Why your AI chatbot failed

Most failures are workflow misfits, not model failures. The chatbot did its job — you picked the wrong job.

Your chatbot didn’t fail because GPT got worse. It failed because you pointed it at the wrong workflow.

The pattern is consistent across every “AI pilot” I’ve audited in the last 18 months. Someone gets excited about agents. They pick “customer support” because it’s visible. They wire up a bot. The bot answers easy questions correctly. Then it hits a tier-2 ticket — a billing edge case, a multi-step migration, a refund involving a partner — and it lies. Users notice. Trust collapses. The bot gets quietly retired. Six months later: “AI doesn’t work for us.”

The agent isn’t the problem. The job is.

A simple test

Before you point an agent at a workflow, ask:

Can a junior employee do this with a written checklist?
Does the workflow live inside one system, or does it span 4+ tools with unstructured handoffs?
When something edge-case happens, is the escalation path clear?

If “no” to any of those, an agent will fail there. Not because the agent is bad. Because the workflow isn’t agent-shaped.

What agent-shaped actually means

The workflows that win with agents look boring. Inbox triage. Invoice classification. Lead enrichment. SDR follow-ups. Status digests. These don’t require judgment — they require repetition done correctly.

The workflows that lose with agents look exciting. Customer success. Sales calls. Strategy. Anything where context outside the system matters.

What to do instead

Pick the most boring workflow that costs you the most hours.

Not the most visible. Not the most “AI-friendly” on a vendor’s pitch deck. The most boring and expensive.

That’s where the first agent goes. That’s where the math works. That’s where the team stops pushing back.

Get one boring workflow right. Then expand.

— Yoann