In 2024-2025 every startup talked "agentic workflows". By May 2026 the noise died down and we can see what really entered Italian SME processes. Short answer: less than promised, more than we expected.
Three use cases that won
1. Customer support tier 1
Chatbots with knowledge-base access + ticket creation. Handles 40-60% of requests without human handover. Not "agents" strictly, but the label stuck in marketing. Cost: $50-150/mo in tokens for a typical SME. Clear ROI.
2. Structured extraction from documents
Invoice PDFs, contracts, orders → structured JSON. With late-2025 multimodal models accuracy crossed 85% on real Italian documents — enough to make replacing OCR + regex with LLMs economically sensible.
3. Sales operations
Lead enrichment, qualification, drafting personalized emails. The pattern: agent works in the CRM, human signs off. Works well for B2B outbound with long cycles.
What we killed
"Do everything" agents
Generalist agents like Devin-for-coding or Auto-GPT-for-business didn't deliver stable value. They fail unpredictably, and debugging a 30-step chain costs more than doing the task by hand.
Autonomous agent-to-agent
Multiple agents talking to each other to solve a problem felt like the future in 2024. In practice it introduces loops, reinforcing hallucinations and runaway token cost. We tried it in 3 projects and removed it from 3 projects.
What changed in the models
Claude 4.6 and 4.7 made tool use far more reliable than in 2024. GPT-5 Vision handles scanned documents at quality that was unthinkable a year ago. Gemini 2.5 closed the long-context gap.
For Italian SMEs: no lock-in. The three players now offer comparable price and quality. Pick by where your data layer lives.
The decisive factor: business context
Agents that work all share one thing: structured access to company data. Without a good data layer — clean CRM, written knowledge base, tagged documents — the agent is a bright intern left without instructions.
70% of our 2026 AI work is data prep, not prompt engineering. Predictable in hindsight, but many clients still thought "putting ChatGPT on it" was enough.
Operational verdict
AI agents in 2026 work in specific, well-fenced use cases. They're not a revolution, they're a new tool for tasks that previously required heavy repetitive manual work. The SME that adopts them well saves time. The SME that adopts them badly burns money on an MVP that never ships.