§ Playbook · Customer services
The AI playbook for Customer services
Customer service is where AI headcount stories are made and unmade: Klarna automated the work of 700 agents, then rehired humans within eighteen months. This playbook covers what resolution rates really look like, which conversations should never meet a bot, what your BPO contract is now worth, and who legally owns what the chatbot says.
Free · no form, no follow-up sequence
A Canadian tribunal ruled that Air Canada owned its chatbot's invented bereavement policy and had to honour it. Your bot's hallucinations are your promises now. Plan accordingly.
700 → rehiring
Klarna's arc: agents' work automated in month one, humans hired back within 18 months.
Bloomberg / Forbes, 2024-25
41% vs 80%
real enterprise median automated resolution vs vendor headline claims.
Zendesk benchmarks, 2026
64%
of customers would prefer companies didn't use AI for service at all; 53% would consider switching.
Gartner survey, 2024
§ Part one
What actually changes.
Klarna is the whole playbook in one company. February 2024: the AI assistant handles 2.3 million chats in its first month, does 'the work of 700 agents', projects $40m of profit improvement. Eighteen months later the CEO is on record - 'we focused too much on efficiency and cost... the result was lower quality' - and Klarna is rehiring humans. The AI was genuinely good at the routine two-thirds. It was deployed on all of it.
The durable pattern underneath: AI fully handles simple, high-volume queries - order status, passwords, billing lookups - and it is now doing about 30% of service cases across the market. Agent-assist delivers the safest returns. And full automation reliably fails on disputes, fraud and hardship: exactly the conversations that decide whether a customer stays.
01
The routine tier genuinely automates
Order status, password resets, standard billing. Salesforce data has AI at ~30% of cases now, heading for half by 2027. This tier is real, measurable, and where all the credible savings live.
02
Agent-assist is the quiet winner
Summaries, suggested replies, auto-wrap-up, live guidance. Less exciting than a bot with a name, and it is where the CSAT holds: the human stays in the conversation, just faster.
03
Voice arrives properly
Sierra stood up a 70%-resolution voice agent for Singtel in ten weeks and now serves 40% of the Fortune 50. Voice automation is no longer the weak sibling - which makes choosing what it should not handle more urgent, not less.
04
Escalation design becomes the product
Zendesk's data shows the AI-to-human CSAT gap nearly vanishing when handover is clean. The bot is not your experience; the moment it hands over is.
05
The honest number is 41%
Vendor claims run 70-80% resolution; the enterprise median is 41.2%, and deflection is not resolution. Whatever you are quoted, ask how 'resolved' is defined and watch the number change.
06
Complex and emotional stays human
Disputes, fraud, bereavement, hardship. This is what broke Klarna's quality scores and what the 64% of AI-averse customers are really talking about. Route it to people, visibly and quickly.
§ Part two - the honest bit
What it costs you.
The headcount reality is messier than the headlines in both directions. Salesforce really did take support from 9,000 to 5,000. But only about 20% of service leaders have actually reduced staffing, Gartner predicts half of those who cut citing AI will rehire under new titles by 2027, and Commonwealth Bank of Australia had to reverse 45 redundancies and apologise when the union showed call volumes were rising. The cuts that stick are the ones made after the resolution numbers were proven - not before.
The commercial shift is just as sharp. A human interaction costs $6-12; an AI one under $1.50. Hedge funds are shorting the BPO industry on that arithmetic, and vendors are hedging their own model - Zendesk now charges $1.50 per automated resolution. Every service contract you hold was priced for a labour model that is ending.
01
L1 shrinks, escalation hardens
The routine tier gets smaller; everything reaching a human is harder and more emotional. The remaining job is more skilled and more draining - staff it, pay it and support it as such, or churn will do your workforce planning for you.
02
New roles appear whether you plan them or not
Conversation designers, bot-transcript QA, escalation specialists, AI supervisors. Salesforce found 81% of reps saying their roles got more specialised. The org chart moves; better it moves on purpose.
03
Your BPO contract is mispriced
Multi-year, per-FTE pricing against sub-$1.50 AI economics. Reopen it: outcome pricing, automation gainshare, quality floors. Your provider has already run this spreadsheet - the only question is which side of it you are on.
04
You own what the bot says
Air Canada: liable for an invented policy. DPD: bot swearing at customers, viral. Cursor: hallucinated policy, cancelled subscriptions, public apology. Legal, brand and service now share custody of every automated sentence.
05
Customers did not ask for this
64% would rather you didn't; 53% would consider leaving. That is not a reason to stop - it is a reason to automate invisibly well and keep the path to a human short and unhidden.
06
The unit economics can turn on you
Gartner warns AI cost-per-resolution may exceed $3 by 2030 - dearer than offshore humans. Model token and orchestration costs per resolution now, or rediscover them at renewal.
§ Part three
What good looks like.
Good customer service is a deliberate split: machines resolving the routine tier at proven rates, humans owning the conversations where trust is won or lost, and a handover between them so clean customers barely feel the seam. Quality is measured in CSAT and retention, deflection is a cost line, and no contract - vendor or BPO - is priced on the old labour model.
01
Automation with an honest scorecard
Resolution measured end-to-end, CSAT tracked separately for bot and human contacts, and the routine tier automated only as fast as quality holds. The board sees the same numbers the floor does.
02
Humans where humans are the point
Disputes, fraud, hardship and anger routed to skilled people fast, with the bot's context attached. The escalation team is senior, supported and staffed for difficulty, not volume.
03
Guardrails on every automated sentence
Grounded answers, tested policies, no invented refunds, and an owner for what the bot is allowed to promise. The Air Canada ruling taped to the wall, metaphorically or otherwise.
04
Contracts priced for the new economics
Vendor deals on outcome pricing with resolution honestly defined; BPO agreements rebuilt around automation gainshare and quality. Nobody pays human rates for machine work - in either direction.
The change, across all five pillars
People
Escalation roles upgraded - senior pay and real support for the hardest conversations, all day, every day.
Processes
Handover designed as the product, and the bot's boundary decided with legal and brand in the room.
Data
Grounded knowledge and transcript QA behind every automated answer the company is now liable for.
Platforms
Vendor deals on outcome pricing, with 'resolved' defined by you rather than the sales deck.
Partners
BPO contracts rebuilt around automation gainshare and quality floors before the margin quietly moves.
§ Part four
The first 90 days.
The order matters more here than anywhere: prove resolution quality first, touch headcount last. Klarna and CBA ran the other sequence so you don't have to.
Ⅰ
Weeks 1-2
Map contacts by what they really are
Volume, cost and emotional stakes per contact type. Draw the line between routine and human-required - with legal and brand in the room for the bot-boundary decision.
Ⅱ
Weeks 3-6
Automate one routine contact type
Grounded knowledge, tested answers, clean escalation, approval gates on anything the bot can promise. Agent-assist for everyone else at the same time - it is the safer half of the value.
Ⅲ
Weeks 7-10
Hold it against real measures
End-to-end resolution, bot CSAT vs human CSAT, escalation quality, cost per resolution including tokens. If quality holds, widen. If not, you just saved yourself a public reversal.
Ⅳ
Weeks 11-13
Then - and only then - the structure
Resize the routine tier on proven numbers, upgrade the escalation roles, open the BPO renegotiation, and tell the team the honest plan including what you will not automate.
The day-90 scorecard · what you should be able to show
§ Part five
The traps, so you can skip them.
Trap 01
Cutting on the projection
CBA cut 45 roles on a bot's promise while call volumes rose, then reversed and apologised. Headcount moves last, on measured resolution, or it moves back - expensively and publicly.
Trap 02
Worshipping deflection
A deflected customer is not a resolved one; sometimes they are just a customer who gave up. Measure whether the problem ended, not whether the ticket did.
Trap 03
Letting the bot near policy
Bereavement fares, refunds, account restrictions - anywhere an invented answer becomes a legal promise. Ground it, gate it, and give legal a veto they can actually use.
Trap 04
Hiding the humans
Customers who want a person and cannot find one become the 53% who leave. A short, visible path to a human is the cheapest retention programme you will ever run.
Trap 05
Leaving the BPO contract sacred
Your provider's costs are collapsing inside a contract priced on FTEs. Every quarter you do not reopen it is margin you donated.
Trap 06
Ignoring what the job becomes
The humans who remain get the hardest conversations all day, every day. Without senior pay, real support and career paths, your quality plan quietly walks out the door.
§ Before you go
Six questions worth an honest hour.
Score yourself before someone with a rate card does. If more than two answers are uncomfortable, that is not a crisis - it is an agenda.
The self-assessment
§ Start with the workshop
Book your
Catalyst Workshop.
Bring the strategic change you have been circling. In 1-2 executive days, we will turn it into the first move, the case for it, and the mobilisation plan.