Nine in ten tickets: what the tenth one teaches.

The agents do the routine work. Nine in ten tickets, no human. That is the headline. The number that matters more is the one underneath it — the tenth ticket, the one the agent did not finish.

We spent the past quarter watching what happens at that edge. Field notes from three customer ops teams running Lumin in production: a logistics floor in Berlin, a finance back-office in Lisbon, and a regional retail support team in Manchester. Different industries, same agent shape, same kind of tenth ticket.

What the tenth ticket looks like

It is rarely exotic. The pattern repeats across customers.

• A shipper email with a margin too thin for the guardrail to clear.
• An invoice that does not match a PO, because the PO never got entered.
• A refund larger than the agent's bounded authority.
• A customer reply that fits two playbooks at once.
• A new edge: a carrier we have never quoted before.

None of these are model failures. The agent did its job — it stopped, wrote a receipt, and routed the ticket to a human. The interesting question is what happens next.

Think about what would happen without a receipt. The dispatcher would open a ticket and see: agent gave up, please review. She would have to reconstruct from the email what the agent saw. She would have to guess which guardrail tripped. She would have to compare the system of record against what she thought the agent would have done. Most teams call this triage. We call it doing the work twice.

What we watched the humans do

Three patterns showed up, in roughly equal share.

1. They finish the ticket and move on.

One-offs. A wrong attachment. A weekend ticket from a carrier the agent does not handle yet. The dispatcher does the work, files the receipt, closes the loop. No follow-up. These are the easy tens — the agent stopped at exactly the right place, the human spent two minutes, the next ticket gets the same shape next month and the agent will pass through.

2. They notice a pattern and tighten the agent.

After the third refund over €500 in a week, an ops lead opened the agent and bumped the guardrail to €750 with a margin check. The change is itself a receipt — agent diff, signed, dated, applied. The next week, that bucket of tickets went auto.

The same Berlin team did this nine times in eight weeks. Each tightening pushed another sliver of the tenth ticket into the auto bucket. The auto rate climbed from 87% to 94% during that window — not because the model got smarter, but because the people running it kept noticing.

3. They find something the system never could.

A customer reply that read fine on the surface but indicated a churn signal. A shipper running an end-of-quarter dump. A vendor whose VAT number changed three days ago. The agent had no way to know. The human did, and now the playbook does.

This is the third bucket and it is the one that pays the salary. The agent never gets credit for these saves, because they were never agent work. But they happen because the agent freed the human to be in the room when the signal showed up. A dispatcher buried in 90 inbox tickets a day does not catch the churn signal. A dispatcher reviewing seven hand-offs catches it on the second cup of coffee.

The receipt is what makes this work

The pattern only holds because every ticket — auto or human — produces the same receipt. The ops lead does not have to remember why she changed the guardrail. The receipt says so. The next dispatcher sees the receipt, sees the diff, and runs the playbook as written.

Without that thread, the tenth ticket is just an interruption. With it, the tenth ticket is the place where the agent gets better. The cost of an exception is a feature request you can read at the bottom of a JSON file. The cost of forgetting an exception is zero, because you do not forget — the receipt remembers.

Nine in ten tickets pay the rent. The tenth one is the product.

If you are sizing an ops team after shipping an agent, do not size for the ninety percent. Size for the ten. You are not paying the dispatcher to handle the routine — the agent does that for free. You are paying the dispatcher to be in the room when the tenth ticket arrives, to spot the pattern, and to tighten the agent. The math is different. The headcount is different. The org chart is different.

That is the field note. The agent buys the team the time to do the work that actually changes the agent. Build for that loop.