Automation should know when to stop
The best systems automate repetitive work without removing human judgment.
There is a common assumption that the goal of automation is to remove people from a process entirely. Measured that way, any system still requiring human involvement looks unfinished.
We think that framing produces worse systems. The useful question is not how much of a process can be automated, but which parts should be.
Repetitive work is the right target
Some work is mechanical, high-volume and identical every time. It benefits from being handled the same way at 3am as at 3pm:
- Noticing that an inquiry has arrived.
- Recording it with its channel, timestamp and content.
- Extracting the details a person would otherwise re-read the thread to find.
- Classifying what is being asked.
- Routing it to whoever should handle it.
None of this requires judgment. All of it degrades when a person is tired, busy or away. This is exactly the work software should absorb.
Some decisions should stay with people
Other work looks similar on the surface but is not. Whether to make an exception for a long-standing customer. How to respond to someone who is upset. Whether an unusual request is an opportunity or a distraction. When a quiet conversation is worth one more follow-up.
These decisions depend on context that is not in the message — the relationship, the state of the business, what happened last time. Automating them does not remove the judgment; it just applies worse judgment faster.
Knowing when to hand over
The quality of a system is largely determined by how well it recognizes its own limits. A system that answers confidently when it should have deferred causes more damage than one that defers slightly too often.
In practice this means designing the handover as a real feature rather than a fallback: the person receiving the conversation should get the full context, not a notification telling them to go and find it. A good escalation feels like a colleague briefing you. A bad one feels like being handed a problem.
It also means defaulting to caution. When a system is unsure whether something is routine, treating it as not routine is usually the cheaper mistake.
The point is visibility, not absence
The aim is not a business where nobody touches customer communication. It is a business where the people involved spend their attention on the conversations that actually need them.
When the mechanical layer is handled reliably, the important work becomes easier to see. That is the real return — not headcount removed, but attention returned to where it matters.