Should Customer Support Be the First Team You Automate?

By Victor Hoang · Updated Oct 11, 2026

Ask a room of executives which team should get AI first, and customer support wins almost every time. The logic is easy to follow. Tickets are countable, the work repeats, the answers mostly exist already, and the savings show up on a spreadsheet by the end of the quarter.

That logic is half right, and the wrong half is expensive. Support is a good place to use AI. It's a dangerous place to start without preparation, because support is the one team where a failure doesn't stay inside the company. It lands on a customer, usually one who was already having a bad day.

Our view is that readiness, not ticket volume, should set the order. Volume tells you where the savings could be. It says nothing about whether your knowledge base is accurate enough to automate from, or whether a customer who gets stuck has anywhere to go. Automate before those two things are fixed and you'll cut cost while quietly raising churn.

Key Facts

• A study of 5,179 customer support agents found that a generative AI assistant raised productivity by about 14% on average, and by 34% for novice and low-skilled workers, with minimal impact on experienced ones. The AI worked alongside the agents, not in place of them.
• In February 2024, Klarna said its AI assistant handled about two-thirds of its customer service chats, work equal to 700 full-time agents, cut repeat inquiries by 25%, and brought resolution time from 11 minutes to under 2.
• In May 2025, Klarna's CEO told Bloomberg the company was hiring human support staff again. As Entrepreneur reported it, he said "there will always be a human if you want."
• A Qualtrics survey of more than 20,000 consumers in 14 countries found that nearly one in five who used AI for customer service saw no benefit, almost four times the rate for AI use in general. Half worried AI would stop them reaching a person.
• The US Consumer Financial Protection Bureau has warned that chatbot "doom loops" are often caused when a customer's issue falls outside the chatbot's limited capabilities.

Why Support Looks Like the Obvious First Move

The case for starting here is real, and it's worth stating fairly.

Support has what AI needs: a large pool of past conversations, repetitive question types, and a clear record of what "resolved" means. The best evidence we have on AI in support is positive. In the study of 5,179 agents, the assistant raised issues resolved per hour, and the largest gains went to the least experienced people. The authors suggest the tool spreads the habits of top performers to newer staff and helps them move down the experience curve faster. The paper also reports improved customer sentiment and higher employee retention.

Read that result carefully, though. The AI was assisting a human who still owned the conversation. A person was checking the suggestion against the customer's actual problem. That's a different system from a bot that answers on its own, and a different risk profile. Plenty of the excitement about "automating support" quietly swaps one for the other.

Klarna's own announcement shows the appeal of the second model. Two-thirds of chats handled, 700 agents' worth of work, resolution in under two minutes. If you were a CFO reading that in early 2024, you'd have asked why your company wasn't doing the same.

What Happens When You Skip the Preparation

Fifteen months later, Klarna's CEO was telling Bloomberg that the company was recruiting human support staff again so that customers could always reach a person. Entrepreneur's account is that AI chatbots were cheaper but produced lower-quality output, and he described investing in the quality of human support as the way forward. Klarna didn't abandon AI. It re-balanced.

We wouldn't read that as "AI failed at Klarna." The more useful reading is that a metric chosen for cost (chats handled, agents replaced) was allowed to define success, and the harder-to-count things, like the customer with an unusual problem, got less attention. The Qualtrics researcher quoted in that company's survey release put the pattern bluntly: too many companies are deploying AI to cut costs, not solve problems, and customers can tell the difference.

Regulators have named the mechanism. The CFPB describes doom loops, where a customer repeats the same request, gets the same canned answer, and can't find a route to a human. Its report adds that chatbots sometimes get answers wrong and that their ability to recognize a dispute may be limited. None of that is exotic. It's what you'd expect when a system reaches the edge of what it knows and nothing catches the customer.

Notice what is common to those failures. They aren't model problems. They're content problems (the system had nothing accurate to say) and routing problems (the system had nowhere to send the customer). Both are fixable, and both are fixed before launch or not at all.

The Two Things That Decide Readiness

The knowledge base is the product

When an AI answers customers, your knowledge base stops being a help-center side project and becomes the thing the customer actually receives. If an article says the refund window is 30 days and policy changed to 14 last spring, a human agent might know better. The AI will repeat 30 days with perfect confidence, thousands of times.

So the first readiness question is dull and decisive: who owns each article, and when was it last verified against the real process? Duplicate articles that disagree, instructions for screens that no longer exist, and answers that live only in a senior agent's head are the usual culprits. We've written about why undocumented know-how becomes the AI bottleneck, and support is where it bites hardest. If you're building retrieval over your help content, the same logic runs through ticket deflection with RAG, and the upkeep work is covered in AI knowledge base maintenance.

Escalation is the safety net

The second question is where a customer goes when the AI can't help. A good escalation path does four things. It triggers on clear signals (repeated rephrasing, negative sentiment, billing or legal keywords, an explicit request for a person). It hands over the full conversation so the customer doesn't start again. It reaches someone with authority to fix the problem. And it's fast enough that the customer doesn't give up first.

Most teams build the bot and treat the handoff as an afterthought. It should be the reverse. The design of the exit determines how bad your worst interactions get. For a view of how an agent can own that routing, see the AI escalation manager agent.

A Readiness Scorecard

Score each line from 0 (not true) to 2 (fully true) for the specific ticket category you're considering. Do it per category, not for support as a whole. "Password reset" and "disputed charge" are different decisions.

Total the scores. A category at 14 or above is a candidate for automation. Between 9 and 13, use AI to draft replies for agents while you fix the gaps. Below 9, fix the foundation first. These thresholds are our judgment, not a standard, so adjust them to your risk tolerance. How much error you can absorb in a given workflow is its own question, and the error budget approach gives you a way to set it.

Sequence by Readiness, Not Volume

Put the scorecard next to your ticket data and a different order appears. High-volume categories are tempting, but volume and readiness often point in opposite directions. The categories with the most tickets are frequently the ones with the messiest edge cases.

Stage one is where the study evidence is strongest, and it's also the best way to find out what's wrong with your knowledge base. Every time an agent rejects a suggestion, you've found a gap for free. That's cheaper than letting customers find it.

One caution on measurement. Cost per ticket and deflection rate are the numbers that make the business case, and they're also the numbers most likely to mislead. A customer who gives up isn't a deflected ticket in any way that matters. Watch repeat contacts within seven days, escalation rate, and, above all, what happens to retention for customers whose issues were AI-handled. Our piece on support signals that predict churn lists the early warnings worth wiring in.

And don't let "a human can step in" become a comfort blanket. If your agents see only the cases the AI already failed, they inherit the angriest customers with the least context. We looked at a related risk in the human-in-the-loop rubber stamp.

What to Do Next

For a COO, head of support or CX leader, a workable plan fits into a quarter.

1. Inventory ticket categories. List the top 20 by volume and tag each by error cost: cheap and reversible, or costly and hard to undo.
2. Score the top ten with the table above. Be honest on content accuracy. Have someone outside the support team spot-check ten articles against the real process.
3. Assign an owner to every article in the categories you plan to automate, with a review date. No owner, no automation.
4. Write the escalation rules before the bot goes live. Test them on 50 real past tickets, including the ugly ones.
5. Start in assist mode. Let agents use AI suggestions for a month and record where they reject them. Fix those articles.
6. Automate one clean category with a visible way to reach a person, and set the pass/fail measures in advance: repeat contact, escalation rate, satisfaction, and churn for that cohort.
7. Review monthly. Expand only when the numbers hold. Pull back when they don't. For the broader picture of AI across B2B service, see AI in B2B customer support.

So should support be the first team you automate? It can be the first team you give AI to. Whether it's the first you hand customers over to depends on a score you can calculate this week.

Learn More

• How Much AI Error Can Your Business Afford?
• The Human-in-the-Loop Rubber Stamp
• Undocumented Know-How: The AI Bottleneck
• AI in B2B Customer Support