Automation that respects the exceptions

Every team has a process that “mostly” works. The order comes in, the invoice goes out, the ticket gets routed. Then there’s the other 10%: the rush order, the partial refund, the customer who replies to the wrong thread.

That 10% is where naive automation falls apart - and where good automation earns its keep.

Automate the path, assist the exception, escalate the rest

We split every process into three lanes:

1. Automate the predictable path end to end. No human needed.
2. Assist the ambiguous cases - the system does the gathering and drafting, a person makes the call.
3. Escalate the genuinely novel, with full context attached, to the right human.

The mistake is trying to automate all three. The win is being honest about which is which.

Make every automated action reversible and visible

If you can’t see what an automation did, you won’t trust it - and you shouldn’t. Two rules we never break:

• Idempotency: running the same step twice never charges twice or sends twice.
• Audit trail: every action is logged with what triggered it and what it changed.

Measure the thing that matters: time returned

The point of automation isn’t to look clever. It’s to give people their week back. So we measure exactly that - hours of manual work removed - and report it. If a workflow doesn’t move that number, it doesn’t ship.

Automation done well is quiet. It removes the drudgery, respects the exceptions, and leaves your team doing the work that actually needs a human. That’s the only kind worth building.