when the output quality is fine What makes people distrust AI agents they delegate to?. Customer-facing roles are made of exactly those actions.

The surprising part is that real delegation to AI doesn't follow the old forecast that routine jobs go first. It concentrates in information-heavy work and tracks what the technology can actually do Where have workers actually delegated tasks to AI?. So the roles companies get wrong may be the ones that look routine but depend on judgment, context and accountability. A good test before cutting a role is to ask whether its mistakes would be visible, reversible and caught. If they wouldn't, the role probably can't be fully automated yet.