After 2+ years building enterprise AI agents, here's what keeps breaking in production

We've been building enterprise AI agents for a little over two years now, primarily around CRM and business workflow automation.

One thing we've learned is that the biggest challenge isn't getting an LLM to answer questions.

It's getting an AI agent to reliably execute an end-to-end business process.

For example, consider a customer refund workflow.

A production agent has to do much more than understand the user's request. It has to:

• Retrieve the correct customer and order information.

• Apply deterministic business policies.

• Decide whether the refund can be approved automatically or requires a manager.

• Pause execution while waiting for approval.

• Resume from exactly the same state after approval (even if it's days or weeks later).

• Update the correct CRM records.

• Produce a complete audit trail explaining every decision.

Most current agent frameworks do a good job of reasoning, tool calling, and orchestration.

Where things become much harder is with long-running, stateful workflows that involve business rules, human approvals, retries, and compliance requirements.

That's where we've spent most of our engineering effort over the past couple of years.

I'm curious how others are approaching these problems.

• Are you persisting workflow state yourself?

• How are you handling human-in-the-loop approvals?

• How do you make workflows resumable after long delays?

• Are you using a policy engine, workflow engine, or something custom?

• How are you generating audit trails for enterprise customers?

I'm not looking for product recommendations as much as I'd like to understand how different teams are solving these challenges in production.

I'd love to hear what has (or hasn't) worked for you.