When should AI hand a ticket to a human?
An AI agent should hand a ticket to a human in four situations: when its confidence in the answer is below the threshold you set, when essential information is missing and clarification did not resolve it, when the user shows frustration or distress, and when the action itself is high-impact or irreversible. Everything else, the repetitive and well-documented work, is exactly what the AI should handle on its own.
The interesting question is not whether AI can resolve tickets. It can, and for a large share of L1 volume it does. The interesting question is whether it knows when it should not.
Why full autonomy is the wrong goal
Vendors love to promise a service desk without people. We think that promise is wrong, and not just technically.
A service desk is a trust function. One confidently wrong answer to the wrong person costs you more goodwill than fifty correct automatic resolutions earn back. The goal is not to remove humans. The goal is to make sure humans only see the tickets where a human actually adds something: judgment, empathy, authority to make exceptions.
That only works when the AI is honest about uncertainty. An agent that always answers is a liability. An agent that knows its limits is a colleague.
Signal 1: confidence below the threshold
Every action our agents take, from setting a priority to sending a customer reply, carries a confidence score. Every action type has a threshold. Reply to the customer: high bar. Suggest an internal note: lower bar. Below the bar, the action does not happen. The AI posts its analysis as a private note and the ticket flows to a person, with all the research already done.
The thresholds are yours to tune. Start strict, watch a few weeks of shadow mode output, and loosen only where the AI has proven itself per category.
Signal 2: the ticket does not contain enough to act
"Printer broken." No location, no model, no error. A human would ask. So should the AI, and it does: one friendly reply with the two or three questions a first-line agent would ask.
But clarification is a single round, not an interrogation. If the answer still leaves gaps, the right move is a handoff with a summary of what is known and what is missing. Users forgive one question. They do not forgive a form disguised as a conversation.
Signal 3: emotion
"This is the fourth time I am reporting this and nobody responds."
That ticket may be technically simple. It is not simple. An automated answer, however correct, tells this user that still nobody is listening. Sentiment detection flags frustration, urgency, and anxiety, and routes those tickets to a person immediately, with a note explaining why. The end-user experience is the product here, not the resolution time.
Signal 4: impact
Resetting one password is routine. Disabling an account, approving an exception, or touching anything that smells like a security incident is not. Some actions deserve a human decision even when the AI is technically certain, because the cost of a rare mistake is asymmetric. Those categories simply stay human, and the AI's job is to arrive with the context prepared.
What a good handoff looks like
A handoff is not the AI giving up. A bad escalation is an empty ticket in a queue. A good handoff means the human starts with a briefing: what the user asked, what the knowledge base says, what was already tried, what the AI recommends, and why it did not act on its own.
That last part matters most. "Confidence 55 percent, two conflicting knowledge articles" tells a person exactly where to look first. The human resolves the ticket faster, and the correction they make becomes new knowledge, so the next occurrence is resolved automatically.
Real service by real people. Administrative work by machines. The handoff is where those two meet, and it deserves more design attention than the automation itself.
Frequently asked questions
What percentage of tickets should AI handle without a human?
There is no universal number. Well-documented, high-volume categories often reach 40 to 60 percent autonomous resolution over time. The honest metric is not the percentage but the error rate within it. A desk that automates 30 percent flawlessly beats one that automates 60 percent with visible mistakes.
Does human-in-the-loop mean a human reviews every AI action?
No. It means every action has a defined level of human involvement: fully autonomous for proven categories, suggest-and-approve for the middle ground, human-only for high-impact work. You decide per category and per action, and you shift the boundary as trust builds.
How do I know the thresholds are set correctly?
Run in shadow mode first and compare the AI's proposals with what your team actually did. Where they consistently agree, raise autonomy. Where they diverge, keep the human in charge and look at why: often the knowledge base is missing an article, not the AI missing intelligence.