How to Automate Customer Support (Without Losing Your Best People)

How to automate customer support without the public failure or the quiet exodus of your best agents comes down to one choice most companies get wrong: pace. The companies that got burned in 2024 and 2025 went from "let's try AI" to "AI is handling tickets" in a single step. The ones it's working for treated 80% automation as a place you arrive, months in, with your own data in hand.

If you lead a support team right now, you're fielding pressure from at least four directions. Your CFO has seen the savings claims (Klarna estimated $40 million of profit improvement for 2024 in its own press release). Your CEO read the Klarna headline and wants to know why you're not doing that. Your best agents read the same headline and are quietly updating their CVs. And the vendors are circling, promising it'll all be fine once you sign.

What makes it harder is that the people with the strongest opinions usually have something to sell. The optimists want your budget. The pessimists want your clicks. Neither has to clean up the mess if they're wrong. I went looking for practitioner views on Reddit while updating this guide. In one thread asking whether anyone had automated support completely and kept quality high, six of the nine substantive replies came from people who work for the vendor they were recommending.

Every vendor demo you've ever seen used cherry-picked tickets. The "where is my order" query with a single tracking number, or the password reset with no complications.

When you're evaluating vendors, ask them this: "Show me the 10% of tickets your AI handles worst. Show me escalation rates from your live deployments, not pilots. Show me what happens when the customer is already angry from a previous failed attempt."

Watch how they respond. The honest ones will have answers. The others will pivot back to the demo.

This playbook is the slower, more boring path. It works.

What went wrong, and what happened next#

Three companies, three different failure modes. All three have had another chapter since the first version of this guide, and the chapters matter more than the failures.

Klarna optimised for cost, then bought the humans back#

In February 2024 Klarna announced that its AI assistant had handled 2.3 million conversations in its first month, two-thirds of its customer service chats, and was doing "the equivalent work of 700 full-time agents". The CEO was everywhere. Hiring was frozen for about a year and headcount fell roughly 22% to 3,500, mostly through attrition.

Then in May 2025 Sebastian Siemiatkowski told Bloomberg the company was recruiting human agents again. His explanation, as quoted at the time: "As cost unfortunately seems to have been a too predominant evaluation factor when organizing this, what you end up having is lower quality." And: "I just think it's so critical that you are clear to your customer that there will always be a human if you want."

The headlines skipped the next part. The AI never stopped working. Klarna's annual report for 2025, filed in February 2026, says the assistant handled 80% of customer service chats that year, did the work of over 850 full-time agents and saved about $59 million. Those are Klarna's own chat-log numbers, so treat them as a company claim, but the direction is clear. Full-time employees went from 4,352 at the end of 2023 to 2,831 at the end of 2025. By February 2026 Siemiatkowski was describing human support as the premium tier: "If AI can do customer service, it means it's going to be the cheap customer service. We said the future of VIP experience will be the human connection."

So the technology did what it was sold to do. The metric was wrong. Cost was the goal rather than a by-product of good automation, and in my experience that's the most common failure pattern there is.

McDonald's picked the wrong channel, then came back with five restaurants#

Some channels are harder than others. Text is easier than voice. Structured requests are easier than open-ended ones. McDonald's learned this in public.

Its drive-thru voice AI, built with IBM and tested for two years at more than 100 US restaurants, was switched off by 26 July 2024 after videos of mangled orders went viral and the trade press reported "concerns about order accuracy". McDonald's own line was gentler: "While there have been successes to date, we feel there is an opportunity to explore voice ordering solutions more broadly."

A drive-thru has background noise, accents, interruptions and ambiguous requests ("make it a meal", but which meal?). The technology wasn't ready for that environment. The same models might have been fine on text-based support, where the input is clean.

What happened next is the interesting bit. In June 2026 McDonald's announced it was testing drive-thru voice AI again, this time with Google, and this time in five restaurants rather than a hundred. Two years, a different partner, a pilot a twentieth of the size. That's what the slow path looks like when a company with more than 13,000 US restaurants does it.

Air Canada: if your bot says it, you said it#

Air Canada's website chatbot told a grieving customer they could apply for a bereavement fare within 90 days of buying the ticket. That policy didn't exist. When the customer asked for the difference back, Air Canada said no. In February 2024 the British Columbia Civil Resolution Tribunal said yes and ordered the airline to pay C$812.02.

The money is trivial. The reasoning isn't. Air Canada argued it "cannot be held liable for information provided by one of its agents, servants, or representatives, including a chatbot". The tribunal member called that "a remarkable submission" and found the airline "did not take reasonable care to ensure its chatbot was accurate". If your AI says something, you said it. No audit trail, no human review, no defence.

It keeps happening. In April 2025 a support bot called "Sam" at Cursor, the AI coding tool, told a user that "Cursor is designed to work with one device per subscription as a core security feature". There was no such policy. The logouts were a side effect of a security update. Users cancelled. The co-founder posted on Reddit: "We have no such policy... Unfortunately, this is an incorrect response from a front-line AI support bot." Cursor refunded people and now labels AI-generated replies. A policy hallucination in a text channel with a technical audience, fourteen months after Air Canada. The lesson hadn't travelled.

What the three have in common#

Each skipped the training phase. They went from "let's try AI" to "AI is handling tickets" without the middle step where humans review what the AI writes and you find out how accurate it is.

Commonwealth Bank of Australia did a version of the same thing with its staffing. In July 2025 it made 45 customer service roles redundant on the strength of a new voice bot. The following month it reversed the decision, saying its initial assessment "did not adequately consider all relevant business considerations and this error meant the roles were not redundant". The union's version: call volumes were rising, staff were being offered overtime and team leaders were answering phones.

Gartner now predicts that by 2027 half of the organisations that expected to significantly reduce their customer service workforce will abandon those plans, and that none of the Fortune 500 will have fully eliminated human customer service by 2028. Kathy Ross of Gartner, quoted in both: "While AI excels at handling routine and well-defined problems, it often struggles with exceptions and high-risk scenarios."

And "escalate to human" is not a safety net. If 40% of your AI conversations end up escalating, you haven't automated anything. You've added a step. The customer explains themselves twice and your agents get half-finished tickets with the context missing. Customers already expect this. In a Liveops survey of 1,000 US shoppers after the 2025 holiday season, 78% had used AI for customer service and 55% had needed to escalate an AI-handled issue to a person. (Liveops sells human agents, so read it as a vendor survey. The escalation figure still matches what I see.)

What's a healthy escalation rate? It depends on your ticket mix, but from the implementations I've been close to, well-run Tier 1 automation sits under 15%. Above 25%, something's wrong with the AI's scope, training or confidence calibration. Above 40%, you're making things worse than having no AI at all. Those are my numbers, not a study's.

When your agents see these stories, they don't think "that won't happen here". They think "I'm next". Given what they're reading, that's rational.

The short version of all three, with the sequel:

If you've been hesitant about AI, that's prudence, and the evidence above backs you up. AI works for support. The question is whether you can implement it without the cost trap or the liability gap, and you can. It just takes longer than the vendors suggest.

One framework: destination, journey, grey zone, constant#

Up to 80% of tickets can be automated. But 80% is where you arrive, not where you start.

The destination: up to 80%, and who says so#

Two sources put the ceiling around 80%, and neither is a vendor blog. Gartner predicts that agentic AI will autonomously resolve 80% of common customer service issues by 2029. Klarna reports its assistant handled 80% of chats in 2025, two years after launch and after it had hired humans back. Both are mature implementations, and neither is a starting point.

The other side of that figure: Gartner's survey of 5,728 customers found that only 14% of customer service issues get fully resolved in self-service today, and just 36% of the ones customers themselves describe as "very simple". Most support operations have barely started. The technology exists. The patience usually doesn't.

The journey#

AI handles volume while your team reviews every outcome, at first. Your agents become trainers, feeding back corrections that improve the system.

This takes weeks to months. That's how long it takes to build a system your team trusts enough to stop reviewing.

(Side note: "weeks to months" drives executives mad. They want a number. Resist. Promising six weeks and hitting ten destroys trust faster than saying "we'll know more after the first month".)

And the 80% moves. Audit it quarterly: the metrics tell you which tickets you can automate, and they keep telling you as the product and the customers change.

The grey zone#

The 80/20 framing suggests a clean split. Reality is messier.

Some tickets are 60% automatable: the AI can gather the information and draft a response, but a person needs to approve it. Some tickets start automatable and shift mid-conversation when the customer reveals complexity. "Where's my order?" is automatable until the customer adds "...and I need it for my daughter's birthday tomorrow and if it's late I'm disputing the charge."

The constant: the 20% that stays human#

Some tickets should never be automated: high emotion, VIP accounts where the relationship drives revenue, legal or compliance exposure, edge cases that fit no pattern. The audit table in Phase 1 has the full row. This is where your best agents shine, and protecting it means routing rules that keep this work away from the AI, plus recognition that rewards a saved account as loudly as a fast queue.

How to automate customer support: the week-by-week roadmap#

Timelines assume a team of 10-20 agents handling 2,000 or more tickets a month. Scale proportionally. Smaller teams move faster. Larger teams need more coordination time.

Phase 1: audit (weeks 1-2)#

Before automating anything, understand what you're working with. Pull your last 1,000 tickets and sort each one into a category:

In my experience most teams find 60-70% of tickets land in the first two rows. If you're seeing less than 50%, either your ticket mix isn't ready for much automation yet or your categorisation is too conservative. Or you have a complex product. Some businesses are mostly edge cases, and that isn't a failure. It just means your ceiling is lower.

While you're auditing, map agent strengths to the human-only work. Who de-escalates best? Who handles the VIPs? Who knows the product deeply enough for the edge cases? This becomes their future focus, and you'll want the list for the conversation later in this guide.

This is the phase most companies rush or skip. Don't.

Phase 2: the fully reviewed pilot (weeks 3-6)#

Start the AI on your Tier 1 tickets, the "automatable now" row from your audit.

Humans review 100% of AI responses before they're sent. Not after. Before.

Yes, this is slower than full automation, and that's the point. You're calibrating. One of the few non-vendor voices in that Reddit thread put it well: "start by having the system work side by side with humans that will provide feedback for improvement, meanwhile gathering metrics to understand its performance, before letting it loose to respond to customers."

Week three, Friday, 4.45pm. An agent opens the queue and sees twelve drafted replies waiting for approval. Nine are fine and go out with a click. Two need a sentence changed. One has confidently quoted a returns window you retired in March. That last one is the whole reason for this phase. It cost you thirty seconds instead of a tribunal.

What you're calibrating#

Start with confidence thresholds. Vendors give you confidence scores, but they're calibrated on their training data rather than yours. A "95% confidence" response might be wrong 30% of the time on your edge cases. Track confidence against actual accuracy for your own tickets. After a few hundred reviews (I'd want 500 before trusting it) you'll know what score actually means "safe to send without review" in your operation. Hay.chat ships with three tiers you can tune: at 0.8 and above the reply goes out, between 0.5 and 0.8 it rechecks against more of your documents before answering, and below 0.5 it hands the conversation to a person. Approval mode holds every AI reply for your sign-off until you decide to loosen it, per agent or across the whole workspace.

Then category accuracy. Some ticket types will perform better than others. "Where's my order" might hit 95%. "Why was I charged twice" might hit 70%. You need this per category to know where to expand next.

And failure patterns. When the AI fails, how does it fail? Wrong answer? Right answer, wrong tone? Hallucinated policy? Each failure type has a different fix, and only the third one ends up in front of a tribunal.

What to log#

This is your audit trail. It protects you legally (see Air Canada) and operationally: when something goes wrong, you need to trace exactly what happened.

Most platforms give you some of this. Reviewer ratings with a comment, recorded against the reviewer and exportable, are the minimum. If a vendor can't show you where this log lives, that's your answer about the vendor.

What will happen that nobody warns you about#

Your agents will test the AI to break it. They'll send the weirdest tickets they can find just to watch the AI fail. It looks like sabotage and it's healthy. I've seen it on every implementation I've been near, and those agents become your best trainers, because they're finding the gaps before customers do. Encourage it.

Training it on your top performers backfires. Your best agents break rules with good judgement. They offer refunds outside policy because they sense a churn risk. Train the AI on their responses and it learns the rule-breaking without the judgement, and you get an AI that gives refunds to everyone. I mix training data at about 60% solid-average performers to 40% top performers, with one caveat: it depends how far your top performers stray from the script. Mostly by the book with occasional judgement calls, weight them higher. Creative improvisers, weight them lower.

Metrics to track#

Reduce the review rate gradually as accuracy improves: 100% to 75% to 50% to 25% to spot-checks. Let the accuracy data set the pace rather than the calendar. For the numbers that matter once the AI is live, our post on customer service KPIs that predict churn covers which ones to watch.

If you'd like to run this phase without building the log yourself, start a 30-day free trial with approval mode switched on. No credit card, and every reply waits for you.

Phase 3: progressive automation (month 2 onwards)#

Expand to Tier 2 tickets. Same pattern: high review rates at first, easing as confidence builds.

This is also when you introduce outcome-based QA. Start measuring whether problems get solved rather than whether replies follow the format. Zendesk's 2026 CX Trends survey of around 11,000 consumers, leaders and agents found that 85% of CX leaders believe customers will drop a brand over an unresolved issue, even on the first contact. First contact resolution is the number that catches it, and the Support QA Playbook has the rubric.

Celebrate human wins publicly. When an agent saves a churning account or lands a messy escalation, make sure the team hears about it. The signal you're sending: human judgement is the scarce resource now.

The AI will make a serious mistake during this phase. A wrong refund amount. A policy that doesn't exist. Something phrased in a way that reads as offensive. Forrester's Kate Leggett predicts that in 2026 "service quality will dip as companies wrestle with the complexity of AI deployment", and she's describing the companies doing it carefully. When it happens:

1. Pause automation on that ticket type immediately.
2. Pull every similar ticket from the last 24-48 hours.
3. Have humans review them for damage.
4. Contact affected customers before they contact you, where it's warranted.
5. Retrain on the failure case.
6. Resume with 100% review until accuracy recovers.

Having this written down in advance prevents panic decisions like "shut down all AI forever". It also prevents the opposite one, which is pretending it didn't happen.

Phase 4: the destination (month 3 onwards)#

Repetitive tickets fully automated. The human team on high-value work. Continuous improvement driven by QA across every conversation, with the escalation rules written down and reviewed each quarter.

Reaching this takes time. But you'll have the thing Klarna had to buy back: a team that trusts the system because they calibrated it.

What this means for your team#

This is the conversation most support leaders avoid.

AI is going to change what your agents do. The team itself stays, and the evidence for that is now reasonably solid. Gartner's March 2025 poll of 163 service leaders found 95% plan to keep human agents, and Forrester expects 30% of enterprises to create parallel AI functions in 2026, with "managers to onboard and coach AI agents, operational teams to optimize AI performance, and specialists to 'unblock' AI when it falters". The job description is what changes.

Stop trying to reassure them it won't. Their concern is reasonable. They've seen the headlines. They watched Klarna. When you say "AI won't affect your job", it sounds like what managers say right before layoffs, and it destroys trust.

Say this instead:

"A year from now, you won't be processing password resets and tracking updates. That work will be automated. You'll be doing the work that actually needs you: complex problem-solving, de-escalation, relationships, training the AI to be better. Some of you will move into roles that don't exist yet. The transition won't be instant, and we'll work it out together."

For some agents this is good news, because the copy-paste tickets aren't why they got into customer service. Others genuinely prefer the predictability of routine work. Either way it needs new skills, and not everyone will want to make the shift.

Three roles that didn't exist two years ago#

Being honest about who won't transition#

Some agents are excellent at repetitive, high-volume work. They're fast and accurate. They don't want complex emotional conversations and they don't want to analyse data.

These agents have fewer options in an AI-augmented team. The honest conversation is: "The work you're best at is the work most likely to be automated. Here's what the new roles look like. Do any of these interest you? If not, let's talk about what makes sense."

This is hard. It's also more respectful than pretending the change isn't coming. Commonwealth Bank made 45 people redundant in July 2025 and had to apologise to them in August. The order of operations matters.

How to spot your future AI trainers#

Look for agents who:

• Catch edge cases others miss
• Ask "why does the system work this way?" rather than just "what do I do?"
• Get frustrated by inconsistency ("why did we handle this differently last week?")
• Document their workarounds
• Are curious about how things work, rather than only how to use them

These are often not your top performers on volume. They may be slower because they're thinking more carefully. That careful thinking is exactly what AI training needs.

Where to go from here#

If you've read this far, you're taking it seriously. Three things you can do this week:

1. Run the audit. Pull the 1,000 tickets. It takes a couple of days, and you'll know more about your operation afterwards than any vendor demo will tell you.
2. Read Safe AI Customer Service Automation: 5 Patterns You Can Defend to Legal, the guardrails this guide assumes, written for the conversation with your legal and security teams.
3. Read the Support QA Playbook for the outcomes rubric that Phase 3 depends on.

And if you'd rather run the pilot on a platform that already holds every AI reply for approval, Hay.chat does that out of the box. Thirty days free, no card.