Cutting Costs with AI Agents Is Impossible... Wait, Actually
OpenAI
Rasa
Amazon Web Services
Paradox.ai
A review of several real-world cases where AI implementation reduced operational costs, including failures and lessons. Examples: Klarna's AI support system backfired, Swisscom's complex agent architecture succeeded, OneDigital's HR AI savings were miscalculated, Duolingo's AI-first strategy had mixed results. Key takeaways: start with simple use cases, measure qualitative outcomes, count real savings.
The article discusses real examples of AI reducing operational costs, highlighting both successes and pitfalls. Klarna replaced over 700 support operators with OpenAI AI, cutting response time by a third and resolving issues 5 times faster, but customer churn rose 15% and losses grew, forcing a return to human agents and later a hybrid model. Swisscom successfully deployed a multi-agent architecture using Rasa CALM and Amazon Bedrock AgentCore, doubling automated resolutions and cutting opex by 50%, though the setup is complex and costly. OneDigital saved up to $170,000 annually on HR scheduling via paradox.ai, but critics note the savings are overstated because employees remained employed and the ATS costs $30k-$100k per year. Duolingo adopted an AI-first approach, boosting revenue to $292 million in Q1 2026, yet admitted at least 20% of AI-generated content was unusable, and internal metrics misleadingly rewarded AI usage. Recommendations include starting with basic cases, evaluating qualitative metrics, counting real savings, and using human-in-the-loop review.
- Abbreviations
- HR = Human Resources
- ATS = Applicant Tracking System
- RAG = Retrieval-Augmented Generation
Source: Habr — хаб ИИ —
original