Table of Contents
At 11:47 on a Tuesday night, a customer opens a chat to ask why her order still says “processing” four days after the confirmation email. Nobody is awake at the warehouse. Nobody needs to be. The reply lands in nine seconds, pulls her order status from the fulfilment system, and offers a refund or a replacement. She picks the replacement. The ticket closes before anyone clocks in.
That isn’t a futuristic demo. It’s a Tuesday. Businesses use AI to improve customer service in exactly this unglamorous way: faster answers, fewer handoffs, and fewer people stuck repeating the same five explanations for eight hours a day.
The Four Places AI Actually Earns Its Keep
Vendor decks make AI look like one big feature. In practice it splits into four jobs, each with a different payback period.
- Self-service deflection. Answering the questions that never needed a human in the first place.
- Routing and triage. Getting each ticket to the right desk before it goes stale.
- Agent assist. Drafting replies, summarising history, and suggesting next steps while a human stays on the thread.
- Prediction. Spotting churn risk, delivery delays, or a backlog forming before customers complain.
Most companies start with the first and quietly get the biggest return from the second. Routing is boring. Routing is where the money is.
Chatbots Finally Stopped Sounding Like Chatbots
Early bots ran on decision trees. Miss three words and the whole thing unravelled, which is why a generation of customers learned to type “agent” as their first message.
Large language models changed the failure mode. A modern support assistant reads the customer’s actual sentence, checks it against the order database and the help centre, and replies in language a person would use. An intelligent agent that reasons about context can also recognise when it’s out of its depth and hand over. That judgement, more than answer quality, is what separates a bot people tolerate from one they resent.
What the numbers look like
Klarna’s published figures are the ones everyone quotes: an assistant doing the work of roughly 700 full-time agents, average resolution time falling from 11 minutes to under two, and a 25% drop in repeat enquiries. Less quoted is that Klarna later rebalanced back toward human agents for complex cases. The lesson isn’t “replace everyone”. It’s “let software take the repeat work and keep humans for the messy 20%”.
Retail and fintech tend to see deflection rates between 30% and 50% on first-line queries within six months, provided the knowledge base is genuinely maintained. Companies that treat the help centre as a write-once archive usually land nearer 15% and conclude, wrongly, that the technology doesn’t work.
Triage: The Unsexy Win
In most contact centres, a quarter to a third of tickets get transferred at least once before reaching someone who can solve them. Every transfer resets the clock and burns patience.
AI triage reads the message, predicts category and urgency, and routes accordingly. It reads context too: two previous contacts about the same dishwasher, a delivery already running late, a contract up for renewal next month. That determines whether the ticket goes to first line, a specialist, or retention. Google Cloud’s contact centre tooling, for instance, pairs call transcription and sentiment scoring with routing rules, so supervisors can see queue pressure building in real time rather than at the end-of-day report.
Teams that attach a short AI-generated summary to every escalation routinely report a meaningful drop in handle time, because the second agent no longer restarts the conversation from zero. Shave 40 seconds off a few million contacts and you’ve funded a hiring round.
Agent Assist: Keeping the Human, Removing the Busywork
Agent assist is the version of AI that support staff actually like, since it shortens their day instead of threatening it. Three features carry most of the value: auto-drafted replies, instant summaries of long threads, and real-time knowledge retrieval that surfaces the right policy page mid-conversation.
Building it is less exotic than it sounds. Most implementations are a search layer over your documentation plus a language model call to draft the response. This practical guide to building with the OpenAI API walks through the pattern: chunk the help content, embed it, retrieve the relevant passages, and constrain the model to answer only from what came back. Agents edit the draft and hit send. Draft acceptance rates of 60% to 70% are common once the knowledge base is clean, and that’s where the time savings live.
Prediction Turns Support From Reactive to Proactive
Reactive support is expensive because you pay for a problem twice, once when it happens and again when the customer calls about it. Predictive models flip that order.
A subscription business notices a customer’s usage fell 60% over three weeks and their last two tickets went unresolved. The model flags churn risk and someone picks up the phone. A retailer sees a carrier running two days behind on one route and messages affected customers before the parcels fail to arrive. A software company spots error rates spiking for a single customer’s integration and opens a ticket on their behalf. None of it requires new data, just the willingness to join up records you already have.
The Data Problem Nobody Wants to Own
None of this works if the assistant can’t see the order, the invoice, and the last conversation in one place. That’s where most projects stall, and it’s rarely a model problem.
Support data usually lives in four or five systems bought by different departments at different times. A returns assistant that can read the warehouse but not the payment system will confidently tell customers the wrong thing, which is worse than saying nothing at all. A single customer record matters more than which model you pick. Choosing a cloud-based ERP that suits a growing company is often where that plumbing gets decided, long before anyone argues about prompts.
Build, Buy, or Partner
There are three routes. Buy an off-the-shelf support platform with AI bundled in, build on a model provider’s API, or hire someone to do it. Buying is fastest and cheapest to start. Building gives you control over tone and integration but demands engineering time you probably can’t spare. On the third route, knowing what an AI agency does and how to vet one will save you from a six-figure pilot that never leaves the sandbox.
A rough cost check: a serious deflection deployment runs from a few thousand to a few tens of thousands a month depending on volume, against a fully loaded human agent cost that often sits above $40,000 a year. The maths only works if the deflection is real, meaning the customer’s problem is genuinely solved rather than the ticket simply closed.
Where It Breaks, and What to Watch
AI handles “where is my order” beautifully. It handles a bereavement call about a joint account badly. The failure modes are predictable, so design for them up front.
- Escalation should be one step, not three. If a customer has to argue with a bot to reach a human, you have spent money to make them angrier.
- Complaints, regulated advice, anything touching money or vulnerability should go to a person by default, not by exception.
- Confidence thresholds need tuning on your own transcripts. Vendors demo on clean English and deploy into customers typing abbreviations on a cracked phone screen at half past midnight.
- Log everything, including the retrieval source and prompt version. When the bot gets something wrong, a transcript alone won’t explain why.
Then measure the right thing. Containment rate flatters everybody in the room. Resolution rate and post-contact satisfaction tell the truth. A bot that closes 60% of chats but pushes a fifth of those customers straight into a second contact has saved precisely nothing, and the dashboard will never mention it.
