Artificial intelligence has delivered dramatic gains across sectors, but for every success story there are cautionary tales. In 2025, 78% of organisations globally already use AI, and 85% have adopted agentic solutions in at least one workflow. Despite this widespread adoption, poorly executed projects continue to remind us that the quality of data, the clarity of strategy and the respect for ethics determine whether AI delivers value or becomes a costly misadventure. Below are three of the most notable failures – and the lessons they teach.

1. DoNotPay's "Robot Lawyer" – Over‑promising and under‑delivering

DoNotPay promoted its AI service as the "world's first robot lawyer", promising to draft legal letters and help consumers fight fines. However, the service struggled to perform simple legal tasks, prompting the Federal Trade Commission (FTC) to accuse the company of misrepresenting its AI's capabilities. DoNotPay agreed to pay $193,000 in refunds, clearly disclose that its service is not a lawyer and stop claiming that its AI can replace professional legal services.

Reasons for failure: The system attempted to automate complex legal reasoning without adequate data or domain expertise. Instead of working with qualified lawyers to refine the model, the company relied on generic templates and unverified claims. The FTC's action underscores that ethical transparency and realistic marketing are essential when offering AI products, especially in regulated industries.

2. Amazon's AI Recruiting Tool – Biased training data leads to discrimination

Amazon's experiment with AI‑driven recruitment ended abruptly when the company discovered that its machine‑learning model discriminated against female job applicants. The tool had been trained on ten years of hiring data dominated by male resumes. As a result, it penalised CVs containing words such as "women's" or names of women's colleges. Researchers later emphasised that the problem wasn't the algorithm itself but the biased training data and the human selection criteria it tried to emulate.

Reasons for failure: The project lacked a robust data governance strategy and fairness testing. By learning from past hiring decisions, the model replicated existing human bias. Amazon's decision to abandon the tool shows that AI systems must be audited for bias and fairness before deployment, and that diverse teams are needed to evaluate training data.

3. IBM Watson for Oncology – Misaligned expectations and contextual mismatch

IBM's Watson for Oncology was launched with grand promises: it would synthesise medical literature, patient data and clinical guidelines to recommend cancer treatments. However, trial deployments revealed that the AI's recommendations were often inconsistent with local clinical practices – the system relied heavily on U.S. guidelines that did not align with treatment standards or drug availability in countries such as India and China. By 2018, reports documented that Watson sometimes provided inappropriate or unsafe treatment suggestions and skepticism grew. Coupled with declining revenues for IBM's health unit, these challenges led to the project's discontinuation in 2023.

Reasons for failure: Watson's developers underestimated the importance of context‑specific data and continuous testing. They trained the model on datasets and guidelines from a limited set of institutions without accounting for global variations in oncology practice. Furthermore, the project emphasised marketing hype over incremental clinical validation, creating unrealistic expectations. This case illustrates that medical AI must be grounded in diverse, high‑quality data and undergo rigorous, peer‑reviewed evaluation.

Lessons learned

These failures share common themes:

Garbage in, garbage out: Biased, incomplete or irrelevant data will produce flawed AI outputs. Rigorous data governance and bias mitigation are non‑negotiable.

Strategy before technology: AI should solve a clearly defined problem and align with organisational goals. Without a business case and user‑centred design, even powerful models can fail.

Ethics and oversight: Transparent marketing, accountability and human supervision are essential. Regulators increasingly demand that AI products be honest about their limitations, respect privacy and avoid discrimination.

Context matters: Domain expertise and regional variations must shape model development. A "one size fits all" approach to AI rarely works.