It's a common scenario: an entrepreneur or CTO at a 70-employee manufacturing SME decides to invest in artificial intelligence to optimize customer correspondence management. After evaluating several options, they choose a model like Anthropic Claude, valued for its ability to handle complex conversations and generate coherent responses. The team successfully implements a prototype that classifies incoming emails, extracts key information, and drafts responses for customer service, reducing request handling time by 30%. It seems like a triumph. Then, in the subsequent months of 2026, something strange begins to emerge: the model's responses become less precise, more generic, sometimes irrelevant. Misclassified emails increase. The need for human review, initially reduced, starts to climb again. The issue isn't outright malfunction, but a subtle yet persistent decline in quality, coupled with rising API costs, because the model 'thinks' less and requires more calls or longer prompts to achieve the same results as before.
This archetypal scenario is not isolated. In the projects we oversee, a recurring pattern we've observed recently is the performance degradation of large language models (LLMs) from certain providers, particularly concerning Anthropic Claude's Fable 5 and Opus 5 versions. Recent analyses and direct user feedback, including from our own teams, indicate a significant drop in reasoning capabilities and the cognitive 'budget' allocated by the model to solve tasks, without clear communication from the vendor. This lack of transparency raises important questions, especially for SMEs investing time and resources in integrating these technologies.
The Damage Beyond Performance: Costs and Reliability
Performance degradation isn't just an output quality issue; it has direct and measurable repercussions on operational costs and the reliability of AI-powered systems. When an LLM becomes less effective, several things happen:
- Increased API Costs: If the model requires more complex prompts or more iterations to arrive at an acceptable answer, the number of tokens processed, and consequently, the cost per call, increases. A company that previously spent 500 euros a month on APIs might find itself paying 700 or more, not due to increased workload, but merely to compensate for the model's diminished efficiency.
- Increased Human Review: The promise of AI is to reduce manual workload. If output quality declines, the time employees spend correcting, verifying, or even rewriting AI-generated responses starts to climb again. This not only negates efficiency benefits but also generates frustration and internal distrust in the AI project. To delve deeper into how commercial LLMs can generate unexpected costs and complexity, we covered the topic in Commercial LLMs for SMEs: When Anthropic Claude Generates Unexpected Costs and Complexity.
- Loss of Trust: Reliability is paramount. An AI system that performs unpredictably erodes the trust of decision-makers and end-users. Projects risk stalling or being abandoned, leading to wasted investments and time. For an SME, every dollar and every hour counts. An AI project that loses reliability represents an incalculable risk.
- Management Complexity: Managing these resources becomes crucial. Robust monitoring systems are needed to track not only usage and costs but also output quality metrics, comparing them over time.
Concrete Strategies to Mitigate Risk
Facing scenarios like the one just described, SMEs shouldn't abandon AI but rather adopt a more resilient and strategic approach. At Logika.studio, we've defined some guidelines to address these challenges:
- Vendor Diversification: Avoid relying on a single LLM provider. Consider integrating multiple models (e.g., Claude for some tasks and GPT or open-source models for others). This strategy reduces dependency and offers greater flexibility in case of degradation or policy changes from a single vendor.
- Continuous Performance and Cost Monitoring: Implement dashboards and alerting systems that track model performance (e.g., accuracy, relevance) and API costs in real time. Tools like n8n or other orchestrators can help detect anomalies and trigger contingency plans.
- Hybrid Approach with Local Models: For specific tasks or sensitive data, consider using smaller, optimized models that can run locally or on private clouds. This not only offers greater control over performance but also enhances privacy and reduces long-term costs. We explored this solution in Hybrid AI Agents: Qwen Code Locally for Optimized Costs in SMEs.
- Prompt Optimization and Engineering: Invest in 'prompt engineering' to maximize model effectiveness with the fewest possible tokens. This includes techniques like context compression, using specific examples (few-shot learning), and structuring responses to facilitate automatic information extraction.
- 100% Human Review (where crucial): Always maintain a component of human review, especially for critical outputs. AI is a tool to increase efficiency, not to replace human judgment. Our approach always includes human review to ensure quality and prevent costly errors.
Transparency from LLM providers is a cornerstone for building trust and stability in the sector. For SMEs, awareness of these dynamics and the adoption of proactive strategies are essential to transform current challenges into opportunities for growth and innovation. AI is a powerful tool, but like any tool, it requires maintenance, monitoring, and a clear strategy to unleash its full potential, without falling into the traps of unexpected costs or loss of reliability.
If you want to explore a similar case, a free 15-minute audit is available at audit — quick analysis, 2-3 concrete points, zero pitch.
