It's a common scenario: an SME, whether with an in-house development team or an external consultant, faces the task of implementing AI solutions. The challenge isn't whether to use AI, but which model to choose, how to optimize it, and what the true cost will be. In the projects we manage, 80% of unbudgeted AI project costs don't stem from insurmountable technical issues, but from ineffective resource management: the wrong model for the task, inefficient prompt engineering, or overly verbose output.
Consider the case of a manufacturing company with around 120 employees, specializing in precision mechanical components. Their goal was to automate the initial screening of quotation requests, distinguishing feasible ones from those requiring immediate human intervention. Initially, the team chose the most powerful model available, convinced that 'bigger is better.' The result: highly accurate responses, but high latency times and tokenization costs that eroded the weekly budget, making the entire solution unscalable for their request volume.
Optimizing Claude Models: The Pragmatic Choice
Not every problem demands the brute force of a 'top-tier' model. Anthropic, with its Claude models (such as Opus, Sonnet, and Haiku, evolving from the AI, Code, and Fable models mentioned in the brief, which continue to follow the same logic of specialization), offers a range of options designed for different balances of capability, speed, and cost. The real intelligence lies in knowing how to select the right one for a company's specific problem.
- Claude Opus: Ideal for complex reasoning tasks, structured data analysis, critical code generation, or deep comprehension of legal or technical documents. It's the most 'capable' model, but also the one with the highest cost per token and slightly longer response times. We use it when accuracy and depth of analysis are a must and the volume of interactions is manageable. As we discussed in a previous article, Claude Opus 5: The Next Leap for AI Reasoning in SMEs, its reasoning capabilities can unlock previously unthinkable scenarios.
- Claude Sonnet: An excellent balance between performance and cost. Perfect for most daily business tasks: email summarization, draft text generation, customer service assistance, categorization. It offers fast and sufficiently accurate responses without burdening the budget. It's often the starting point in our projects.
- Claude Haiku: The fastest and most economical model, ideal for high-volume, low-complexity activities where speed is crucial and a minimal error margin is acceptable. Think basic FAQ chatbots, rapid entity extraction, or preliminary data filtering. Often, an internal team will use it to quickly test an idea or prototype a solution, reducing the initial investment.
For the aforementioned manufacturing SME, the solution was simple yet effective: switching from the Opus model to Sonnet for the initial categorization of quotes. The result? A reduction in cost per request of over 60% and halved response times, with no noticeable drop in the accuracy of preliminary categorization. More complex quotes requiring in-depth analysis were then redirected (guided by Sonnet) to the human team for finalization.
Controlling Verbosity and Optimizing Costs
Another critical, often underestimated, element is output verbosity. The more words the AI generates, the more tokens it consumes, and the higher the costs. At Logika.studio, we've observed that managing verbosity can reduce inference costs by 20-30% in high-volume projects.
- Clear and Concise Instructions: Always specify the desired output length (e.g., 'respond in a maximum of 50 words', 'provide only the 3 main points').
- Structured Format: Requesting output in specific formats like JSON or Markdown (even with
claude.md) helps limit rambling and obtain precisely the necessary information. Theclaude.mdformat is particularly useful for engineering complex prompts, allowing sections and output constraints to be defined clearly and comprehensibly to the model. - Prompt Iteration and Fine-tuning: This is a continuous process. Testing different prompts for the same task helps find the formulation that returns the desired output with the fewest tokens.
For example, in a logistics and transport startup, we were tasked with daily summarizing hundreds of shipping reports into a few key points. Initially, the model generated summaries of 200-300 words. We implemented a prompt system that specified the JSON format for key fields and limited the 'summary' field to a maximum of 50 words. The impact on monthly costs was clear, with an estimated saving of about 300 euros/month on a volume of 500 reports daily.
Governance, Privacy, and Code Ownership
When discussing AI integration, especially for SMEs, governance and privacy issues arise. Using models like Claude via API is generally secure, but understanding how data is processed is crucial. Anthropic's policies are clear on not training models with customer data, a critical aspect for intellectual property protection and GDPR compliance.
At Logika.studio, we adopt an approach that aims to guarantee maximum code ownership for the client. Regardless of the cloud provider (or even in on-premise environments), the client retains total control over their infrastructure and data. Our 100% human review at every project stage, from design to implementation, ensures not only technical quality but also alignment with the company's strategic and regulatory needs.
Integrating an AI model into a business process doesn't have to be a leap of faith. With the right strategy and a careful selection of tools, the benefits in terms of time and resources saved can be significant, often quantifiable in weeks of automated work or thousands of euros saved each month. Implementing solutions like those described typically takes 2-4 weeks, including tuning and testing phases.
If you want to delve deeper into a similar case, a free 15-minute audit is available at audit — quick analysis, 2-3 concrete points, zero pitch.
