Straight answer: AI customer service replies are getting faster and more polished, but customers can spot them, and a spotted AI reply does more damage to repeat business than a slow, slightly clumsy human one ever did. The fix isn’t ditching AI, it’s using it to draft and a real person to finish, every single time, with no exceptions for “simple” tickets.
The one-star review that made me look at this
A client of mine runs a skincare brand doing around £40,000 a month on Shopify. Small team, three people, all working from home, all stretched thin. Last year she plugged Gorgias into her store and set up AI-drafted replies for refunds, shipping delays and product questions. Response time went from four hours to under a minute. She was thrilled. I was thrilled for her, for about six weeks.
Then a customer left a one-star review that had nothing to do with the product. It said, word for word: “Asked about my delayed order and got a reply that was clearly written by a bot. Felt like nobody cared, just wanted me to go away quietly. Won’t be ordering again.” The order itself arrived two days later, fine, no damage done there. But that customer had bought three times before, average order value around £68, and her lifetime value to that point was roughly £340. Gone, over a reply that took forty seconds to generate and read like every other AI customer service reply on the internet.
That’s the bit nobody selling you AI customer service tools tells you upfront: speed was never the thing customers were complaining about. Slow replies annoy people. Robotic replies make people leave.
The number that should worry you more than response time does
Everyone tracks response time because it’s easy to measure and easy to brag about. Four hours down to forty seconds looks brilliant on a dashboard. What almost nobody tracks is repeat purchase rate broken down by whether the customer’s last support interaction was human-edited or sent as-is from the AI draft.
We started tracking it for three of my consulting clients last year. Rough pattern across all three, small sample, but consistent enough to take seriously: customers whose support tickets were resolved with a human-edited reply repurchased within 90 days at roughly 22 to 28 percent. Customers whose tickets got the raw AI draft, no edits, sat at 11 to 14 percent. Half. The AI replies were factually correct in almost every case. They were just not felt as care, they were felt as processing.
If you’re running customer service the way a lot of small businesses now do, chasing the response-time number and quietly letting the AI drafts go out untouched because everyone’s busy, you might be optimising the exact metric that’s costing you the money.
Why customers can tell, even when the grammar is perfect
I’ve read hundreds of these replies now, across ecommerce, coaching businesses, agencies. There are tells, and once you see them you can’t unsee them:
- “I completely understand your frustration” appearing in a reply about a two-day shipping delay, which is not a frustrating problem for most people
- Three-part structure every time: acknowledge, apologise, resolve, in that exact order, with no variation in tone from ticket to ticket
- Replies that answer the question asked but never the question underneath it, like a customer asking “will this arrive before my daughter’s birthday” getting a shipping estimate and nothing about the birthday
- Perfect punctuation and zero personality, when the brand’s own marketing, on Instagram or in email, has plenty of both
That last one is the real giveaway. If your Instagram voice is warm and specific and your support replies read like a compliance document, customers notice the gap even if they can’t name it. Airbnb worked this out years ago with their host response guidelines, they push hosts hard toward personal, specific replies because a generic one erodes trust in a way a slightly late one doesn’t. Worth looking at how they built that expectation into their whole platform in the Airbnb marketing strategy breakdown I did, it’s the same principle just applied to a different channel.
The two-line fix I now give every client
This isn’t complicated and it isn’t expensive. Here’s what I tell people to do, step by step:
- Keep the AI draft, don’t throw the tool out, it saves time on the boring 80 percent of tickets like “where’s my order”
- Before anything sends, one real person reads it and does two things only: adds one specific, human detail (the customer’s name used naturally, a reference to what they bought, an acknowledgment of the specific situation), and deletes one stock phrase (“I understand your frustration”, “Thank you for reaching out”, “I hope this finds you well”)
- Cap the edit time at 30 seconds per ticket, this is not a rewrite, it’s a five-word fix
- For anything involving a complaint, a refund over £50, or a customer who’s bought more than twice, the reply gets written by a person from a blank page, no AI draft at all
- Once a week, pull five random sent replies and read them out loud, if they sound like they could have gone to anyone, they’re not ready
My skincare client put this in place in under a week. No new software, no extra headcount, just a rule that AI drafts get a human pass before sending. Three months later her repeat purchase rate for support-touched customers had moved from around 13 percent back up to 24 percent. Same AI tool, same team, one rule changed.
Where AI earns its place in customer service
I’m not anti-AI here, I use it every day and I’d never tell a small business to switch it off. It’s brilliant at the parts customers don’t emotionally invest in: order status, returns policy lookups, sizing charts, opening hours, tracking numbers. Nobody feels unloved by an automated tracking link. The mistake is using the same tool, the same tone, for the tickets where a person is upset, confused, or spending real money based on trusting you.
If you’re mapping out where automation belongs in your business more broadly, not just support, the same logic applies to email. A brand’s Mailchimp marketing strategy can automate welcome sequences and abandoned cart nudges perfectly well because those are low-stakes, expected, transactional. The moment a customer emails back with a real question or complaint, that’s not a job for the automation, that’s a job for a person, ideally quickly, using the AI draft as a starting point rather than a finished product.
What this means if you’re running this alone from home
This is where it gets uncomfortable for a lot of solo owners and tiny teams, because I know exactly what it’s like to be the only person answering support emails at 9pm because that’s when you finally got to your inbox. The temptation to let the AI reply go out untouched isn’t laziness, it’s survival. You’re tired, there are forty tickets, and the draft looks fine.
But here’s the uncomfortable bit most people don’t want to hear: if you don’t have capacity to add that 30-second human touch to every reply, you have too many tickets for one person and no automation, however clever, fixes that. It masks a staffing problem rather than solving it. I’ve watched founders convince themselves that a good AI setup means they can keep running support solo indefinitely, when the volume passed “one person, part-time” months ago and the AI replies are just hiding the strain from the founder, not from the customers.
If that’s you, the honest move is working out the real hourly cost of a part-time support person against the lifetime value you’re losing to flat, unedited replies, not layering on more automation to cope. It’s a calculation worth doing rather than guessing, and it’s the kind of thing an AI consultant for small business can help you model in an afternoon, comparing the cost of a few hours of human editing time against what you’re currently losing in repeat purchases.
If you want to make this tangible for your own numbers rather than take my word for it, build yourself a rough version of what you’re losing. Multiply your monthly support tickets by the percentage you’d estimate go out unedited, then by the difference in repurchase rate you’d expect, then by your average order value. It’s the same kind of exercise as building an interactive calculator for your visitors, except this one is just for you, and it usually produces a number uncomfortable enough to change behaviour that week.
The bit people skip because it’s not exciting
None of this needs a bigger tool, a fancier prompt, or a new platform. It needs a rule, written down, that a human reads and edits every reply before it sends, and that the tickets which matter most never touch the AI draft at all. Alex Hormozi talks a lot about the businesses that win being the ones that do simple things consistently rather than chasing the next clever tactic, and there’s a version of that in the business lessons from Alex Hormozi that applies exactly here. This isn’t clever. It’s a 30-second habit. Most businesses won’t bother building it because it doesn’t feel like progress, it feels like admin. It’s the admin that keeps your customers coming back.
Frequently asked questions
Can customers really tell if a reply was written by AI?
Often, yes. The tells are specific phrasing patterns (“I completely understand your frustration”), a rigid acknowledge-apologise-resolve structure, and answers that address the literal question but miss the underlying concern. Customers may not consciously identify it as AI, but they register it as generic, and generic reduces trust and repeat purchases even when the information given is accurate.
Should small businesses stop using AI for customer service entirely?
No. AI is useful for low-stakes, transactional queries like order status, returns policy, and tracking numbers. The risk is using the same automated tone for complaints, refunds, and higher-value or repeat customers, where a human edit or a fully human reply protects the relationship far more than speed does.
How much time does editing an AI reply take?
In practice, about 30 seconds per ticket: adding one specific, personal detail and removing one stock phrase. It’s not a rewrite. Clients who added this single step saw repeat purchase rates for support-touched customers roughly double within three months, with no new software or headcount.
What’s a warning sign that AI customer service is masking a bigger problem?
If you’re a solo founder or tiny team and you don’t have time to add a human touch to every reply, that’s usually a sign your support volume has outgrown one person’s capacity, not a sign you need cleverer automation. In that case it’s worth costing out part-time support help against what unedited AI replies are likely costing you in lost repeat business.
Want this done for you? See AI automation for ecommerce stores.
Related reading: Why Your AI-Written Marketing Sounds Like Everyone Else’s (And How to Fix It) and Why Your AI-Written Emails Are Quietly Killing Your Reply Rate.
If you want the full breakdown, here is everything I know about AI marketing.
