AI in PMI: il ROI non è nell'agente, è nell'orchestrazione
Why small businesses automating with AI fail — and how to avoid it
There's a data point that AI vendors prefer not to highlight: according to the McKinsey Global Survey on AI (2023-2025 editions), the majority of AI automation initiatives in SMEs do not reach break-even in the short term. The main causes are not related to technology limitations, but to how it is implemented: as an isolated component, not as part of a system.
The typical mistake is buying or integrating an AI agent and treating it as a virtual employee to whom you entrust a task. "Manage the emails. Generate the reports. Respond to customers." And then being surprised when the result is a plausible but useless text, a pipeline that breaks at the first unexpected input, a customer who receives a perfect answer to the wrong problem.
The Real Problem: The Skeleton Is Missing
What's missing in most SME implementations is not a more powerful model. It's an orchestration architecture.
An AI agent without an orchestrator is like a surgeon without an operating team: they know what to do, but there's no one to hand them the scalpel, monitor the parameters, or manage emergencies. The result is that the surgeon operates at 40% of their capacity — or worse, they perform well in controlled conditions and crash at the first unforeseen event.
In the work we do on Silicea — building a multi-agent system with a kernel, persistent memory, tool execution layer, and cognitive monitoring — we learned this lesson firsthand. It was not enough to give Silicea the ability to reason. We had to build the system that allows her to act in the right way, at the right time, with the right data. That is orchestration.
What "Orchestration" Means in Practice
Three concrete components:
1. Intelligent task routing. Not every request goes to the same agent. An SME handling 200 orders a day doesn't need an LLM that analyzes every order with the same depth. It needs a system that classifies: this order is standard → automated workflow. This order has an anomaly → escalation to the agent. This customer has a pattern → notify the sales rep. Routing is 60% of the value.
2. Persistent operational memory. An agent that doesn't remember what it did yesterday is an agent that asks the same questions, repeats the same errors, and loses context every session. For an SME, this translates into hours of human work wasted re-contextualizing. A well-designed memory system — like the one we're building with our Memory Server — eliminates this friction.
3. Guardrails and fallback. Every agent must have a defined behavior for when it doesn't know. It doesn't "make things up." It doesn't "try." It says "I don't know, here's what you can do." This seems like a limitation. It is actually the feature that prevents economic damage.
The Immediately Applicable Insight
If you're an SME evaluating AI integration, before choosing the model, map your five highest-volume repetitive processes. For each process, ask yourself: "What is the hourly cost of this activity? How many times a day is it performed? What is the current error rate?"
Then, and only then, ask yourself: "Which of these five processes has a decision flow structured enough to be automated with a simple agent, without needing human judgment at every step?"
That is your starting point. Not the most spectacular. The most profitable.
On Progetto Siliceo we are building the orchestration infrastructure that allows an AI agent to operate reliably in a real-world context: persistent memory, tool execution, fallback, and monitoring. The kind of architecture you need when AI is not an experiment but an operational asset.
If you're evaluating how to bring AI into your company and want to avoid the mistakes that cause most projects to fail, let's talk about it.