AI in Customer Support: Your Trail Guide to the Summit
Exploring the asymmetry of consumer expectations for AI in customer support
The chatbot love/hate paradox
Here’s something strange. If you were to survey your customers and ask them how they feel about AI in customer service, most of them will tell you to keep it away from them. Gartner put the question to 5,728 people and found that 64% would rather companies didn’t use AI for service at all. In fact, more than half said they’d switch to a competitor over it.
Now ask those same customers how they spend their Tuesday afternoon. They’re chatting with an AI, happily, to reschedule a flight, check on a package, or summarize Amazon reviews on some new tech they’ve been scoping. Zendesk’s research pegs the share of consumers actively wanting to use AI assistants for their own service tasks at 67%.
The contradiction resolves the moment you stop treating “AI” as one thing. Customers don’t hate AI. They hate AI deployed at them instead of for them. They hate being stuck, and putting a robotic bouncer that is having trouble interpreting their request adds insult to injury.
Here’s the challenging reality: a good AI interaction is a shrug; a bad one is an exit. Seventy percent of consumers say a single bad AI experience is enough to send them shopping for a new brand. A single great one? Barely moves the loyalty needle. The gains are linear. The losses are off a cliff.
Think of AI as a shortcut up the mountain. It’s a real shortcut — real time saved that can be reinvested — but only if you take the time to build the trail carefully. If you don’t, your customers don’t get a mildly inconvenient Tuesday. They go missing.
Depending on what business you’re in, you might be thinking: my customers are younger, more tech-savvy, they’ll love it. The data has a twist for you. Younger customers are more willing to use AI and more willing to walk when it fails. Gen Z and Millennials are enthusiastic adopters — as long as they can escape to a human. For younger generations, your web experience is your storefront. The experience they have at the digital doorstep says more to them about your brand and the value you place on them than you might think. The asymmetry is amplified.
The view from the top
Before we talk about the pitfalls — and there are a few — let’s be clear about why anyone would try to summit this treacherous mountain in the first place.
A well-built AI stack at a 10-to-100-person company will reclaim roughly 10 to 20 hours a week per user. That’s not marginal. That’s a part-time employee, freed up, pulled off your most expensive headcount. The math is compelling.
Here’s the fork in the trail. Most companies, when those hours come back, pocket them as margin and reduce headcount. Run a tighter ship. Do more with less. It looks fine on a spreadsheet — right up until a competitor decides to reinvest their reclaimed hours strategically.
The alternative is to focus those hours into the moments your enterprise-scale rivals cannot possibly replicate. For example:
First purchase. A hand-signed note in the box. A short welcome video. A founder call within the first week. Amazon can’t do this. You can.
Cancellation attempt. Not a form, not a survey. A real human on the phone within the hour, authorized to listen and authorized to fix. Winning a customer back is a fraction of the cost of finding a new one — and it’s effectively impossible to automate well.
Complaint recovery. A real person acknowledging the problem within the day, with the authority to make it right. This isn’t where loyalty is tested. It’s where loyalty is built.
Onboarding. A 15-minute call in the first week that nobody asked for. Your largest competitor cannot make this call at scale. You can make it effortlessly.
The annual check-in. Once a year, a human reaches out to every meaningful customer — not to sell them anything, just to check in. The cumulative effect over three years is enormous.
In a February 2026 HBR piece, Horst Schulze (the Ritz-Carlton cofounder) and Micah Solomon argue the same thing from the enterprise side: as AI becomes commoditized, human hospitality becomes the moat. Your enterprise competitors are spending billions trying to scale intimacy — trying to make AI feel human. You don’t have that problem. Your team already is human. The game isn’t scaling intimacy — it’s creating the time to deploy it where it matters. AI is how you fund that.
AI handles the routine so your humans can handle the signature. The question isn’t how much of your support to automate. It’s which stretch each interaction lives on.
Three stretches of trail
So where does AI actually belong on your trail map? Think of it in terms of terrain.
The green zones are the flat, paved parts of the trail. Routine, deterministic, low-emotion. Order status. Returns. Password resets. Scheduling. When a customer asks when their package arrives, they want the date. They don’t want a relationship. They don’t want empathy. They want the date, instantly, at 11 p.m., in whatever language they prefer. AI is genuinely better here — not because it’s cheaper, but because it’s faster.
The yellow zones are the rough patches where the trail is overgrown or worn out. You may need a guide to show you where to step, and it needs to be carefully maintained. These are the oddities. A billing question with one eyebrow-raising detail. A return that falls in a gray zone. This is the copilot zone — AI drafts, a human reviews. AI summarizes, a human decides. Done well, this doesn’t replace anyone. It sharpens them.
The red zones are the parts where you hand your gear to the Sherpa and let the expert lead the way. This is a full handoff from the AI to the rep, but with AI-enabled tooling supporting them in the background — retrieving specific policy wording, pulling up the customer’s history, drafting a response for the rep to approve. Cancellations. Complaints. Anything emotionally loaded. The conversations that decide whether the relationship survives. When a customer starts explaining why they want to cancel, they’re not really asking for a refund — they’re asking whether you’re the kind of company worth sticking with. No amount of clever prompting makes an AI the right voice for that.
Climbing without handrails
A few climbers have already fallen off this mountain. Worth a glance before you lace up.
Air Canada, February 2024: a grieving customer asked the airline’s chatbot about bereavement fares. The bot invented a policy that didn’t exist. When the airline refused to honor it, a Canadian tribunal ordered Air Canada to pay up and flatly rejected the argument that the chatbot was “a separate legal entity.” Your bot’s hallucinations are your hallucinations. That’s a wake-up call.
Around the same time, Klarna — the buy-now-pay-later fintech — announced its AI assistant was doing the work of 700 customer service agents. Fifteen months later, CEO Sebastian Siemiatkowski quietly reversed course and started rehiring humans. His own words: “We focused too much on efficiency and cost. The result was lower quality — and that’s not sustainable.”
And DPD, the parcel carrier, had a customer coax its chatbot into writing a poem about what a terrible delivery firm DPD was. Screenshots hit 1.3 million views in a single day.
The technology worked in all three cases. What failed was intent. One was optimizing for deflection. One for cost. One for volume. None of them was optimizing for the customer in front of the screen — and customers read that signal almost instantly.
Five handrails to build before you open the trail
If you’re building the trail, build these handrails first. You can skip any of them. You’ll regret every one you skip.
1. Start with deflection, not replacement. Target the routine 60–80% of tickets. Keep humans as the default for anything emotional, regulated, or relationship-defining. This is the biggest philosophical fork in the road. Replacement logic is what sent the Klarnas of the world scrambling to rehire. Deflection logic quietly delivers the same savings without breaking the brand. The vendor landscape has largely caught up with this framing — Intercom Fin, Zendesk AI Agents, Ada, Decagon, and Sierra all position around resolution of well-scoped tickets with a clear human fallback, rather than “replace your team.”
2. Invest in the knowledge base before the bot. Audit your existing help docs, old tickets, and support macros. Consolidate. Clean. Make it retrievable. A messy knowledge base doesn’t just limit AI — it amplifies your weak spots. The bot will confidently say the wrong thing, in your voice, faster than any human ever could. Research on retrieval-grounded AI is unambiguous: the knowledge base is the ceiling.
Good news — this is also the step where AI is most useful to you. Export a year’s worth of support tickets to a CSV and hand it to Claude Projects, ChatGPT with Advanced Data Analysis, or Google’s NotebookLM. Ask it to cluster the top 20 recurring themes, flag the questions your docs don’t answer well, and rank them by volume. You’ll get a prioritized work list for your content team in an afternoon — work that used to be a month of manual tagging. This is AI as a tool for you, before you point it at your customers.
3. Design the escalation path first, not last. Most of what customers call “a bad AI experience” isn’t about a failed answer. It’s about a failed escape. They asked for a human and couldn’t find one. Define your triggers (low confidence, emotion detected, specific keywords), define how the conversation context transfers to the human, and define the SLA for the follow-up. If the escape hatch is clean, customers will forgive a lot. Most major platforms now expose handoff logic and context-passing out of the box — use them, and test them with a stopwatch.
4. Match pricing to volume shape. Stable, predictable volume? Per-seat pricing is probably fine. Seasonal or spiky? Outcome-based pricing — roughly a dollar per resolved conversation, depending on vendor — finally aligns your vendor’s incentives with yours in a way per-seat never did. Re-evaluate at every 2× growth milestone. The alignment can flip on you fast if your volume changes shape.
5. Protect the signature moments. Identify the three to five interactions that define your brand — the ones a customer would tell their friend about. Route them to humans, enhanced by AI, never replaced by it. These moments carry wildly disproportionate weight in whether a customer stays.
Better questions, better directions
The wrong question is the one everyone asks first: how much of my support can I automate? It points you at the trail. It measures success in tickets deflected and dollars saved.
The right question is quieter: which interactions define my brand, and how do I use AI to give my humans more time and sharper tools to nail those? It points you at the summit.
If you want a low-friction place to start, try this. Open ChatGPT, Claude, or Gemini and paste in the following:
Act as a customer experience strategist interviewing me. I run a [describe your business in one sentence]. I want to identify the three to five “signature moments” in my customer journey — the interactions that most shape whether someone stays a customer, recommends us, or leaves. Ask me one question at a time. Start with the customer’s first contact and walk me through the relationship chronologically. After about 10 questions, summarize what you’ve heard, propose my top signature moments, and flag which ones a competitor three times my size couldn’t replicate.
Spend 20 minutes on that conversation. Then do a second pass with your last 90 days of support emails or tickets pasted in, and ask: “Where in this data do you see the most friction, confusion, or repetition? Cluster it by theme and estimate how many of these could be handled by a well-built AI response, versus which ones signal a signature moment I’d want a human to own.”
Those two conversations, back-to-back, will give you a sharper map of your own trail than most consulting engagements produce in a month.
The companies still asking “how much can I automate” are planning their own headcount reversal. The ones asking “what do I want my humans free to do” are building something enterprise-scale rivals cannot replicate — and quietly capturing the savings anyway. The shortcut was never the point. The time at the top was.
This post was collaboratively brainstormed, researched, and drafted alongside AI (Claude Opus is our go-to lately for this sort of thing), then revised and edited heavily by real humans.
Conversint is a consulting firm helping SMBs and everyday humans unlock real value from AI and other modern tech. To learn more about our services and schedule a free initial consultation, visit our website at Conversint.ai.
If you learned something useful, consider sharing this post with someone else who may find it useful as well.
For more insights and guidance like this delivered to your inbox regularly, consider subscribing.