AI Agents in Customer Support: Where It Breaks
Most tickets are the same forty questions.
The loop, as it actually runs
Most customer support processes are a fixed sequence. Writing it out is the fastest way to see where the time goes:
- Ticket arrives
- categorised
- routed
- answered
- resolved
- tagged and closed
Where it breaks
The failure is almost never the steps themselves. It is the joins between them — the points where a person has to notice something and act:
- Repeat questions consume the queue
- Routing is done by whoever reads it first
- Context is lost on escalation
- Tagging is skipped so reporting is useless
Every one of those is a noticing problem rather than a thinking problem. That distinction matters: software is reliable at watching continuously and unreliable at judgement. Automate the watching, keep the judgement.
What an agent takes over
A customer support agent sits on the joins. It watches the systems of record continuously, moves each item to its next state when the conditions are met, chases what has stalled, and escalates the genuine exceptions to a person with the context already assembled. It runs inside Zendesk, Slack, Shopify, Stripe, Gmail.
What it does not do is make the calls that need judgement. Those still route to a person — just faster, and with the file already complete.
How to tell if this is worth doing
- Does the process run more than weekly? Below that, the build cost rarely pays back.
- Can you write the rules down? If two people on your team would handle the same case differently and both be right, it is a judgement call and should stay with a person.
- What does an error cost? High-volume, low-error-cost work is the sweet spot. High-error-cost work needs a human approval gate, which is fine — it just changes the design.