Most teams plan an AI customer agent the way they plan a product catalog. They list what they sell, then try to cover it.
I've found it works better to start with what customers actually say in their chats.
As the ecommerce operations manager of ECD Digital Strategy, I lead AI automation for our DTC brands. Klaviyo Customer Agent is the AI customer agent that works from the same customer profile as your email, text messaging, web chat, and WhatsApp.
That shared profile is what makes the powerful use cases below possible.
The strongest Customer Agent builds share a few traits:
- Volume: Zac Fromson, cofounder of Lilo Social, describes a fragrance sample brand with thousands of monthly conversations where "most are repeat questions with known answers." Questions like that are the easiest to hand to an AI agent.
- A clock: Fromson points out that on a subscription brand, "most subscription questions are routine and come with a deadline." If nobody answers before the next charge, the order ships anyway. An AI agent that replies in seconds closes that gap, preventing the shipment from shipping if there are issues.
- Access to the systems where the answer lives: Roddy Smith, general manager North America of Andzen, an award-winning CRM lifecycle agency for DTC and ecommerce brands, built a connection to a studio's booking platform so the AI agent and the marketing team "share the same live view of each member." Once an AI agent can see and act on live data, a lot more of your conversation volume becomes resolvable.
- Written rules: Each build below spells out what the AI agent can say, what it can do, and when it hands the conversation to a person. That's what lets the team trust it with real customers.
Here are 7 builds from agencies in the field, which cover use cases ranging from order tracking to class bookings to the follow-up emails that go out after a chat ends.
1. Start with the questions customers already ask
Fromson's team works with a DTC fragrance sample brand that gets thousands of support conversations a month. About 36% of them need live order data to answer, such as:
- "Where is my order?"
- "It hasn't arrived."
- "How do I cancel?"
How it works: live order data and late-order alerts
Fromson and his team at Lilo Social use Customer Agent's out-of-the-box skills for looking up and changing orders, drawing on the brand's Klaviyo and Shopify data. In addition, they’ve built a custom Claude skill will connect the brand's warehouse system and add two things:
- A projected delivery date before a tracking number exists
- Flags for orders that are running late
Late orders also trigger a Klaviyo flow message before the customer writes in. Fromson says this approach "cuts the number of tickets instead of only automating the replies." When Customer Agent cancels an order, it "confirms the cancellation in every downstream system before telling the customer it's done."
When Customer Agent cancels the order, it confirms the cancellation in every downstream system before telling the customer it's done.
Lilo Social
2. Put the brand's rules in writing
For the same fragrance brand, Fromson says about 37% of inquiries are knowledge questions. Customers ask about sample authenticity, sizes, and shipping costs by market.
Another 11% are loyalty questions about points balances, expiry, and redemption. Previously, Lilo Social reps opened Klaviyo daily to look up points and promo codes by hand.
How it works: brand rules loaded into AI agent guidance
Fromson and his team loaded the brand's rules around the following into the Customer Agent's guidance:
- Exact product size names
- Final-sale terms on samples
- No guaranteed shipping speeds
- Separate free-shipping thresholds for US and international orders
- A list of banned words
This way, “Customer Agent never promises something the brand can't deliver," Fromson explains. A custom skill (which lets brands automate their unique use cases) lets it check balances, process redemptions, and explain why a promo code doesn't apply, all inside the conversation.
3. Let the conversations decide what you build
For a welding and pipeline safety gear brand, my team at ECD stopped planning Customer Agent around the catalog and started planning it around the chats.
How it works: a weekly review of every conversation
Every Monday, an AI routine we built on Customer Agent APIs reads every conversation from the past week. It tags each one by topic, outcome, and buying intent. It flags where the AI agent got stuck, gave a wrong answer, or let a ready-to-buy shopper leave without a product link.
Small fixes, like knowledge snippets and skill instructions, go live the same week. Before they do, we replay real shopper questions through the AI agent in preview mode, both before and after the change. Customer Agent drafts bigger builds, like new skills and follow-up flows, for our team to approve with a one-line email reply.
The chats tell a different story than the catalog. Umbrella and shade questions were the biggest unmet need, with over 40% handed to the team. A gift loop we built from the catalog got zero mentions. The weekly loop also catches small gaps fast, like the AI agent saying a bundle didn't exist when 5 did.
The results: higher scores on test conversations and nearly 3x conversion
In the first 3 weekly cycles, 9 improvements went live, each tested on real shopper questions first. Here’s what we discovered:
- Up to 26 points better: Customer Agent's score on replayed shopper questions rose from 63% to 89% in a single cycle.
- Nearly 3x conversion where the conversations pointed us: Chats handled by the umbrella skill turned into orders at 19%, compared to 7% for chats overall.
- 67–78% resolved: That's the share of finished chats the AI agent resolves on its own, without the human team getting involved.
4. Book and cancel classes in-chat
A boutique fitness studio runs every class, credit pack, and membership in their booking platform, Mariana Tek. Customer Agent couldn't reach it out of the box, and Klaviyo couldn't see it either. The team at Andzen built a custom middleware that connects the booking platform with Klaviyo, providing 3 years of data along with continued real-time updates on every member.
How it works: booking, canceling, and live member data
Signed-in members can see the live schedule, check their credits, review upcoming bookings, and book or cancel a class without leaving the chat. Behind the scenes, Andzen backfilled 3 years of history into Klaviyo.
Now, 12 event types land within seconds: bookings, check-ins, cancellations, no-shows, orders, and membership changes. Credits, membership status, and attendance sit on every profile.
The studio's rules live in Customer Agent's guidance, so every action follows them. For example:
- Customer Agent repeats the class, date, time, and instructor, then "waits for a clear yes before acting," Smith says.
- Inside the 12-hour window, Customer Agent explains the late-cancel fee and asks again.
- Membership cancellations and freezes go to the memberships inbox.
- Customer Agent "only ever acts on the signed-in member's own bookings," Smith says.
That shared view unlocks flows the studio couldn't run before, like first class attended, no-show follow-up, and credits about to expire.
Customer Agent only ever acts on the signed-in members own bookings.
GM, North America, Andzen
The results: 679 bookings in two weeks and 100% pass rate on test scenarios
- 679 class bookings were attributed to Klaviyo in the first two weeks of the data connection, up from zero before.
- About 200,000 events were backfilled across 3 years, and about 2,000 studio members were added as profiles.
- 15 of 15 Customer Agent test scenarios passed in the live account. They covered booking, late cancels, other members' bookings, memberships, privacy, and brand voice, with zero member bookings changed.
679 class bookings were attributed to Klaviyo in the first two weeks of the data connection, up from zero before.
5. Handle subscription changes before the next charge
A DTC pet food brand sells freeze-dried and kibble dog food. A large share of their revenue comes from orders to their
Customers may want to:
- Skip or delay a shipment because the bag isn't empty yet.
- Move their next order up.
- Change their flavor or bag size.
- Update their address.
- Ask why a discount didn't apply.
"If nobody answers before the next charge, the order ships anyway," Fromson says. That can lead to a refund request or a cancellation.
How it works: shipment changes and saved conversations
Lilo Social pairs Customer Agent with a custom Claude skill connected to the brand's subscription platform. In the conversation, the AI agent checks subscription status, next charge date, and order history. It can then handle several changes:
- Skipping, delaying, moving up, or changing the frequency of the next shipment
- Swapping products or bag sizes
- Telling the customer the new ship date and price before confirming anything
When a customer asks to cancel, Customer Agent asks why first. Then, it changes its response depending on the reason:
- Overstocked: It offers a longer interval or a skipped shipment.
- Dog doesn't like the food: It suggests a different protein.
- Health concern or complaint: It hands off to a person.
Customer Agent follows the brand's rules. It makes no nutrition or health claims beyond what's on the product page. It keeps product line names separate. And it describes subscription offers as redeemed in the subscription portal, never "at check-out."
Because subscription and service data sit in the same Klaviyo account as marketing, each outcome can update the customer's profile. A "too much food" save can trigger a reminder before the next shipment. A retained subscriber who switches flavors can get a check-in a few weeks later. Customers with an open support ticket don’t receive promotional sends until it's resolved.
Because subscription and service data sit in the same Klaviyo account as marketing, each outcome can update the customer's profile. A "too much food" save can trigger a reminder before the next shipment.
6. Assess damage claims, then feed the results back to marketing
About 15% of the fragrance brand's inquiries are damage, wrong-item, or warrantee claims. Fromson says these inquiries cost the team the most to handle, and customers sometimes submit AI-generated photos to get refunds.
How it works: profile risk checks before any refund
Customer Agent collects the order number and photos, then checks profile signals:
- Prior claims and claim rate
- Order value
- Account age
- Whether the customer purchased insured shipping
- The carrier's delivery scan
Customer Agent resolves first-time, low-risk claims automatically under the published policy. It hands off repeat or suspicious claims to a person, summarizing red flags. If a request isn’t eligible for a refund, Customer Agent explains the guarantee options. High-trust VIPs skip the queue.
The customer service data lives in the same Klaviyo account as marketing, so Lilo Social holds promotional sends while a customer has an open complaint. Once it's resolved, the customer moves into a recovery flow. A positive CSAT rating triggers a review request.
7. Follow up after the chat ends
Customer Agent creates events in Klaviyo: a conversation started, a product recommended, a chat handed to the team. For the welding and pipeline safety gear brand, we use them as flow triggers.
How it works: chat events as flow triggers
- "Chatted, didn't buy": The flow sends a recap with a link to what the shopper asked about, then the questions other welders ask most. It exits as soon as they order.
- "Got a recommendation, didn't buy": The flow sends the exact products Customer Agent suggested.
- "Handed to the team, then quiet": The flow sends a service-first check-in 3 days later with tracking, returns, and warranty help.
Each email has an A/B variant ready. Many shoppers buy a recommended product days later, and a timely follow-up meets them right there.
The results: more shoppers in flows and 32% more revenue
Measured over a rolling 30 days, within two weeks of the flows going live, here’s what we found:
- 75% more shoppers in follow-up flows: The count went from 28 to 49, and every identified chatter now qualifies, up from fewer than 10% before.
- 32% more revenue from follow-up flows: Customer Agent follow-up flows also drove 3x the orders.
- 19% lift in chat-to-purchase rate: The share of chatters who buy within 7 days went from 5.9% to 7.0%.
- First-week conversion: The "chatted, didn't buy" flow converted in its first week live.
- About 9 in 10 purchases after the chat ends: That's for Customer Agent-recommended products, which shows how far the agent's influence reaches beyond the conversation.
ECD Digital saw 32% more revenue from follow-up flow: Customer Agent follow-up flows also drove 3x the orders.
Start with your own chats
Pick the one conversation type your team answers most, write down the rules, and test Customer Agent on real shopper questions before it goes live. Then read what customers actually say and let that guide your next build.
