AI for customer service: results or hype?

The most oversold use case in the field and also one of the few with genuinely reliable returns. Those are not contradictory: the returns are real and they are concentrated in a narrower band than the marketing suggests.

Here is the split, based on what actually holds up in production.

Where it works

Questions with one correct answer. Opening hours, whether you cover an address, whether you handle a particular product, what is included in a service, whether you take a payment method. High volume, no judgment, and it consumes a startling share of a support desk's day.

Acknowledging that something arrived. Not resolving it. Just responding immediately so nobody is sitting in silence wondering whether they were heard. Cheap, and it moves satisfaction more than it should, because the alternative in most companies is hours.

Collecting information before a person gets involved. Order number, address, what happened, when it started. A human picking up a case with that already gathered starts several minutes ahead and does not have to make the customer repeat themselves.

Scheduling. Offering real availability, taking a confirmation, sending a reminder. Contained, verifiable, and it reduces wasted appointments.

Routing. Getting a case to the right team on the first attempt rather than the third.

Notice these are mostly not "AI answers customer questions". They are logistics, and logistics is where the reliable value is.

Where it fails

Anything where somebody is angry. A frustrated customer needs a person with authority quickly. A system that keeps trying to resolve them is making it worse in a way that shows up in your reviews rather than in your metrics.

Anything requiring an exception. Support is largely the business of deciding when to bend a rule. A refund outside policy for a customer worth keeping, a delivery escalated because of a circumstance nobody anticipated. Those judgments are the job, and they are not automatable in any way I would defend.

Anything where being wrong is expensive. Commitments about money, dates you have to honor, legal or safety information. If a wrong answer creates an obligation, that category needs review before it reaches a customer.

The conversation that is the relationship. Some contact is the product. Removing a human from it is a cost recorded as a saving.

The failure that damages companies

Not a wrong answer. The loop with no exit.

A customer cannot get to a person, keeps rephrasing, and the system keeps helpfully offering things that do not address the problem. This is the experience that produces genuine anger at automation, and it is a design choice rather than a technical limitation. Somebody decided containment was the metric.

That is the thing to refuse. If the target being optimized is the percentage of conversations resolved without a human, you will get that number and you will pay for it in customers who left without telling you.

Why deflection is the wrong metric

Which follows directly. Deflection rate measures how many people did not reach you, and it cannot distinguish between a question answered well and a customer who gave up.

Measure resolution instead, and measure abandonment. Track how often somebody asked for a person and how quickly they got one. Track your error rate deliberately, because a system that improves speed while quietly degrading quality looks excellent for a quarter. Metric design.

What the honest version looks like

Instant response on everything, always, so nobody waits in silence.

Full resolution on the narrow band of questions with fixed answers, which is a large share of volume and a small share of difficulty.

Information gathered and the case handed to a person for everything else, with the context attached so the customer does not repeat themselves.

An obvious, fast route to a human that is never obstructed, and a system that takes it on its own initiative when the conversation turns emotional or unusual.

That is a smaller claim than "AI handles your customer service" and it is the version that does not cost you goodwill.

Channel matters more than people expect

One practical note. Written channels work considerably better than voice for this, because the customer stays in control of the pace, there is no mishearing, and the exchange leaves a record they can look at again.

Automated voice on an inbound line is the version customers have learned to resent, and the resentment is earned. The comparison in detail is written for one trade but the reasoning is general.

What we would tell you

We build agents that handle inbound conversation, so this is our own category and you should discount accordingly.

What we hold to: the escalation design gets more attention in a build than the conversation does, the route to a person is never obstructed, and we do not optimize for containment. Orelle is the product and the consulting side will tell you if your support volume does not justify it, which for smaller operations it sometimes does not.