AI and customer experience: what your customers notice
AI improves customer experience when it resolves, not when it deflects. Your customers notice three things: whether the answer is right and can be checked, whether the AI admits when it doesn’t know, and whether they have to repeat themselves when they’re handed to a person. Everything else, as far as they’re concerned, doesn’t exist.
What do customers notice when a company puts AI in front of them?
Your customers don’t see your model, your architecture or your automation rate. They see whether their problem got solved, what it cost them and whether they were treated as someone with a problem or as a number in a queue.
Our team includes people who come from systems architecture for multichannel contact centres and from creative and UX direction at a national telecoms operator; you can see who’s on the team. From both sides you learn the same thing: what wears customers down usually isn’t the wait, it’s having to explain their problem all over again.
And a fast but wrong answer is worse than a slow, right one: it generates a second call, and that one arrives angry.
AI can improve all of that, or make it worse at a new speed. It depends on what you ask of it.
What does good AI customer service do?
It answers with the source
Every answer rests on a specific document: the contract terms, the returns policy, the equipment manual. The customer can check it, and so can the person who picks up if the conversation is handed over. It’s the principle we apply in RAGNIX, answers with verifiable citations: cite or abstain, and never half a procedure.
It says “I don’t know” in time
“I don’t have that information; let me put you through to someone who does” costs a minute. A confident, wrong answer costs a complaint. Honest abstention is customer experience, not a system failure.
It hands over to a person, with context
When the AI escalates, whoever picks up receives what the customer has already said, what has already been checked, the details already identified and the reason for the handoff. The customer doesn’t repeat anything. And the way out to a person is visible and available, not hidden behind three menus.
It doesn’t promise what it can’t deliver
A refund, a delivery date, an exception to a tariff: these are decisions for a system, or for a person with the authority to make them. A scoped agent looks them up in the system that decides them, or escalates; it doesn’t improvise them. An invented promise becomes a commitment somebody will have to keep or break.
It’s the same on every channel
If the chat says one thing, the phone another and email a third, customers stop trusting all three. One knowledge source for every channel, with the same answers and the same citations.
It can explain what it decides
If the AI prioritises a ticket, flags a transaction as suspicious or turns something down, a person has to be able to explain why to the customer. A decision nobody can explain can’t be defended in front of them.
It only sees the data it needs
What your customers tell you doesn’t have to leave your network or end up in a service you don’t control. Decide where that data lives and what each agent can see: the one reporting on a fault doesn’t need payment details. An agent that sees less also has fewer places to go wrong.
Which shortcuts damage customer experience?
Most projects that go wrong start with good intentions and a shortcut. These are the most common:
| Shortcut that damages the experience | Alternative |
|---|---|
| Measuring success by the share of contacts deflected | Measuring resolution and recontact |
| An agent with no scope that answers everything | A written scope: what it answers, what it looks up and what it escalates |
| Confident answers even without support | Cite or abstain |
| Escalating without context, so the customer starts from scratch | A handoff with a summary, the details identified and what has been tried |
| Hiding the person behind menu loops | A way out to a person, visible at all times |
| Letting the AI promise refunds or dates | Commitments come from the system that decides them, or from a person with authority |
| One assistant per channel, each with its own version | One knowledge source for every channel |
The first does the most harm, because it steers the whole project the wrong way. A deflected contact that comes back by phone the next day isn’t a saving: it’s the same contact, more expensive and with a more annoyed customer.
How do you measure whether AI improves customer experience?
With metrics that look at the customer, not just the contact centre. These are the five we propose in any pilot:
- Resolution. Was the problem solved? The customer confirms it, or no new contact arrives for the same reason.
- Recontact. How many customers come back about the same thing in the following days, on any channel? This is the metric that exposes apparent deflection.
- Handoff quality. Did the person who picked up the case have to ask anything again? Did the case reach the right team?
- Unsupported answers. How many times did the AI answer without a source? How many times did it abstain, and how many of those abstentions were right?
- Corrected promises. How many times did a person have to undo something the AI said or committed to?
Deflection can stay on the dashboard as a capacity figure. What it can’t be is the goal.
What changes for the people who answer?
If the AI handles the repetitive work, your team gets the hard cases, and gets them already under way. Their job changes: less retyping what the customer already said, more deciding.
For that to work, whoever answers needs to see what the AI has done and what it relied on, to be able to correct it without asking permission, and to know how far to trust what it suggests. If they don’t, they either trust it too much or work around it, and either way the customer notices.
That’s where customer experience (what the customer goes through) and user experience (what the person using the tool goes through) meet: they’re two sides of the same conversation. We cover it in AI and user experience.
Where do you start?
Pick a single flow with volume and clear rules: a billing query, the status of a fault report, a change of details. Write down the agent’s scope, connect its sources with citations, design the handoff before the conversation and measure resolution and recontact from day one.
If you’d like to do it with the people who design it and build it, start with our applied AI, from pilot to production.