Gartner's survey of customer service leaders found 85% planned to explore or pilot customer-facing conversational generative AI in 2025. On the ground, that has largely happened: AI in some form is now close to standard kit in the contact centre, not a differentiator.
What hasn't happened at anything like the same rate is redesigning the process underneath it. McKinsey's Global AI Survey found only 21% of organisations using generative AI have fundamentally redesigned any of their workflows around it. That figure isn't specific to contact centres - it spans every function McKinsey surveyed - but there's no reason to think customer service is meaningfully ahead of that average, and plenty of reason from what we see in engagements to think most deployments still look bolted on rather than rebuilt.
That gap is not a rounding error. It is the single best predictor of whether an AI deployment in customer experience produces a measurable result or just a more expensive version of the same problem.
What "bolting it on" looks like
Most AI in contact centres today sits on top of a process designed for a world without it. The routing tree still sends every complex query down the same path it always did. The QA scoring model still grades average handle time the same way whether or not an agent had AI assistance. The escalation policy still requires human sign-off on categories of query the AI is now perfectly capable of resolving unsupervised - because nobody went back and rewrote the escalation policy.
The result is an assistant that drafts a response faster, inserted into a process that was never shortened to take advantage of it. You have added a step, not removed one. The call still takes roughly as long. The agent is now also proof-reading the AI.
Where the current AI actually sits
Look at what most contact-centre AI is actually doing today: drafting a reply for an agent to check, summarising a call after the fact, surfacing a customer's history before the conversation starts. All genuinely useful. All of it, by design, sits on top of the existing workflow rather than changing it. None of it touches routing, escalation authority, or how success gets measured.
That is assistance, not redesign - and assistance alone tends to plateau quickly, because it is still bounded by the shape of the process it was dropped into.
The three things that actually need rewriting
1. Routing logic
The question is not what the AI is technically capable of resolving. It is what it is authorised to resolve unsupervised, and that authorisation has to be an explicit design decision, not an accident of the old routing tree still being in place.
2. QA and scoring
If AI-assisted calls are graded on the same handle-time and script-adherence metrics as unassisted ones, you are optimising for the wrong thing and will not see it in the numbers until someone asks why the "AI-powered" contact centre isn't actually any faster.
3. Escalation and override authority
Someone needs explicit, named authority to override what the AI suggests - and overriding it has to be faster than following it blindly, or agents will default to compliance over judgement, which defeats the point of having a human in the loop at all.
The bottom line
Adoption was never the achievement. Every contact centre serious about customer experience will have AI live somewhere by now - that bar is already cleared. Workflow redesign is the unglamorous 80% of the actual work, it is significantly harder than switching on a tool, and by McKinsey's measure it is precisely the part nearly four in five organisations adopting generative AI are currently skipping. That is exactly why it is where the returns still sit unclaimed.