If you want customers to use your support AI, give them an easy way to leave it. Gartner surveyed 3,566 B2B and B2C customers in early 2026 and found that 50% say GenAI makes customer service easier, yet 87% say access to a human agent is essential when a company uses it. Those two numbers don’t conflict. People like the speed when it works, and they won’t commit to a channel that can trap them when it doesn’t. I think that’s the most useful thing to understand before you decide how much of your front line to hand to a bot.
Why does a way out make customers use the AI more?
Because the fear isn’t the bot, it’s being stuck with it. The same Gartner research found that among customers who won’t use AI for service, the most common thing that would change their mind is being able to switch to a person. Gartner’s analyst also advises against making GenAI a mandatory first step, because forcing customers through repeated failed AI turns makes them less likely to use it again. A bad first run costs you the second attempt, and the third, and eventually the customer’s trust in the channel. It’s the same pattern as an IVR menu that traps callers, just with a friendlier voice.
The competition for that first attempt is also changing. Customers were about three times more likely to have used a third-party AI like ChatGPT, Gemini or Copilot than the company’s own chatbot in their latest service interaction. They already have a place to ask questions without a hand-off to anyone, so your channel has to offer something theirs can’t: it can do the task, and it can bring in a person who already has the context. That’s the case for skill-based routing once the AI has done its part, and the argument only works if a customer can actually reach those people. If the route to a human is buried, you’ve built a worse version of a free tool they already use.
Trust is the other half of it. A separate Gartner survey of 5,801 US customers in early 2025 found that 54% trust humans more than AI for recommendations, against 32% who trust AI more. That’s a majority leaning toward people, and it isn’t a gap a better prompt will close. It means that when the stakes are higher, a refund dispute, a billing error or a service that’s been down for a day, the customer’s instinct is to look for a person, and a design that pretends otherwise just frustrates them earlier. The practical reading for a support leader is that the human route is part of how the AI earns its keep, because it’s what lets a nervous customer try the bot in the first place.
What did Klarna’s reversal teach everyone?
Klarna is the cautionary tale most of us have heard, with the caveat that it’s second-hand and from 2025. In 2024 the company said its AI assistant handled 75% of chats and did the work of about 700 agents. By May 2025, according to a report citing a Bloomberg interview, its CEO said the push had gone too far, that cost had been over-weighted and quality had dropped, and the company started hiring humans again through an on-demand remote model. The same report says Klarna wanted customers to always be able to reach a person. I’d treat it as a reported reversal, not a verdict on the current state of the company.
What stays with me is how a cost metric ran the show. A deflection number is easy to celebrate, and it’s easy to miss what it’s hiding, which is the overreach I was getting at in my reality check on agentic AI. My own view is that AI can do a lot, and it keeps improving, but it doesn’t have the human factor that matters most when a customer is upset or the problem is odd. I’d also push back a little on the cost story itself. With the price of AI tokens and data centres climbing, it may turn out to be cheaper to keep human talent than we assumed a couple of years ago, and the business case for replacing people could look different in a few years than it does today. Either way, the sensible approach is a real partnership between AI and people, not a choice between them.
How do you build the way out in practice?
Start with an audit, and make it boring. Pick every entry point a customer can use, the website, the app, the phone tree and the chat widget, and count how many steps it takes to reach a person from each. Flag every flow where the AI is a mandatory first step, and look at how many customers who touched it came back about the same problem within a few days. That re-contact number tells you more about the AI’s real performance than the deflection rate ever will, because it counts the customers who gave up and called anyway, the same trap as a KPI that passes the test and fails the customer.
Then build the trigger into the product. Nubank, which ran five production support agents for more than 100 million users, listed frustration detection and escalation among the improvements to its card-delivery agent. Its A/B tests showed gains of 37 percentage points in transactional NPS and 29 in self-service over the legacy system. It also noted that these two measures can pull against each other: sending hard cases to humans lifts satisfaction but lowers the self-service number. That’s why I’d always pair the metrics, so a bot doesn’t get rewarded for hanging on to a conversation it can’t win. Let the AI gather context and intent first, attempt a resolution when it’s confident, and pass everything it learned to the person so the customer doesn’t have to start over. If you’re still deciding where the AI belongs, my write-up on implementing AI ticket triage covers the earlier-stage version of that question.
A sane pair of targets helps here. Pick a self-service or containment goal, and put a satisfaction goal right next to it, so neither can be hit at the other’s expense. Add a quality guardrail beside the cost metric in the business case too, which is the step Klarna’s story suggests got skipped. And review a sample of failed AI conversations every month by hand. Reading twenty transcripts where the customer typed “agent” three times in a row will teach you more about where the way out is hidden than any dashboard, and it gives your agents something useful to do with the context the bot gathered.
One more thing to watch is the wording on the way out. “Talk to an agent” should mean exactly that, not a second round of questions the bot has already asked. If the customer has to repeat their account number, their issue and what they already tried, the handover failed even though it technically happened. Treat the context pass as part of the way out, and test it by sitting in the agent’s seat and seeing what actually arrives on screen.
What happens to your agents when the bot takes the easy work?
Probably less than the headlines suggest. A Gartner survey of 321 service leaders found that 31% have made or plan AI-driven frontline layoffs through the first quarter of 2027. But 85% are adding responsibilities to agent roles, 75% are moving agents into new roles, and 63% are reducing headcount gradually through attrition. Role redesign is far more common than elimination, which fits what I see in practice: the simple contacts go to AI, and the person left on the line has the harder, more emotional conversation. That’s a different job, and it deserves different training.
AI can also sit behind the human rather than in front of them. The NBER study of about 5,000 agents at a software company found that an AI reply assistant raised throughput by roughly 14% while customers were less likely to ask for a supervisor and CSAT held steady. Whichever shape you pick, the way out should be visible, short and obvious, and your own team should test it every quarter by trying to reach a person the way a frustrated customer would.
Hutch Morzaria is a Director-level CX and Support Leadership professional with 19 years of experience building global support organizations across SaaS, Fintech, and enterprise technology. He has hired dozens of support and CX leaders across his career and holds ITIL Expert certification across V3 and V4.
