Tier-1 support: which requests can an AI agent handle on its own?

Those whose answer already exists in your documentation and whose action is bounded: order status, a procedure, a recurring question, routing to the right team. A dispute, a refund outside the rules or an angry customer goes up to a human, with the summary and the proposed next action.

The Monday morning queue

Monday, 8:40 am. The weekend has left behind a queue of tickets longer than three advisers will clear before noon. A good half asks the same thing: where is my order, how do I change my billing address, how do I reset my access. The advisers know it, they know the answers by heart, and that is exactly what wears them down: you spend the morning repeating what is written somewhere, while the customer with a real problem waits behind.

There is a second consequence, less visible. When ten advisers answer the same question, they phrase it ten ways, and sometimes with ten different contents. A customer who writes twice receives two answers that do not say the same thing, and they are the first to notice. Industry analysts see the same movement coming: according to Gartner, by 2029 agentic AI will autonomously resolve 80% of common customer service issues without human intervention. It is a forecast, not a finding, and above all it says where the line sits: common requests.

The grid: handle alone, prepare, escalate

For each type of request, two questions are enough. Is the answer written down somewhere, and up to date? And is the action that follows bounded, meaning it leaves no room for judgement? Two yeses, the agent handles it alone. One, it prepares and an adviser approves. None, the request goes straight to a human.

The grid is filled in with the support team, on a sample of real tickets, not in a management workshop. An hour with the last hundred tickets of a given month usually gives the split, and a few surprises: a request everyone thought delicate takes three lines, another, judged harmless, hides a special case every time.

The condition: up-to-date documentation

The agent answers from what you have written: help articles, internal procedures, delivery terms, refund policy. If the article is out of date, the answer will be too, and delivered with confidence. Documentation is the real limit of the agent, not the model that reads it. It is the part of the project teams underestimate, and the one that pays off most.

One detail changes the daily routine: the agent flags the questions for which it found no article. Each week, the list of requests with no written answer becomes the list of documentation to write, and it proposes a first draft that the person in charge reviews. The knowledge base fills in where customers really get stuck. The sources stay where they are, in your tools, and the agent reads them with the rights it has been given, as the article on the Knowledge Hub describes.

The escalation rule

Tuesday, 5:55 pm. A customer writes that they were charged twice and intend to cancel. The agent does find the article on direct debits, but the customer’s message contains a duplicate charge, a threat to leave and a rising tone. Three signals, none of which belongs to tier 1. The request goes to an adviser, with a summary of the exchange, the documents found in the case and the reply the agent would have proposed, marked as a draft.

The rule comes down to three conditions, written by the team and changeable by it. A confidence threshold: below it, the agent does not answer, it hands over. A list of subjects that always go up, whatever the confidence: dispute, refund outside the rules, personal data, formal complaint, and anything the team considers sensitive. And an explicit request from the customer, who wants a human and gets one without having to justify it. In every case the adviser receives the full context; nobody asks the customer again what they have already answered.

Practitioners make the same choice. In March 2025 Gartner polled 163 customer service leaders: 95% of them plan to retain human agents. Kathy Ross, Senior Director Analyst at Gartner, sums it up: “A hybrid approach, where AI and human agents work in tandem, is the most effective strategy for delivering exceptional customer experiences.”

Answers that no longer vary

Once the documentation is in place, the same question gets the same answer, whatever the day and whoever asked it. Each answer cites the article it comes from, so the adviser can check at a glance and the customer can read further if they wish. The tone is the company’s: the style guidelines are written once, and the agent keeps to them at 3 am as well as at 10.

When an answer is wrong, it happens, the fix is made at the source: you rewrite the article or sharpen the rule, and every later answer benefits. Correcting ten advisers one by one has never produced that result.

What the supervisor sees

Every request opens a case in Cases, and the case keeps what happened: what the agent read, the answer it gave or prepared, the rule that triggered the escalation, the adviser’s approval or rejection. The supervisor does not have to take the agent’s word for it. They open a case at random, or the one of an unhappy customer, and read the sequence.

From those cases come the numbers that matter for steering: how many requests go up to a human and why, which ones were routed to the wrong team, which subjects keep coming back with no article to answer them. When the escalation rate jumps, the cause is most often a product that has changed and documentation that has not followed. The principle is detailed in the article on Cases.

Measure, before and after

Four measures are enough, recorded over a quarter before the agent and then over the following quarter:

• The first-response time, as a median, by channel: the average is skewed by the Friday-evening ticket handled on Monday.
• The share of requests handled without human intervention, measured on cases closed without reopening within the week.
• Satisfaction on escalated requests: that is where the relationship is won or lost, and it is the measure to look at first.
• The number of misrouted requests, readable in the cases.

If the share handled alone rises but satisfaction on escalations falls, the agent holds on too long to requests it should have passed on: tighten the confidence threshold. If it stays low while the documentation is complete, the grid is too cautious, and you open one more type of request.

Frequently asked questions