There’s a phrase that keeps coming up in customer service conversations that I think we need to retire, or at least interrogate a bit more aggressively: deflection.
For the better part of two decades, we’ve treated deflection as the goal. Fewer calls, fewer contacts, fewer humans involved. It shows up nicely in dashboards. It makes the math look better. It gives the illusion of control. But it also quietly shifts the burden onto the customer. And at some point, that trade starts to break.
In my conversation with Pasquale DeMaio, who runs Amazon Connect Customer at Amazon Web Services, you can feel that shift happening – not as a headline, but underneath the way they’re actually building and deploying these systems.
What they’re really talking about isn’t deflection at all. It’s resolution. And more specifically, what it takes to resolve something at the right speed and with the right level of intelligence for that moment. That sounds obvious until you start pulling on it.
Because most contact centers are still wired to optimize for things like handle time and cost per contact. Those aren’t wrong, exactly – they’re just incomplete. They were built for a world where every interaction required a human, and the only lever you really had was efficiency.
AI changes that. Not just by making things faster, but by changing what the interaction actually is.
There are moments where the fastest possible answer is the best experience you can deliver. No waiting, no escalation, no friction. Speed, in that case, is the service. And then there are moments where speed is the wrong answer – where what the customer actually needs is context, judgment, or just someone who understands what’s going on. That tension is where things start to get interesting. And messy.
You can hear it in the way Pasquale describes the system. There’s this implicit continuum – some things should be fully automated, some things should be AI-assisted, and some things should stay human. But that’s not a philosophical model. That’s an operational problem. Because now you have to decide, in real time, what kind of moment you’re in.
And that’s where most organizations start to wobble. Not because they lack AI, but because they haven’t rethought the system around it.
A lot of what’s happening inside platforms like Amazon Connect looks, on the surface, like incremental improvement – better routing, real-time agent assist, automatic summaries so people aren’t stuck doing wrap-up work all day. All of that matters. But the more interesting shift is where that work is going.
The system is starting to take on the operational overhead that used to sit on the agent.
Which raises a question that I don’t think the industry is fully grappling with yet: if you strip out the repetitive, procedural work, what’s left of the role?
“If you step back from all of it, the shift here is less about adding AI to the contact center and more about rethinking what the contact center is actually for.”
Because what’s left isn’t simpler. It’s harder. It’s more judgment-driven, more situational, more human in the ways that are difficult to standardize and measure. And that doesn’t map cleanly to how most contact centers are staffed, trained, or managed today.
There’s a similar gap in how people are talking about agentic AI. A lot of the conversation swings between extremes – either full autonomy or full control. Let the system run everything, or keep it locked down so nothing breaks.
What you actually see in practice is much more grounded. You let these systems operate with some degree of flexibility, but you put hard boundaries around the things that matter – payments, identity, security, anything that carries real risk. It’s not as exciting as the fully autonomous story, but it’s how you build something that people will actually trust.
And trust ends up being the throughline here, whether it’s stated explicitly or not. Not brand trust. System trust. Does this thing do what it’s supposed to do? Does it handle sensitive moments correctly? Does it know when to stop and bring a human in? Because if the answer to those questions is “not reliably,” then none of the higher-level promises really matter.
The other piece that comes through pretty clearly is just how fragmented most CX environments still are. Different tools solving for different slices of the problem, all stitched together just enough to function. AI doesn’t fix that. If anything, it exposes it faster. You can make each piece smarter, but if the system itself isn’t coherent, you just end up with smarter fragmentation.
And that’s probably the part that gets glossed over the most. Everyone wants to talk about the intelligence layer. Very few want to talk about the underlying architecture that intelligence depends on.
If you step back from all of it, the shift here is less about adding AI to the contact center and more about rethinking what the contact center is actually for. Not to deflect demand. Not even just to handle it efficiently. But to resolve it – properly, completely, and in a way that doesn’t create more work downstream.
That’s a much higher bar. And it forces a different set of decisions than the ones most organizations are used to making. That’s really what this conversation gets into. Not the surface-level “AI in CX” story, but the harder question of what happens when the system itself starts to change.
And whether most organizations are actually ready for that.
