Customer Support
Deflect repetitive enquiries automatically, equip reps with instant answers, and cut resolution times.
Updated 4 days ago · 10 sources
Proof over hype
- Learn from thisEnterprise
Commonwealth Bank of Australia
Commonwealth Bank rolled out an AI voice bot to handle simple, routine phone queries, and told the public it had cut call volumes enough to justify cutting 45 customer service roles.
All 45 roles reinstated and the bank apologized, within about a week of the union's numbers becoming public.
So: Count call volume yourself for a full quarter after go-live, overtime and managers on phones included, before you cut a single role — a wrong internal number is what caused this, and it took a union to catch it.
The takeaway The bot wasn't necessarily the failure, the claim about it was. Nobody at the bank checked whether call volume had actually dropped before using that claim to justify cutting people.
How they did it
The bank announced the cuts in late July 2025, saying the AI bot had freed the remaining team from the simple calls. Australia's Finance Sector Union checked the actual numbers and found the opposite: call volumes had risen, not fallen, and the bank was already running remaining staff on overtime and pulling managers onto the phones to cover the gap the bot was supposed to have closed. Within about a week of the union surfacing that, the bank reversed the decision, reinstated all 45 roles, and apologized to the staff it had already let go.
- Learn from thisEnterprise
Klarna
Klarna built an AI customer service assistant and, in 2024, cut staff and paused hiring around it as part of preparing to go public.
Two-thirds of chats handled by AI at the assistant's peak, work equivalent to 853 employees claimed — but support and operations costs still rose to $50M in Q3 2025 from $42M a year earlier, once rehiring began.
So: Send anything complicated or multi-part to a person from day one, and price out what rehiring would cost before you use projected AI savings to justify cutting agent headcount.
The takeaway The AI wasn't wrong to exist, it was deployed past its competence: genuinely excellent at the repetitive two-thirds, genuinely bad at the rest, and Klarna cut headcount before drawing that line.
How they did it
The assistant ended up handling two-thirds of all customer chats, which Klarna said was doing the work of 853 full-time employees, up from 700 at the start of the year, and by most accounts it was genuinely good at what it was built for. Analyst Kate Leggett's read on why matches what Klarna itself found: 'A lot of the questions they get are highly repetitive and fairly simple, and that's a perfect candidate for AI.' The problem showed up on the other side of that split: customers with anything sophisticated or multi-part got generic answers that didn't actually resolve what they'd asked, and complaints piled up. In May 2025, CEO Sebastian Siemiatkowski reversed course publicly: 'customers should always have the option to speak with a human,' describing the new target model as more like an 'Uber type' hybrid than a pure-AI one, and Klarna started actively recruiting human agents again. The financial picture tells the same story: Klarna had claimed roughly $60 million in AI savings, but its actual customer service and operations costs still rose to $50 million in Q3 2025, up from $42 million a year earlier, once the rehiring started.
- Enterprise
Salesforce
Salesforce ran a roughly 9,000-person support team and wanted its own Agentforce AI agents to carry a real share of that volume, not just run as a pilot on the side.
Support headcount fell from about 9,000 to about 5,000; support costs down 17%; half of customer conversations now go to AI agents, half to humans.
The takeaway Benioff's own numbers don't fully add up in public: 'hundreds' redeployed against roughly 4,000 fewer heads leaves most of the reduction unexplained. That gap is worth asking about before taking any company's AI-driven headcount story at face value.
How they did it
Starting in early 2025, Salesforce put Agentforce live on its own help site and let it take routine customer questions first, with people handling whatever it couldn't close. Over six to nine months, the company says it handled more than a million customer conversations that way, cutting support costs by 17%. CEO Marc Benioff described the current split plainly: '50% are with agents, 50% are with humans,' and gave the headcount number without softening it: 'I've reduced it from 9,000 heads to about 5,000, because I need less heads.' Salesforce says it redeployed 'hundreds' of the affected employees into professional services, sales and customer success — a real number, but one that leaves a gap: out of roughly 4,000 fewer support heads, only hundreds are accounted for as redeployed, and the company hasn't said publicly what happened to the rest.
Tool used Agentforce (Salesforce)
- Enterprise
Home Depot
Customers calling a Home Depot store had to work through a phone menu, picking options and waiting, before they ever reached someone who could actually answer their question.
Reaches a resolution about four times faster than the old phone-menu system, across a 50-store pilot ahead of the company-wide rollout.
The takeaway The design choice worth copying isn't the AI itself, it's that a live associate stayed one step away the entire time. The speed gain came from skipping the menu, not from removing the option to reach a person.
How they did it
Home Depot piloted AI voice agents on store phone lines across 50 stores: instead of navigating a menu, a caller just says what they need in plain language, and the system figures out why they're calling within about 10 seconds. It handles real tasks on its own — checking an order's status, confirming whether a product is in stock, giving store hours and location, texting a product link to a customer's cart, and completing a purchase — and callers always keep a direct path to a live associate for anything the agent can't resolve. After the pilot, Home Depot committed to rolling the same system out to every US store over the following year.
- Enterprise
SiriusXM
SiriusXM had heavy chat volume from subscribers on billing, technical issues and basic account questions, and wanted a channel that could actually resolve things rather than just route them to a person.
The highest-rated and lowest-effort service channel SiriusXM runs, per Sierra — SiriusXM has not published the underlying numbers independently.
The takeaway The conversational-understanding detail, correctly routing a DJ's name to the right channel, is a better signal of quality than the 'highest-rated' claim itself, because it's specific and checkable in a way a vendor superlative isn't.
Reported by Sierra · checked Dec 2025
How they did it
SiriusXM built a chat agent called Harmony with Sierra, as one of Sierra's original design partners since its Agent OS platform launched in February 2024. Harmony handles subscription administration, technical troubleshooting like satellite signal resets, and content recommendations, and understands conversational, non-obvious requests — asking 'where can I find Howard or Andy?' gets correctly routed to Howard Stern's or Andy Cohen's channel rather than failing on a literal keyword match. Sierra says Harmony is now SiriusXM's highest-rated and lowest-effort service channel, ahead of the company's other support channels on customer satisfaction, resolution and ease of use. The two companies have since added what Sierra calls an Agent Data Platform, a memory layer that lets Harmony remember a subscriber's history across conversations instead of treating every chat as a one-off, with SiriusXM as the platform's first adopter.
Tool used Sierra
Try this today
Group last month's tickets by real cause
20 minYou see the handful of reasons behind most of your volume, and which ones a help article prevents.
Copy the prompt
You are a support operations analyst. Below are subject lines and first messages from recent tickets. Customer names, emails and account numbers have been removed. Group them by the underlying reason the customer got in touch, not the words they used. For each group give: the reason in one sentence, the number of tickets, and whether a help centre article could have prevented it. Sort by ticket count, highest first. Put anything you cannot place confidently under 'Unclear' rather than forcing it into a group. Return the result as a plain table. Tickets: [PASTE TICKETS]
One check first. Strip names and account numbers first. Ask your security team before putting customer data into a consumer AI tool.
Rewrite the help article customers still ask about
30 minA shorter article your customers can follow, plus the questions it never answered.
Copy the prompt
You are a support content editor. Below is one help centre article, then five real customer questions it was meant to answer. First, list every question the article does not answer. Then rewrite the article so a customer with no product knowledge can follow it: one task per section, numbered steps, no more than 15 words a step, and no internal terms. Keep every fact and every warning from the original. Where a step needs information the original does not give, write [NEEDS CHECK] instead of guessing. Article: [PASTE ARTICLE] Questions: [PASTE FIVE QUESTIONS]
One check first. The model fills gaps if you let it. Read every [NEEDS CHECK] and confirm the steps yourself before publishing.
Draft a scoring guide for AI answers
25 minA one-page checklist your team can use to grade what the bot sent your customers.
Copy the prompt
You are a customer support quality lead. Build a one-page scoring guide for reviewing answers that AI sent to our customers with no person checking them. Cover five things: factual accuracy, whether it followed our policy, tone, whether it handed off to a person at the right moment, and whether the customer had to come back. For each one give a 1 to 5 scale, plus a plain sentence describing what a 1, a 3 and a 5 look like in a real reply. Then list the kinds of answer a person must always read before it goes out. Our product is [WHAT YOU SELL]. Our customers get in touch most often about [TOP THREE REASONS].
One check first. This is a first draft. Have your policy owner and two senior agents mark it up before you score anyone against it.
Specialized tools, and what to ask vendors
| Tool | What it does | Setup | Best fit |
|---|---|---|---|
| Finpay per resolution, about $0.99 each | Answers customer questions in chat and email from your help centre and past tickets, then hands off when stuck. | WeeksIT sign-off | Most of your volume is repeat questions and your help centre articles are already decent.Skip it ifYour help centre is thin or out of date. It answers from your content, so weak content gives weak answers. |
| Sierraenterprise, priced on outcomes | Runs branded voice and chat agents that follow your policies, act in your systems, and escalate to people. | MonthsIT sign-off | You have high phone volume and the calls need real account changes, not just answers.Skip it ifSmall teams. Setup is a build project run with the vendor, and the starting price is high. |
| Decagonmid-market to enterprise, annual contract | An AI agent — software that answers customers without a person — across chat, email and phone, run by your support managers. | WeeksIT sign-off | You have several product lines and your support policies change often.Skip it ifYou have one simple product and one help centre. You would pay for flexibility you never use. |
| Zendeskper verified resolution, plus Zendesk seats | AI agents built into Zendesk that answer, sort, and route tickets, billed only for resolutions Zendesk verifies. | Under a weekIT sign-off | Zendesk is already your help desk and you want the shortest path to something live.Skip it ifYour product is complex and your answers live in many systems. Test what it resolves before you commit. |
Questions to ask before you buy
Fin — 7 questions to ask them
Also used by. Software and subscription businesses with a large self-serve customer base
- What exactly counts as a billable resolution, and what happens when the customer comes straight back?
- Show me the resolution rate — the share of conversations the bot closes on its own — for a customer whose help centre is as thin as mine.
- When Fin does not know the answer, what does the customer see and what does my agent see?
- How does the handoff to a human work, and does that person get the whole conversation?
- If Fin tells a customer something wrong and we honour it, who carries that cost?
- What does my bill look like in a month where volume doubles, and is there a cap?
- Which customers switched Fin off, and what reason did they give?
Sierra — 7 questions to ask them
Also used by. SiriusXM and other large consumer subscription and services businesses
- How do you define an outcome I get billed for, and who checks that number?
- What does the agent say on the phone when it cannot do what the caller asked?
- How long before the agent can change something in my billing system, not just talk about it?
- What did your last customer of my size have to staff, in people and in weeks?
- How do you stop the agent promising a refund or a policy we do not offer?
- When the agent escalates mid-call, does the caller have to repeat everything to my person?
- Which deployments underperformed, and what was different about them?
Decagon — 6 questions to ask them
- Who on my team writes and changes the agent's rules, and how long is the training?
- Is the number you just quoted deflection — tickets the bot ended — or resolution, and how do you tell them apart?
- How do you handle one product line whose policy differs from all the others?
- What does the agent do with an angry customer, and how does it detect one?
- What does the contract say about accuracy, and what do I get when the agent is wrong?
- How does the price move if my ticket volume drops because the agent is working?
Zendesk — 6 questions to ask them
- What does your verification model actually check, and can I see the cases it rejected?
- How does a verified resolution differ from a ticket that simply closed itself?
- Now that you own Forethought, which product am I buying and how long is it supported?
- What resolution rate do customers on my plan and my ticket mix actually reach?
- What does my bill look like if resolutions double, and where is the ceiling?
- When the agent escalates, what does my person see, and does the customer wait again?
What everyone is asking
- NewJul 2026
Best AI for ticket deflection (tickets a bot ends without a person)?
It depends on where your answers already live. Fin and Zendesk agents sit on your help desk and go live fast. Sierra and Decagon build deeper into your systems and cost more.
What to watch for
Trade press turned sceptical of deflection in 2026, because a customer who gave up still counts as deflected.
Also worth a look. Fin, Sierra, Decagon, Zendesk AI agents
- NewJul 2026
Does AI support actually cut costs or just move them?
Both happen. Salesforce cut support headcount from about 9,000 to about 5,000. Klarna cut agents, hired people back, and its support and operations cost rose year over year.
What to watch for
The companies that kept the savings kept people on the hard tickets. The one that cut deepest ended up rehiring.
Also worth a look. Salesforce, Klarna
- NewJul 2026
What resolution rate (share of chats a bot closes alone) should I actually expect?
Ask for resolution, not deflection. Klarna's assistant handled two thirds of chats. Salesforce says half of customer conversations go to AI agents. Both companies reported those numbers themselves.
What to watch for
Every published rate is self-reported. Nobody audits these the way an auditor would check a financial figure.
Also worth a look. Klarna, Salesforce
- NewJul 2026
Will customers hate it?
They mind far less when a person is one click away. Five9 research reported by CX Dive found four in five people will use AI support when that route exists.
What to watch for
The same research says 41% are less likely to use a company that uses AI support, rising to 53% with no human option.
Also worth a look. AI-only support, AI with a visible route to a human
- NewJul 2026
Should I cut support headcount when I turn this on?
Do not cut on a forecast. Commonwealth Bank cut 45 roles on a call-volume claim that turned out to be wrong, then apologised and reversed the decision.
What to watch for
A union and a tribunal forced those numbers into the open. Most companies never have to show that data.
Also worth a look. Cut headcount first, Hold headcount and measure for a quarter
Worth following
Support Driven
A Slack community of customer support professionals, with events and a salary database.
daily · community
Why them
Practitioners compare notes on what AI tooling does to their queues and their jobs.
Shep Hyken
Customer service speaker and author who runs an annual customer experience research report.
weekly · newsletter
Why them
He publishes what customers say they want from AI service, not what vendors say.
CX Today
Trade publication covering contact centre technology, vendors, and AI product news.
daily · publication
Why them
It catches vendor changes and acquisitions early, and those move your contract terms.
CX Dive
News site from Informa TechTarget covering customer experience and service operations.
daily · publication
Why them
Original reporting on named companies with real numbers, not recycled vendor claims.
Jeff Toister
Customer service consultant, author, and LinkedIn Learning instructor on service culture.
weekly · blog
Why them
Practical writing on what happens to your team once the easy tickets disappear.
Fin AI research
The engineering research blog of Fin, the vendor that used to be called Intercom.
monthly · vendor research blog
Why them
Detailed write-ups on why support agents fail. This is a vendor, so read it that way.