Where AI actually helps in customer support

Most support inboxes are not full of hard problems. They're full of the same ten questions, asked by different people, over and over. "Do you ship to Ireland?" "How do I cancel?" "Is there a student discount?" "Where's my order?" None of these need a human to think. They need a human to look something up and paste it back, which is a waste of a human.

That's the part AI is genuinely good at. Not "customer support" as a whole — the repeated lookup. If you go in expecting the first and get the second, you'll do fine.

Start by reading your own transcripts

Before you automate anything, spend an hour reading your last few hundred support conversations. It's tedious and it's the most useful hour you'll spend. You're looking for two things: which questions repeat, and how you actually answer them. Not the polished help-center version — the real reply your team sends at 4pm on a Tuesday.

When we did this for our own inbox, a third of everything was two questions. Not two categories. Two questions. That's the shape you want to find, because that's where automation pays off immediately and where a wrong answer does the least damage.

What AI should take, and what it shouldn't

Here's roughly where we draw the line.

Good candidates for an AI answer:

• Anything with a factual, stable answer that lives on your site already. Shipping zones, return windows, opening hours, what's included in a plan.
• "How do I..." questions where the steps don't change. Reset a password, change an email, export data.
• The 20 questions your team is quietly sick of answering.

Bad candidates:

• Anything involving money moving the wrong direction. Refunds, disputed charges, "I was billed twice."
• Anything where the customer is already angry. An AI reply, however correct, reads as a brush-off.
• Anything you'd want a record of a human deciding. Cancellations of large accounts, anything legal, anything about someone's data.

We wrote a whole separate piece on what an agent should refuse, because it matters more than the list of what it should answer. The short version: a confident wrong answer about a refund costs you more than fifty right answers about shipping earn you.

The thing that makes or breaks it

An AI support agent is only as good as what it's allowed to read. This is the part people underestimate. If your help content is thin, out of date, or contradicts itself across three pages, the agent inherits all of that and delivers it faster. Speed doesn't fix wrong.

Engine64 answers only from the content you give it, and it shows the source on every reply. We built it that way on purpose — not because it's a nice feature, but because the alternative is an agent that invents a return policy you don't have. When it can't find an answer in your content, it says so and hands off, rather than guessing. That "I don't know, let me get someone" is doing more work than it looks like. It's the difference between a tool your team trusts and one they quietly turn off after a week.

So the real prep work isn't technical. It's writing down the answers to your top questions clearly enough that both a customer and a machine can follow them. If you've never written your return policy in one plain paragraph, you'll discover you don't actually have one — you have three slightly different ones on three pages. Fixing that helps your human team too.

Don't measure the wrong thing

The tempting metric is "how many tickets did the AI deflect." It's tempting because it's big and it goes up. But a deflected ticket where the customer left confused isn't a win, it's a delayed, angrier ticket. We'd rather watch two things: how often the agent hands off cleanly when it should, and whether the questions reaching humans are actually harder than they used to be. If your team is still fielding "do you ship to Ireland?" after a month, the automation isn't working, no matter what the deflection number says.

And keep a human in the loop on the escalations from day one. The failure mode we've seen most is a support agent that answers the easy 80% brilliantly and then strands the hard 20% in a dead end with no way to reach a person. That last 20% is where your reputation actually lives. Automate the boring part so your team has more time for it, not less.