---
Title: "5 things NZ customer experience leaders are actually doing with AI – and what's not working"
Url: "https://devrev.ai/blog/nz-cx-leaders-ai-customer-experience"
Published: "2026-06-15"
Last Updated: "2026-06-15"
Author: "DevRev Editorial"
Category: "Blog"
Excerpt: "Insights from sixteen senior NZ enterprise CX leaders on what's working with AI in customer experience, where they're hitting walls, and what's next."
Reading Time: 8
---

# 5 things NZ customer experience leaders are actually doing with AI – and what's not working

_Michael Robinson, Head of NZ Business, DevRev_

_Jeff Smith, Office of the CTO, DevRev_

A report published last month by [One NZ](https://media.one.nz/ai-trust-report-2026) found that 62% of New Zealanders would stop using a product or service if they had concerns about how an organisation was using AI. 70% of AI users in New Zealand experienced problems with AI in the past 12 months. And yet, 76% used an AI-powered tool or service in the past year.

That tension of high exposure, rising scrutiny, and eroding trust is exactly the environment enterprise CX leaders in New Zealand are navigating right now. Customers are using AI-powered experiences whether they know it or not. They're also paying closer attention when those experiences go wrong.

Last month, we sat down with sixteen senior CX, product, and technology leaders from some of New Zealand's largest enterprises for a candid roundtable on AI in customer experience.

We covered what's actually working in enterprise AI deployments, where organizations are hitting walls, and what the next 12 months realistically look like.

What follows is our attempt to faithfully capture what came up.

## **1. Agent assistants are delivering – and raising the bar**

The most consistently successful AI use case in the room? Co-pilots that sit inside the agent desktop, surfacing relevant knowledge in real time during customer interactions.

These aren't replacing people. They're making experienced agents faster and new agents competent sooner. Handle time drops. First-contact resolution improves. Training ramp shortens.

The pattern is simple: RAG-powered search over internal knowledge, surfaced contextually during a live conversation. It's not glamorous. But across the room, it was the use case that had moved furthest from "pilot" to "production."

Several organizations are now extending this into autonomous resolution – letting AI handle defined, high-confidence query types end-to-end without human intervention. But the consensus was clear: you earn the right to automate by first proving the AI can assist reliably.

## **2. Knowledge quality is the real blocker – not the model**

Every leader in the room said some version of the same thing: _"Our AI is only as good as our knowledge base, and our knowledge base isn't good enough."_

This wasn't a technology complaint. It was an operational one. Years of accumulated documentation – inconsistent, duplicated, outdated, scattered across wikis and PDFs and tribal knowledge – means even a well-configured AI retrieves unreliable information.

The organizations making the most progress had invested in what one attendee called a "KM triage" – auditing their highest-traffic knowledge areas first, cleaning those, and letting AI serve from a curated subset rather than the entire corpus.

The takeaway wasn't "wait until your knowledge is perfect." It was: **fix the foundations on your highest-impact channels first** (email, chat, KB search), prove value there, then expand.

## **3. Trust and governance are slowing everyone down – and that might be okay**

In a room that included leaders from insurance, banking, telecommunications, and energy, the governance conversation was predictably intense. But what surprised us was the nuance.

Nobody was saying "we can't use AI because of compliance." They were saying: _"We don't yet have the testing frameworks to prove AI outputs meet our bar."_

The gap isn't policy – it's measurement. Most organizations still lack mature ways to evaluate AI accuracy, catch hallucinations before they reach customers, and demonstrate compliance to regulators. Several leaders mentioned deploying AI internally first (agent-facing, not customer-facing) specifically to build confidence and evidence before going external.

Practical governance patterns that came up:

- Human-in-the-loop approval gates for anything customer-facing.
- Data masking and anonymization as non-negotiable baseline requirements.
- Logging and observability as a prerequisite for any production deployment.
- Starting with assistive AI (suggestions an agent can accept/reject) rather than autonomous actions.

The message: governance isn't the enemy of speed. Done early, it _accelerates_ scaling because you don't hit a compliance wall at the worst possible moment.

## **4. Token costs and AI economics are a real conversation now**

This is the talk of the town right now – from Uber to Microsoft, the economics of running AI at scale have moved from theoretical to urgent. Multiple attendees raised this not as a future concern, but as something they'd already been burned by or were actively managing."

The pattern is familiar: you pilot an AI capability, it works well, leadership says "scale it" – and suddenly your inference costs are an order of magnitude higher than anyone budgeted for. Surprise bills. Unobservable token consumption. No clear way to attribute cost to value.

The consensus was that the industry needs to move toward more efficient retrieval architectures – systems that don't pump millions of tokens through an LLM just to answer a structured question. Shared retrieval layers, pre-built knowledge graphs, and deterministic data access (SQL over structured data rather than AI inference over raw documents) were all cited as part of the solution.

One attendee put it plainly: _"We need to know what the AI is doing, how much it's costing, and whether the answer is actually right – before we give it to a customer."_

## **5. The real ambition is proactive, cross-system CX – but almost nobody is there yet**

When we asked where leaders _wanted_ to be in 12 months, the answers were remarkably consistent:

- Proactive outreach before customers contact you – detecting issues from observability data and resolving them preemptively.
- Personalized, context-aware interactions that span channels – a customer who started in chat shouldn't re-explain themselves on the phone.
- AI that connects CRM, engineering, and support to surface insights that no single system can see alone.

These aren't unrealistic visions. They're architecturally achievable. But they require something most organizations don't yet have: a **shared data substrate** that links customer identity, product context, and interaction history across systems.

The gap between "we have AI summarizing calls" and "we have AI predicting which customers are about to churn based on a combination of support patterns, product usage, and engineering issue data" is not a model gap. It's a data integration and memory gap.

## **What's not working**

The room was equally candid about failures and dead ends:

- **"Bolt-on" AI that doesn't connect to anything.** Point solutions that summarize or deflect but can't access the broader context of the customer relationship. They help in isolation but don't compound.
- **Scaling without governance.** Several leaders had seen pilots succeed and then stall at the point of enterprise rollout because security, privacy, and approval processes hadn't been designed in from the start.
- **Underestimating change management.** The technology is often the easy part. Getting frontline teams to trust AI suggestions – and getting senior leadership to accept that roles need to be redesigned, not just "augmented" – is harder.
- **Testing in production.** Multiple attendees mentioned not having robust pre-deployment testing frameworks. AI gets evaluated on vibes rather than metrics, which erodes executive confidence when it inevitably makes a mistake.

## **Where this leaves us**

The NZ CX community isn't behind. If anything, the pragmatism in the room – the insistence on measurable outcomes, clean data, and governance-first thinking – positions these organizations well for the next phase.

The NZ CX community is marching forward with a healthy dose of pragmatism. The insistence on measurable outcomes, clean data, and governance-first thinking positions these organizations well for the next phase.

That next phase isn't about whether to use AI. It's about how to build the _infrastructure_ for AI: shared memory across systems, deterministic retrieval for structured data, governance by design, and operating models that pair human judgment with machine speed.

The organizations that get there first won't be the ones with the most AI features. They'll be the ones that fixed their knowledge foundations, proved value in narrow use cases, and built the connective tissue between their systems before trying to scale.

_This piece reflects themes from a private leadership roundtable held in Auckland in May 2026, hosted by DevRev and Inner Circle Events._

__