Oros Digital
All articles
aiagents

Why 70% of AI agent projects in SMEs fail

Four specific reasons — and the one simple thing the 30% do differently.

Most AI agent pilots in Swiss SMEs never reach production. That is our observation after a dozen audit conversations over the past six months. Four reasons come up again and again — always the same ones.

1. Automating the wrong thing

The most common mistake: picking a visible workflow instead of a costly one. A team wants to "automate customer support" because that is what everyone is talking about, when the thing actually costing them money is manual follow-up on unpaid invoices. Agents shine where volume is high, work is repetitive, and impact is direct.

The simple test: which workflow is your team losing the most time on this week — not in theory, in practice?

2. No human-in-the-loop guardrails

Human-in-the-loop means a human reviews or approves the agent's output before it takes effect. An agent acting without that check on sensitive topics (money, legal, health) is an incident waiting to happen. Almost every pilot that makes it to production has a step where a person validates before anything is sent. That is not a compromise — it is what makes an agent deployable in a regulated context.

3. No feedback loop

A feedback loop is the mechanism that tells you whether the agent is doing its job correctly. Without one — error rate, response quality, human validation rate — you are flying blind. Serious pilots log every action, sample a fraction for manual review, and adjust instructions accordingly.

4. No maintenance budget

An agent in production is software. It needs monitoring, testing when models change, and updates when your business evolves. Failed projects all follow the same pattern: a big pilot, then nothing. The 30% who succeed have budgeted 10–20% of the initial cost per year for maintenance.

What the 30% do differently

They choose one narrow, measurable workflow with high manual cost. They add a human in the loop from day one. They measure. And they maintain.

That is all. No magic. Just engineering discipline applied to a new tool.


If you want to identify the right agentic candidates in your business, a free 30-minute audit is the fastest way to get there.