There’s a particular kind of helplessness that hits when you’re deep in a customer service chat, you’ve explained your problem three times, and the bot responds with: “I understand your frustration. Let me look into that for you!” — then gives you the same FAQ link it gave you two messages ago.
This isn’t a glitch. It’s the feature.
The Promise vs. The Product
The pitch for AI in customer service is seductive: faster resolutions, 24/7 availability, lower wait times, happier customers. And in narrow, well-scoped scenarios — “what’s my account balance?”, “when does the store close?” — chatbots work fine. These are lookup problems. The bot retrieves a value. Done.
But real customer service problems aren’t lookup problems. They’re exception problems. Something broke, something didn’t ship, you were charged incorrectly, your account is in a weird state. Exception problems require human judgment, system-level access, and authority to deviate from the default. Most chatbots have none of those things.
What they have instead is a decision tree and a policy manual — neither of which covers the situation you’re actually in. So the bot loops. It asks you to rephrase. It offers you the escalation path that leads to a 45-minute hold. It provides the email address of a support team that takes four business days to respond. It does everything except change anything.
Friction as a Strategy
Here’s the uncomfortable subtext: for many large companies, friction in customer service isn’t a bug — it’s a design choice. If 40% of customers give up before reaching a resolution, that’s 40% of refunds you don’t issue, 40% of complaints you don’t log, 40% of policy exceptions you don’t grant. AI chatbots can scale that friction in a way that a human call center never could. A human agent, faced with a clearly frustrated customer, often just solves the problem. The bot doesn’t feel bad. It doesn’t have a bad day and cave. It runs its script at perfect consistency, forever.
The chatbot isn’t replacing customer service. It’s replacing the awkward moment where a company would have to explicitly say no.
When the Underlying Service Fails
The worst version of this is when the underlying product or service genuinely fails — an outage, a billing error, a logistics breakdown — and the chatbot sits in front of it like a polite paper wall. The bot can’t acknowledge the failure because it hasn’t been told about it. It can’t issue a credit because it lacks the authority. It can’t tell you when things will be fixed because no one connected it to the status feed. All it can do is confirm that yes, it received your complaint, and yes, it has escalated it, and no, it cannot tell you anything else.
Meanwhile the problem grinds on. The package doesn’t move. The charge stays on your statement. The service stays down. The chatbot remains cheerful.
The Accountability Gap
Part of what makes this so corrosive is what it does to accountability. When a human tells you “I can’t help you with that,” there’s at least a face attached to the failure — a name, a department, an implied escalation path. When a bot tells you the same thing, there’s no one to ask for a supervisor. The bot is the supervisor. The company has inserted a system that diffuses responsibility across an algorithm while retaining the appearance of responsiveness.
“We have 24/7 support” is technically true. That support just can’t do anything.
A More Honest Conversation
I’m not arguing that AI has no place in customer service. But there’s a meaningful difference between AI that genuinely resolves issues — with real system access, real authority to make exceptions, and real feedback loops that improve its accuracy — and AI that exists primarily to intercept and exhaust customers before they reach someone who can actually help.
If you’re building one of these systems, ask yourself: what percentage of the issues your chatbot handles get resolved in that conversation, without human escalation? If you don’t know that number, or if it’s embarrassingly low, you’re not running a customer service AI. You’re running a complaint filter.
And if you’re a customer reading this: what’s been your experience? Have you found AI chatbots that actually helped, or is it friction machines all the way down? I’m genuinely curious whether good examples exist at scale — or if the incentive structure makes them basically impossible.