Hey team,
We are back in the US after a week of gallivanting through Sicily and Malta, which was beautiful, chaotic, and a helpful reminder that vacation with kids is mostly parenting with extra carbs.
Feeling inspired, slightly tired, and very ready for childcare so I can catch a breather.
Over the last few weeks, I’ve gotten some version of the same question a few times:
Where should I use AI in CX, and where should I avoid?
Most teams start with “what can we automate?” which sounds practical until order tracking, refund exceptions, angry VIP customers, damaged birthday gifts, and cancellation saves all get dumped into the same giant bucket called “support.”
That is how brands get into trouble. AI is great at reading the pile, spotting patterns, pulling context, routing work, and drafting the boring stuff. The wheels come off when it starts making promises, denying refunds, apologizing to angry customers, or deciding whether a long-time customer deserves an exception.
BFCM makes this worse because volume hides bad judgment. A weak AI resolution in August is annoying. The same weak resolution in November gets multiplied across thousands of tickets while everyone is staring at response time, deflection, and queue size like those numbers tell the whole story.
They don’t.
So this week: five CX jobs AI should probably own already, and five I’d keep it far away from for now.
Let’s get into it.
This week’s newsie is brought to you by Kim.cc
BFCM doesn’t create new support problems. It exposes the ones you already had.
The tickets your AI half-resolves in October come back at 2-3x volume in November, on the most expensive week of your year.
A closed ticket isn’t necessarily a resolved ticket, and when the first answer doesn’t land, the brand pays twice: once for the AI interaction, again for the human cleanup. Tickets closed by AI carry an average re-open rate of 50%+. Going into peak, that’s a bill you haven’t seen yet.
Here is an interesting approach that 200+ Shopify brands are taking for this BFCM season: Kim.cc. An AI-native customer support service for ecommerce.
AI handles the volume and Sentinels, trained CX, QA, and prompt experts, own the outcome: they implement the system, step in where judgment is needed, and teach the AI how to handle the hard ones via proper prompt engineering so the same ticket doesn’t come back next time.
This way more issues are resolved the first time, fewer customers coming back (because suboptimal resolutions received by standard AI tools in the market), and an operation that gets sharper as volume climbs instead of buckling under it.
This BFCM brands should really try to look at the problem differently, i.e, the number that matters isn’t how many tickets close. It’s how many stay closed.
You can now book a pre-BFCM re-open rate teardown. This team is pulling a sample of your AI-resolved tickets and show you your real resolution rate — and where peak volume will hurt — before the rush hits.
You can book a 20-min call with their CX solutions expert to get your analysis done:
First, stop worshipping deflection rate
Deflection rate is one of those metrics that looks great in a board deck and gets much weirder the second you read the actual conversations.
Leadership loves it because it is clean “cost savings.” A ticket did not reach a human, the customer got some kind of answer, and the dashboard moved in the right direction. Everyone gets to feel like the operation got more efficient, which is a fun feeling until you remember the customer may have hated every second of it.
A ticket can close cleanly while the customer decides they are done with you.
I remember reading a story about Air Canada slop a few years back. Its chatbot gave a grieving customer bad information about bereavement fares. The company tried to argue the chatbot was responsible for its own actions, which is an incredible attempt at putting a name tag on the Roomba and sending it to court. The tribunal did not go for it.
Then there was the Chevrolet dealership chatbot that got prompt-injected into agreeing to a 2024 Tahoe for $1. Obviously nobody got a Tahoe for a dollar, but the point is still useful. The bot was allowed to speak in a place where the words had financial and reputational weight, and nobody had built a strong enough fence around what it was allowed to say.
AI should help your team see more, move faster, and spend less time doing soul-crushing admin work. It should not be handed the moments where money, emotion, judgment, or public embarrassment are on the table.
1. AI should tag and categorize tickets
This is boring, which is usually a good sign.
Most helpdesks have tags that look like they were built by five teams, two agencies, one intern, and someone who left the company in 2021 but apparently had very strong opinions about capitalization.
AI should clean that up. It should tag tickets by issue type, product, urgency, sentiment, customer type, subscription status, and root cause. It should turn the inbox from “general chaos with search functionality” into something your team can learn from.
No customer sees this or gets offended. Nobody screenshots your bot being weird. Your team just gets better data and fewer tags named “misc-urgent-final-v2.”
Good AI job.
2. AI should summarize messy ticket history
A customer writes in after three previous conversations, two replacements, one refund, and a subscription change nobody fully understands.
The agent opens the ticket and gets to choose between spending seven minutes spelunking through the helpdesk or answering with the confidence of someone who absolutely did not read the file.
AI should solve that.
Give the agent the customer history, previous issues, order timeline, subscription status, prior concessions, and the reason this ticket probably matters. Do not make the human dig through fourteen messages to learn that the brand already messed this up twice.
This is one of the best uses of AI because it makes the rep sharper before they reply. The customer gets a better answer, the agent gets the full picture, and nobody has to pretend that opening six tabs is a customer experience strategy.
3. AI should find patterns across tickets
This is the one I wish more brands cared about.
If ten customers mention leaking caps, someone should know. If five customers say the new formula tastes weird, someone should know. If subscription customers keep saying they have too much product, someone should know before the next retention meeting where everyone pretends churn is mysterious.
AI should be reading across tickets, reviews, surveys, and social comments to flag patterns your team might miss while they are buried in the queue. Broken zippers, melted product, weird smell, missing scoop, late shipments from the same warehouse, confusing onboarding, broken discount code, whatever.
A good CX person can catch some of this manually. They should not have to rely on memory and vibes while 300 tickets are waiting.
Let AI find the smoke and put it in front of the people who know what it means.
4. AI should route work to the right place
A lot of bad CX happens because the ticket takes a scenic tour through the wrong queue before anyone useful sees it.
AI should route billing issues to billing, product-quality issues to the right escalation path, angry subscription issues to the retention team, VIP issues to someone who understands why they are VIP, and low-risk questions to automation.
This will not make a sexy LinkedIn post. It will make the operation better, which is usually more useful.
Good routing means faster replies, cleaner ownership, fewer missed escalations, and less “who owns this?” theater in Slack. It also means your best people spend more time on the tickets that actually need them, instead of being human traffic cops for a helpdesk that looks like it was configured during a fire drill.
5. AI should draft low-risk replies
Drafting is great. Sending is where brands start lighting candles around the helpdesk.
AI should draft the boring replies: order status, return instructions, product FAQs, warranty eligibility, subscription instructions, basic policy explanations. Nobody needs to handwrite “you can start your return here” for the 4,000th time. That is how good agents slowly become houseplants.
The risk level should decide how far the draft goes.
If the customer is calm, the issue is simple, and the answer is low-stakes, let AI help. If the message includes a refund denial, an apology, a dollar amount, a health concern, a damaged gift, a public complaint, or someone who is clearly already angry, do not let the bot auto-send its little masterpiece.
That is how you create screenshots.
Now, the five things I’d keep far away from AI
The simplest rule I’d use: if the moment requires judgment, keep a human close.
1. Refund and return exceptions
If the policy cleanly says yes, automate it.
If the customer is asking for an exception, do not let the bot freestyle.
The policy might say no while the context says obviously yes. A customer outside the return window might be unreasonable, or they might be a great customer with one weird issue your policy did not account for. They might be technically wrong and still worth taking care of.
AI can show the order history, LTV, prior issues, and suggested options. Great. The actual exception should belong to a person until you are extremely confident the logic reflects how your best CX people would make the call.
Most brands are not there. They just have a refund threshold someone invented in a meeting.
2. High-LTV customers
Your top 1% customers should not get the same bot treatment as someone who bought once with a 40% off code and is now trying to return a final-sale item from the Obama administration.
This sounds obvious, and yet plenty of automations do exactly that because the ticket type is the same.
AI should identify the VIP. It should surface customer value, order history, subscription status, prior issues, and anything else that changes how the brand should handle the moment.
It should not quietly decide that your best customer deserves the cheapest workflow because the workflow technically works.
3. Angry customers
AI is very bad at knowing when it is making the situation worse.
It keeps the same tone while the customer gets more annoyed. It apologizes in that weird smooth way where every sentence sounds like it was dipped in customer-service lotion. It says “I understand how frustrating this must be” and then sends a help-center link, which is how brands create screenshots with captions.
Angry customers do not need perfect grammar. They need to feel like someone competent is actually handling the issue.
AI can summarize the problem and suggest next steps. A human should own the reply.
4. Product-quality judgment
AI should flag product-quality issues. It should not be the final judge of whether they matter.
A bot sees the ticket in front of it. A good CX person remembers the last ten and knows when the pattern feels too familiar to ignore.
Let AI group the signals. Let humans decide whether product, ops, quality, or leadership needs to hear about it before Reddit does.
5. Generic cancellation saves
Cancellation is where brands love pretending every problem is a discount problem.
Too much product gets 20% off. Too expensive gets 20% off. Did not see results gets 20% off. Going on vacation somehow also gets 20% off, because apparently every subscription issue can be solved with the same sad coupon.
Too much product needs a pause, skip, or better frequency. Price might need a smaller bundle. Product confusion needs education. Poor fit needs a swap. Travel needs a delay. A customer who is actually done needs a clean cancellation, because making cancellation painful is how you turn “maybe later” into “never again.”
AI can help here when it is designed around the reason, the customer context, and the right save path. It should not treat every cancellation like a discount negotiation with a pulse.
The actual rule
The Extinguisher Rule is the simplest way I can think about this.
AI can help you find the fire. It can read more tickets than your team can read, spot the pattern, summarize the context, route the issue, draft the boring reply, and make the human faster.
When the customer is standing there watching to see if the brand actually cares, keep a person holding the extinguisher.
Not because humans are inherently always magical. Plenty of humans are also bad at this. But your best CX people understand context in a way most AI workflows still do not. They know when the policy answer is technically correct and still expensive. They know when the customer is asking for information versus fairness. They know when a ticket is the first visible sign of something upstream breaking.
Before you buy another AI tool or expand the one you already have, run a simple audit. Pull the last 20 tickets AI resolved end to end. Read them like a customer would. Ask whether you would be comfortable with each exchange being posted publicly. Check whether any of them involved an apology, refund denial, dollar amount, angry customer, VIP, health concern, damaged gift, or product-quality issue.
If more than two make you wince, your deflection rate has probably been eating your judgment while everyone admired the silly little dashboard.
Find the fire with AI. Hand the extinguisher to a person.
That’s it for this week!
Any topics you’d like to see me cover in the future?
Just shoot me a DM or an email!
Cheers,
Eli 💛
