A shopper is one tap from buying. The cart's loaded, the card is out, and they have one last question before they check out: can they return this if it doesn't fit?
So they ask the chatbot, and it sends them a link to the FAQ page. They click, scroll, don't find a straight answer, and close the tab. The bot logs the chat as handled. The customer logs it as a reason to shop somewhere else, and you never even see it happen. There's no ticket for the shopper who gave up and left.
The chatbot recognized the word return and fired off a generic policy link, and it had no idea who it was talking to. It couldn't see that this shopper is a VIP with a longer return window, or that the item is final sale and can't come back at all. It answered the keyword. It missed the customer.
Customers feel that, and they act on it. In a December 2025 SurveyMonkey study of more than 2,000 US adults, 89% said companies should always offer the option to reach a human, and 79% said they prefer a person over an AI agent. That's what bad automation earns you. When the bot leaves people stuck, they go looking for a human, and if they can't find one, they go looking for another brand.
The technology to do better already exists. Whether you have it comes down to one thing, and it isn't the AI. It's the data underneath it.
What separates a chatbot from an AI customer agent
A scripted chatbot matches keywords to canned replies. An AI customer agent reads the situation, pulls up the customer's actual data, and does something about it. Same chat window, but a completely different machine behind it.
Say a shopper asks whether their order will arrive before Saturday. The chatbot hears the words order and arrive and serves up the shipping policy page, same as it would for anyone. The AI agent looks at this specific order, sees it shipped yesterday on 2-day delivery, and tells the shopper it's on track to land Friday. One of them answered a question category. The other one answered a person.
Most chatbots can't do that because of how they're built. They sit on top of your site like a widget bolted to the bumper, walled off from everything you actually know about the customer. They can't see that someone just placed an order, that they've been circling the sale section for 20 minutes, or that they emailed support about a sizing issue last week. So they ask for an order number you've already given them, twice, and treat a 10-year regular exactly like a first-time visitor who wandered in off an ad.
An AI agent that runs on a real customer profile starts from the opposite place. It already knows the order, the history, and the loyalty tier before the customer finishes typing, so the conversation picks up right where a chatbot's tends to stall out.
The 4 things that actually separate the two
Plenty of tools will tell you they have AI. What you want to know is whether that AI can actually do the 4 things below, because that's where deflection and resolution part ways.
- It reads the live customer profile, not last night's export. When someone asks about a return, the agent already knows what they ordered, when it landed, and what their window looks like, because it reads the current state of their account instead of a data sync that ran at 2 a.m. The customer gets a straight answer in one message instead of a request for an order number you already have.
- It handles how people actually talk. Customers misspell, change the subject mid-chat, and ask follow-ups that point back 3 messages. The agent has to hold the thread across the whole conversation instead of resetting with every reply.
- It takes action, not notes. The most frustrating chat ever written ends with the bot promising it has created a ticket for the team. An agent that can process the return, apply the credit, cancel the item, or fix the shipping address closes the issue instead of forwarding it.
- It hands off with the whole story attached. When a conversation does need a person, that person inherits the full chat, the customer's history, and whatever the AI already figured out. Nobody should have to re-explain a problem they just spent 5 minutes explaining to a bot.
Klaviyo's Customer Agent does all 4 because it runs on the same customer data that powers your marketing. So your shopper asks about their order and gets an answer from something that works like a person at your store who already pulled up their account, not a script that's never met them.
When AI should resolve, and when a human should step in
The goal was never to automate everyone off the team. It's to hand the repetitive stuff to AI so your people get their time back for the conversations that actually need a person.
AI is good at the questions that have a clear right answer: order status and tracking, product availability and specs, simple account changes, return-policy lookups, a discount code, a product recommendation pulled from what someone has browsed and bought. High volume, low ambiguity, and fast to settle.
People should keep the rest: the complaint that needs empathy and a judgment call, the exception that lives outside the policy, the high-value relationship where a human touch is the whole point, the upset customer who needs to be heard before anything else, the edge case nobody trained the AI on yet. You decide where that line sits, and you set the escalation rules. Nothing routes to the AI, or goes live at all, without your sign-off.
The handoff runs both ways. When Customer Agent resolves something, that interaction feeds back into the profile. A customer who just had a return processed probably shouldn't get a hard upsell email tomorrow. A note thanking them for their patience, with a loyalty perk attached, reads the room a lot better. The service conversation makes the next marketing message smarter, and the marketing data makes the next service reply more relevant.
Why this only works when the data is connected
A bolted-on chatbot feels broken because it's cut off from everything that would make it useful. It can't see what you just bought, which emails you've gotten, what segment you're in, or what you're worth over a lifetime. It's stranded on its own little island of data, which is why it treats a first-time browser and a decade-long regular the same way.
Customer Hub and Customer Agent run on the same data as your marketing. When a customer opens a chat, the agent already sees their loyalty tier, what they browsed today, any open tickets, their order history, and how they engage with email. All of it shapes the reply.
So a loyalty member starts a chat about a jacket they looked at this morning. In the first reply, the agent knows their VIP status comes with free expedited shipping, sees the size they want is running low, and flags that the jacket qualifies for their member discount. That answer is built for this one person, and that's the part a keyword chatbot can't fake.
The brands doing this have the numbers to show for it. Protective-case maker Nanuk trained an agent on their catalog, and it now resolves 84% of inquiries on its own. Tea brand Harney & Sons ran the same play on a sprawling catalog, where the hard part is steering a shopper to the right product, and the agent drove a 77% increase in product recommendation queries it resolved in 60 days. Neither number comes from a smarter script. They come from an agent that can actually see the customer it's talking to.
What the shift actually looks like on the team
Support roles don't disappear in this model. They move up. Your agents spend less time answering the same question for the 50th time and more time on the hard cases, and on training the AI to handle more of the easy ones. Most brands get there the same way.
Start with the high-volume, low-complexity queue: order status, shipping, and return policy. These are the conversations eating your team's hours without needing anyone's judgment, so let the agent earn its trust there before it touches anything else.
Then measure the right number. Deflection rate is the vanity metric, the one that counts a customer who gave up as a win. Resolution is the number that matters. Did the customer actually get their answer? Did they finish the purchase? Did they come back? Once the agent is closing those reliably, widen its lane: more query types, more actions it's cleared to take, more nuance in how it responds. It gets sharper as it handles more volume and as your team tightens its training.
The bigger payoff shows up later, once the 2 sides start feeding each other. Your marketing gets sharper because you know who hit a problem and how it got solved. Your service gets sharper because the agent knows what someone's been eyeing before they ask. It's the same customer and the same data, working both sides.
The difference is whether the AI can see your customer
The choice was never humans versus AI. It's AI that deflects versus AI that resolves, and what divides the two is whether it can read the person on the other end of the chat.
The brands pulling ahead on customer experience aren't swapping people for bots. They're giving their AI agents the data and the authority to handle what they're good at, so their people can spend their hours where humans win: the complicated, the emotional, the high-stakes conversations that need a real person on the other end.
Remember the shopper with the return question? An agent connected to real customer data would have seen their order, known their return window, and answered in seconds. The sale closes, the customer stays, and no one ever sees the FAQ page.
Klaviyo's Customer Agent runs on the same real-time profile as your marketing, so every service conversation makes the next marketing message smarter, and every marketing signal makes the next service reply more relevant. Get started with Klaviyo Service.
Related content
