AI hype is loud. AI ROI is quieter. Here are the use cases where small and mid-size businesses are getting real money back, and the ones that are still pre-revenue.
Where AI is paying back today
Three categories pay back consistently in 2026: document processing (invoices, receipts, contracts), customer support triage (categorize and draft responses), and lead enrichment (clean inbound, score, route). The pattern is the same: human-in-the-loop, narrow scope, clear measurement.
Where AI is still hype
Open-ended customer-facing chatbots, AI-driven sales agents, and broad replace-a-team-member systems. These are still mostly demos. They work in a controlled demo and break in real-world variance.
Why narrow beats broad
Narrow automations have measurable success rates. You can build an eval set, score it, and improve. Broad agents fail in ways you can't see until customers complain. Narrow ships; broad burns.
Cost per task, not cost per token
The right metric is cost per completed task vs. the same task done by a person. Most useful SMB automations cost five to fifty cents per task when measured correctly. That's about a hundredth to a thousandth of human cost. The savings are real but the engineering is non-trivial.
Evals are the work
What separates a working automation from a hallucinating one is the eval set. A few hundred labeled examples with edge cases. Without evals, you're shipping vibes.
Guardrails are non-negotiable
Anything customer-facing or money-handling needs guardrails. Human-in-the-loop for high-stakes outputs. Tool restrictions. Output validation. Audit logging. The MSP and software providers who skip these create new risk faster than they reduce old cost.
Practical SMB starting points
Three good first projects: invoice intake from email to AP system, support ticket triage and draft reply, and lead enrichment from public sources. All three pay back in months and teach the team how to scope future automations.