We help Italian SMEs design and implement their Artificial Intelligence strategy: proprietary AI systems built on open source or commercial LLMs, operational AI agents, and team training. We work with one core principle — AI does not replace people, it empowers them.
Artificial Intelligence is not a replacement for human labor: it is a tool that allows people to focus on what they do best. Deciding. Relating. Creating. Solving complex problems. Everything else — summarizing documents, answering repetitive questions, retrieving information, compiling reports — can be delegated to a well-designed AI system.
It is a profound paradigm shift, but it is not a revolution that erases roles. It is an evolution that redraws tasks. Those who have been with the company for twenty years and know the field will know more, not less, after the introduction of AI: they will have more time to think and more powerful tools to act.
This is the way we work. No promises of "AI managing the entire company for you." Only strategy, implementation, and an intelligent redistribution of work between people and machines.
Assessment of the company's AI maturity, mapping of candidate processes for AI augmentation, use cases prioritizing for impact and feasibility, and building a 12-24 month AI roadmap.
Technical architecture of the proprietary AI system, choice between open source (Llama, Mistral) or commercial LLMs (Claude, GPT, Gemini), implementation of operational AI agents, and integration with existing company systems.
Corporate AI governance, compliance with the European AI Act, change management, technical and cultural team training, and defining internal policies on responsible AI usage.
There is no single answer for everyone. We help the company choose the right model based on the use case, data sensitivity, budget, and internal capabilities.
| Dimension | Open source LLMs (Llama, Mistral, Qwen) | Commercial LLMs (Claude, GPT, Gemini) |
|---|---|---|
| Data sovereignty | Total: the model runs on your servers (private cloud or on-premise) | Data passes through the provider's servers (with enterprise options for isolation) |
| Quality on complex tasks | Good, rapidly improving (Llama 3.3 70B is competitive) | Superior for advanced reasoning, long-form generation, multilingual |
| Volume costs | Lower at scale (fixed infrastructure costs, no per-token fees) | Variable per token: cheap at low volume, expensive at large scale |
| Time-to-value | 4-8 weeks (requires infrastructure setup) | 1-2 weeks (API key setup) |
| Vendor lock-in | None: the model is downloadable and portable | Present: changing provider requires prompt re-engineering |
| EU Compliance | Simpler: data remains within the company | Possible with dedicated setups (Azure OpenAI EU, Vertex AI EU) |
| When to choose | High volumes, sensitive data, regulated sectors (health, legal, finance) | High-value use cases with low frequency, critical quality, rapid prototyping |
In practice, the most effective choice is often **hybrid**: open source LLMs for high-volume tasks on confidential data, commercial LLMs for high-value, lower-frequency tasks where model quality makes the difference. Designing this hybrid architecture is a core part of our service.
An AI agent is a system that combines a language model with the ability to use tools (databases, APIs, business software) to complete tasks autonomously. It is the evolution of the chatbot: it does not just answer, it acts.
Weeks 1-4. Analysis of business processes, assessment of AI maturity, mapping of available data, and interviewing 5-10 key figures. Output: assessment document with action priorities sorted by impact.
Weeks 5-8. Strategic workshop with top management. Definition of the 2-3 priority use cases, business case with expected ROI, 12-24 month AI Roadmap, choice of technical architecture (open source / commercial / hybrid).
Weeks 9-20. End-to-end implementation of the first use case. Development, integration, testing with real users, iteration. Goal: bring a working and measurable system to production.
Months 6-18. Extension of AI to other processes, progressive team training, definition of governance policies, continuous model monitoring, and AI Act compliance check.
Virtual assistants are becoming autonomous agents. See how to manage them in our special on AI Agents Integration: news presented at Google I/O.
The methodological framework we use is inspired by the best of Italian managerial literature on AI, integrated with operational experience on the field. Among the main sources we refer to:
A proprietary AI is an artificial intelligence system tailored for a company: trained on its data, integrated into its processes, and governed according to its policies. It can be built starting from open source LLMs (Llama, Mistral, Qwen) installed on servers controlled by the company — private cloud or on-premise — or by using commercial models (Claude, GPT, Gemini) via APIs but with isolated company data. The opposite is \"using ChatGPT.com for work stuff\", which is not an AI strategy: it is an individual tool that exposes data and processes without control.
It depends on the use case. Open source (Llama 3.3, Mistral, Qwen) offers total data sovereignty, no vendor lock-in, and lower operational costs at scale, but requires infrastructure and internal expertise. Commercial (Claude, GPT-4, Gemini) offer superior quality on complex tasks, rapid time-to-value, and immediate scalability, but with variable costs and vendor dependence. The right choice is often hybrid: open source for high-volume tasks on sensitive data, commercial for high-value, lower-frequency tasks.
An AI agent is an autonomous system that combines a language model with the ability to use tools (databases, APIs, corporate software) to complete complex tasks without continuous human intervention. Practical examples for SMEs: an agent that handles first-level customer service inquiries, an agent that retrieves information from corporate document archives, or an agent that qualifies sales leads by reading emails and CRM data. They are the natural successor to chatbots, but with real action capabilities — not just conversation.
No, and no well-designed AI implementation sets this goal. The correct approach is augmentation: AI manages low-value, repetitive tasks (document compilation, information search, customer first-response, report summaries) freeing people to focus on high-value tasks (decisions, relationships, creativity, complex problem solving). The result is not less human work: it is more qualified and less frustrating human work. Almost all companies that tried to \"replace\" customer care with pure chatbots in the years 2020-2022 have backtracked: the standard today is AI as a first tier, human on complex cases.
The initial assessment requires 4-6 weeks. The first operational pilot on a specific use case typically arrives within 8-12 weeks from starting. Scalability across multiple business processes happens over the following 6-18 months. Beware of anyone promising complete AI implementations in 30 days: it is almost always a pre-packaged solution disguised as a custom project, with inevitable vendor lock-in.
A strategic assessment costs 4,000-8,000 euros and lasts 4-6 weeks. A pilot implementation on a single use case costs 15,000-40,000 euros depending on complexity. Scaling projects across multiple processes range from 50,000 to 200,000+ euros. For Italian SMEs, we recommend starting with an assessment + a targeted pilot (totaling 20,000-50,000 euros in the first 4 months) before committing to larger budgets. This is the \"validate before scale\" approach that reduces the risk of investing in technologies not suited to the specific context.
The EU AI Act (EU Regulation 2024/1689) classifies AI systems into 4 risk levels: unacceptable (prohibited), high (mandatory audit and compliance), limited (transparency), minimal (free use). Most AI applications for Italian SMEs fall under limited or minimal risk, which primarily requires transparency towards users (declaring when they interact with an AI) and technical documentation. High-risk systems — such as recruitment, credit scoring, or critical infrastructure — require formal conformity assessments and periodic audits. We integrate compliance into the design phase, not as a later patch.
Yes, if the architecture is designed correctly. With self-hosted open source LLMs (in a private cloud like AWS, Azure, or on-premise), data never leaves the company perimeter. With commercial LLMs, there are enterprise options (Anthropic Claude for Work, Azure OpenAI, Google Vertex AI) that guarantee zero data retention and total isolation, with contractual SLAs. The choice of the right architecture is a fundamental part of the initial consulting: without this, any AI implementation is a compliance and reputational risk.
Free 60-minute AI assessment to evaluate maturity, concrete opportunities, and action priorities. No generic pitches, just analysis of your specific context.
Book assessment