L'AI come capacità architetturale

Tipo

Concetto

Artificial Intelligence is not just a feature, a model, a chatbot or a vendor capability. For a Solution Architect, AI should be treated as an architectural capability: something that changes how a business process makes decisions, produces recommendations, automates work, handles uncertainty and interacts with people. The first architectural question is not which model to use. The first question is what decision, task or user outcome AI is expected to improve. AI is useful when the system must classify, summarize, search, recommend, extract, predict, generate or assist with work that is too variable, language-heavy or context-dependent for simple deterministic rules. It is weaker when the organization actually needs traceable business rules, exact calculations, strict repeatability or legally explainable decisions. AI systems introduce specific architectural concerns. Data quality becomes part of behavior, not just reporting. Prompt design, model selection, retrieval strategy, evaluation, latency, cost, privacy, security and human review all become design decisions. A generated answer may be fluent but wrong. A model may work in a demo but fail with edge cases, ambiguous user intent, missing context, outdated knowledge or adversarial input. This means AI needs boundaries, validation, fallback behavior, observability and clear ownership. A Solution Architect should distinguish three things: the AI capability, the system that hosts it and the business process that relies on it. The AI capability may generate a draft, classify a request or suggest an action. The hosting system must enforce permissions, protect data, validate inputs and outputs, log decisions and manage errors. The business process must define who remains accountable, when human review is mandatory and what happens when confidence is low. In enterprise architecture, AI should rarely be allowed to silently become the system of record. It can assist, enrich or accelerate decisions, but authoritative data, approval rights and audit trails must remain explicit. A good AI architecture makes uncertainty visible. It records what context was used, what output was produced, which checks passed, who reviewed the result and whether the output was accepted, rejected or corrected. The common mistake is treating AI adoption as a technology integration only: calling an API, sending a prompt and displaying the response. The better architectural approach is to design an AI-assisted workflow with clear intent, data boundaries, evaluation criteria, human accountability, monitoring and rollback options. AI should serve people and decisions, not hide responsibility behind automation.