Our CTO, Dane Maddox, has a line he uses when customers ask him how much of their contact centre they can automate. He says: "Ask me a better question."
He is right, and it took me a while to admit it. The percentage question is the wrong question. It treats automation as a dial you turn up until the finance director stops asking about headcount. What it never asks is the only thing that matters: which conversations should a machine be handling, and which ones should absolutely not.
So let us be honest about where the line sits. Not diplomatic. Honest.
AI earns its place where the answer is knowable
There is an enormous category of customer contact where the correct answer already exists somewhere in your systems, and the only thing standing between the customer and that answer is a queue.
Where is my order. What is my balance. Can I move my appointment. Have you received my document. Reset this for me. Confirm that. Cancel that.
These are not conversations. They are lookups wearing a conversation costume. Putting a human on them is not a kindness to the customer, who wanted the answer at 11pm, and it is certainly not a kindness to the human, who did not take this job to read balances aloud four hundred times a week.
This is where automation genuinely earns its place, and where NeonNow Agentic does its best work. Not as a wall to hide behind, but as the thing that gets the knowable answers out of the way fast, at any hour, so that the humans are free for the conversations that need them.
Humans earn their place where the answer is a judgement
Then there is the other category. The customer whose payment failed because her husband died in March and the direct debit was in his name. The business owner who has been trying for three weeks and is now, understandably, furious. The person who is technically outside policy but obviously deserves an exception.
No model resolves those. Not because the model is not clever enough, but because the resolution is not an answer, it is a decision. It requires someone with the authority to bend, the judgement to know when bending is right, and the human instinct to hear what is not being said.
Every organisation that has automated its way over that line has ended up walking it back, usually loudly and usually in public. The pattern is always the same. Cost comes down for two quarters. Complaints go up. Retention goes quiet, then goes down. Somebody senior gets asked a difficult question in a board meeting.
The resolution is not an answer, it is a decision.
The handover is where most deployments actually fail
Here is the part that gets skipped in almost every AI business case I read.
Both categories exist in the same conversation. A customer starts with a knowable question and ends with a judgement call. She does not announce the transition. She just gets to a point where the automation cannot help her any more, and what happens in the next fifteen seconds determines everything she thinks about your company.
If the handover drops the context, you have not built an AI contact centre. You have built a very sophisticated hold queue.
The research is unambiguous. Gartner found 62% of channel transitions are high effort, and that easy transitions save around four minutes of agent time per journey precisely because agents are not re-collecting what the customer already told the machine. It also found that customers do not mind switching channels when their context travels with them and the issue gets resolved in one continuous interaction.
So the design principle is not "how much can we automate". It is "what does the human receive at the moment of handover".
What the human should receive
When an agent picks up a conversation the automation has been running, three things should already be on the screen before he says hello.
- Who this is, verified. Not "a customer calling from an unknown number".
- What has happened so far, summarised in a sentence, not a transcript he has to speed-read while she waits.
- What she is actually trying to do, and how she is feeling about it.
That is what NeonNow IQ is for. It runs across every channel in NeonNow CX and puts the three things above in front of the agent: who the customer is, what she is trying to do, and a summary a human can use in three seconds. Automation without that is not intelligence. It is deflection with better marketing.
Underneath, the platform matters just as much, and this is where we work side by side with Twilio. Identity has to be established once and trusted afterwards, which is what Twilio Verify handles across SMS, voice, email and silent network signals. Voice automation has to be able to hand a live call to a person without dropping the thread, which is what Twilio ConversationRelay is built for. The context has to survive the crossing, which is the job of Twilio Conversation Memory. And messaging has to be two-way and branded so the customer can actually reply, which is where Twilio Messaging comes in. Get that foundation right and everything above it works.
You are probably measuring the wrong thing
Most AI business cases in customer service are built on containment. What percentage of contacts did the bot handle without a human. It is a simple number, it goes up over time, and it is close to useless.
Containment tells you how many customers the automation kept. It tells you nothing about how many it should have kept. A bot that traps a grieving widow in a loop about her direct debit has a beautiful containment rate and has done real damage to a real person and a real relationship.
Here are three numbers I would put on the board instead.
- Escalation quality. Of the conversations handed to a human, what proportion arrived with verified identity, a summary and an intent already attached? If the answer is not close to one hundred per cent, that is the first fix.
- Repeat rate. How often does a customer supply information the business already had? This is the number nobody tracks and everybody feels.
- Return to self-service. After a customer is escalated to a human, does she come back to the automated channel next time? Gartner found that 74% of customers who get an easy transition do exactly that. If yours do not, your handover is telling them something.
Containment measures how well you avoided the customer. These three measure how well you served her, which is a different and considerably more useful thing.
The honest version of the business case
If you want the version I would give a board, it is this.
Automation will handle a large share of your contact volume, and it should, because most of that volume is not conversation. It is retrieval. Do that well and you will free up real capacity and give people answers at 11pm, which they will thank you for.
It will not reduce your need for good people. It will change what you need them for, and it will raise the bar on what "good" means, because every conversation that reaches a human will now be one that a machine could not solve. Those are harder conversations. That is the point.
And the whole thing lives or dies on whether the conversation remembers itself as it moves between the two. Get that right and the customer never knows where the line is. Get it wrong and she finds it immediately, usually while spelling her surname for the third time.
One more thing worth saying out loud
There is a version of this conversation that treats agents as a cost to be minimised until the technology catches up. I want to be clear that we do not build for that version, and we are not interested in customers who do.
Not because it is unkind, although it is. Because it does not work. The organisations getting the most out of automation right now are the ones that used it to give their people harder, more interesting work, not fewer hours. They kept the judgement in the building. They gave their agents better information and better tools and then trusted them with the conversations that were always going to need a person.
Their contact centres are cheaper to run and better to work in, which is not a coincidence. It is what happens when you put the machine on the lookups and the human on the decisions, and you build something that remembers what happened in between.
Where does your line sit?
Most contact centres have never actually mapped this. They have a bot doing some things, humans doing others, and a handover in the middle that nobody has looked at closely since it went live.
If that sounds familiar, book a demo and we will walk one conversation across the line with you, from automation to human, and you can see exactly what survives the crossing. Or talk to our team if you would rather start by working out where your line should sit in the first place. That conversation is free, and in our experience it is the one that saves people the most money.
.svg)
