The problem
Deflection is not service
Here's the problem. Many teams adopt AI in customer service with one metric in mind: how many conversations can we keep away from a human? The result is familiar to anyone who has tried to cancel a subscription. Loops. Canned answers. A "talk to an agent" option buried three menus deep.
Meanwhile, the actual pain in most support teams is somewhere else.
What's really happening
Agents spend a large share of their day on work that isn't helping anyone: reading long threads to understand what happened, searching for an order, finding the right policy paragraph, rewriting the same answer for the fifteenth time. The complex, emotional cases — the ones that decide whether a customer stays — wait in the same queue as "what are your opening hours?"
Who experiences it
Customers wait too long for simple answers and get poor answers to complex ones. Agents burn out on repetition. Team leads can't see what's driving volume. The business pays for both the support team and the churn.
The AI thought
Help the agent first
This is where AI changes the equation — if it's pointed at the right job.
Language models are now genuinely good at three things support teams do all day: understanding what a message is about, summarising a long history, and drafting a clear reply from known information. Grounded in your own help centre and policies, they can do this accurately enough to save real time.
So start there. Use AI to make every agent faster and better-informed. Automate a reply only when you have evidence — from weeks of approved drafts — that it's reliable for that kind of question.
What should NOT be automated
Refunds and compensation decisions. Complaints. Conversations with vulnerable customers. Anything with legal implications. And never hide the human: a customer should always be able to reach a person in one step.
If customers are trying to escape your AI, it isn't customer service. It's a wall.
Design for the moment someone needs a human, and everything else gets easier.
The solution
An AI layer that works behind the agent
Triage
Every incoming message is tagged with intent, urgency and sentiment. An angry customer whose order is late jumps ahead of a general enquiry.
Context in one place
Before the agent opens a ticket, they see a three-line summary of the history, the relevant order, and the policy that probably applies.
Drafts grounded in policy
A suggested reply, written in the brand's voice, citing the relevant help article. The agent edits and sends — and each edit teaches you where the knowledge base is weak.
Careful automation
For a small set of simple, high-volume intents, answers go out automatically — with a clear way to reach a person.
How it works
From problem to outcome
Problem
Support teams are overloaded with repetitive questions while complex, emotional cases wait in the same queue.
Input
Incoming emails, chats and messages; order and account data; the help centre; and past resolved conversations.
AI / Agent
A model classifies each request by intent and urgency, gathers the relevant context, and drafts a reply grounded in policy.
Decision
Simple, low-risk questions are answered directly; anything ambiguous, emotional or high-stakes goes to a person with the draft attached.
Action
Agents review, edit and send. Customers can always reach a human in one step.
Outcome
Faster answers to simple questions, more agent time for hard ones, and fewer customers trapped in a loop.
Level 1
Manual
Every ticket read, researched and answered by an agent.
Level 2
AI-assisted
AI triages, summarises and drafts; agents edit and send.
Level 3
AI-automated
Simple, well-defined intents answered automatically with clear escalation.
Level 4
AI-agentic
An agent can take bounded actions — rebook, reissue, update an address — within strict policy, with logging and review.
Implement this
0/8 done
Identify the workflow
Take a week of tickets and sort them by intent. Most teams find a handful of categories make up the bulk of volume.
Identify repetitive tasks
Order status, returns policy, password resets, opening hours — plus the hidden work of looking up context for every ticket.
Define required inputs
A clean, current help centre is the foundation. Add read access to order or account data for the categories you automate.
Select the appropriate AI capability
Classification for triage, retrieval-grounded generation for drafts, and summarisation for long threads.
Build the workflow
Start with agent-assist: AI drafts, humans send. Only automate replies for categories where drafts are consistently approved unchanged.
Add human approval where required
Refunds, complaints, vulnerable customers and anything legal always go to a person. Make 'talk to a human' easy to find.
Measure the result
Track first-response time, resolution time, re-contact rate and customer satisfaction — not just 'deflection'.
Iterate
Review a sample of AI-handled conversations every week. Fix the knowledge base before you tune the model.
Business impact
How to know it's working
Measure outcomes for customers, not activity for the system.
- Speed. First-response and resolution times.
- Quality. Re-contact rate — did the answer actually solve it?
- Experience. Customer satisfaction on AI-assisted versus fully manual conversations.
- Team health. Time agents spend on complex cases versus repetitive ones.
The takeaway
The goal isn't fewer conversations with humans. It's better ones.
Share this thinking
The newsletter
Think better about AI.
One practical AI idea. One real-world problem. Delivered periodically.