We Tried Replacing Employees with AI It Broke

Klarna: The Most Public Reversal

In 2024, Klarna CEO Sebastian Siemiatkowski became one of the most-quoted AI optimists in the industry. His claim: AI was doing the work of 700 employees. The company had stopped hiring 'largely due to AI.' The messaging was confident and the press covered it extensively as a proof point for the AI transformation thesis. Within roughly twelve months, Siemiatkowski had publicly admitted the company 'went too far' and 'focused too much on cost.' Customer satisfaction scores had declined sharply. Users described AI support responses as 'generic, repetitive, and insufficiently nuanced.' Klarna began rehiring humans.The Klarna case is important not because it was unusual but because Siemiatkowski was unusually honest about what happened. Most companies that overdeployed AI and quietly reversed course did not issue public statements. The CEO of Klarna's transparency made visible what is almost certainly a much more widespread phenomenon: organisations that announced AI-driven headcount reductions as strategic decisions and are now managing the customer experience consequences of those decisions without acknowledging them publicly.

McDonald's McHire: The Security Failure

In mid-2025, McDonald's AI-powered hiring platform McHire built on Paradox.ai, since acquired by Workday was found by security researchers to have a test administrator account secured with the default credentials '123456/123456' and no multi-factor authentication. The exposed data was linked to approximately 64,000 applicant records. The exposed information included full chat transcripts from the 'Olivia' hiring chatbot and responses to personality assessment questions.This incident illustrates a consistent pattern in enterprise AI deployment: AI systems are treated as products to be turned on, not infrastructure to be secured. The organisations deploying these systems inherit the security posture of the vendor's implementation defaults which are often consumer-grade rather than enterprise-grade. The McHire breach was not sophisticated. It required no hacking skill. It required knowing that a publicly accessible admin panel existed and that the default credentials had never been changed.

Workday: Algorithmic Discrimination at Scale

A US federal court allowed a nationwide class action to proceed against Workday in May 2025, alleging that its AI-powered resume screening system discriminated against applicants over 40, and against Black and disabled applicants. The lead plaintiff a Black male jobseeker over 40 applied to over 100 jobs through companies using Workday's AI screening and was automatically rejected every time. One rejection arrived at 1:50 AM, less than an hour after he applied. No human could have reviewed the application.The legal theory: Workday's AI baked in existing bias by training on historical hiring data that reflected discriminatory human decisions. The AI did not invent new discrimination. It automated and scaled the discrimination that already existed in the training data. Thousands of companies use Workday's tools to screen candidates with no human in the loop. The SafeRent AI tenant scoring system settled a $2.2 million lawsuit in November 2024 for the same structural failure in housing access decisions. The pattern AI trained on historically biased data, deployed at scale with no human review, producing discriminatory outcomes is repeating across industries.

Block: The 4,000-Person Layoff That Immediately Required Rehiring

Jack Dorsey cut 4,000 employees in February 2026 and attributed the decision entirely to AI capability advances. Within weeks, Block began quietly rehiring some of the same employees. At least four former workers returned to the company. One departure was acknowledged by Block leadership as a 'clerical error.' Another was described by the company as 'operational necessity.' The speed of the reversal in some cases less than two weeks suggests the cut was not as precisely calibrated to AI capability as the public framing implied. Dorsey's own internal memo from March 2025 had explicitly stated that the round of cuts then underway was 'not trying to hit a specific financial target, replacing folks with AI, or changing our headcount cap.' Eleven months later, AI replacement was the stated justification for a cut four times larger.

Air Canada: The Chatbot That Created Legal Liability

Air Canada's AI chatbot told a customer that bereavement fare discounts could be applied retroactively a policy that did not exist. When the customer attempted to claim the discount, Air Canada refused, arguing the chatbot had made an error and the company was not bound by it. A Canadian civil tribunal ruled against Air Canada: the company was responsible for information provided by its own AI system, whether or not a human authorised the specific statement. The ruling established a legal precedent that AI chatbot outputs are the company's outputs. The liability follows the deployment, not the individual erroneous statement.

What the Failures Have in Common

• Every significant AI failure involves a high-stakes, judgment-dependent context hiring decisions, customer service commitments, security of personal data where the cost of individual errors is high and the volume of decisions means errors compound before anyone notices
• The failures occur when AI makes or communicates decisions without a human review step before consequences are delivered to the affected person
• Training data that encodes past discrimination reproduces that discrimination at scale not because of malicious intent, but because historical data reflects historical bias that the model has no mechanism to correct
• Security treated as an afterthought produces breaches that require no sophistication to execute default credentials, missing MFA, no access reviews
• The timeline from pilot to production affecting real people compresses in AI deployments in ways that do not happen with traditional software, and governance structures have not kept pace with that compression
• Companies that move fast on AI replacement and slow on AI governance produce the specific combination of conditions high deployment scale, low oversight, high error cost that makes failures severe rather than contained

The Lesson the Cases Teach

MIT's 2025 review of over 300 AI implementations found that the biggest failures were not technical. They were organisational: weak controls, unclear ownership, and misplaced trust in AI systems that had not been stress-tested in production conditions. The companies that have made AI replacement work are the ones that defined a specific problem, assigned a named owner, invested in data quality, maintained human review for high-consequence decisions, and measured outcomes against the baseline. The companies that failed are the ones that deployed first and governed later.Given that AI systems can make thousands of decisions per minute, governing later means thousands of errors before the problem is detected. Klarna's customer satisfaction data degraded across hundreds of thousands of support interactions before the reversal decision was made. Workday's hiring AI screened out hundreds of qualified candidates before the class action was filed. The scale and speed at which AI systems operate is precisely what makes ungoverned deployment so damaging and precisely what makes the governance investment non-optional rather than nice-to-have.