CX Network's flagship annual research report, CX Horizons: The Global State of CX in 2026, found agentic AI and AI agents to be the second most significant trend influencing the work of CX practitioners. At the same time, consumer privacy appeared in the top 10 trends for the first time since the research launched 11 years ago.
All Access: AI + Data in CX 2026 brought together thought leaders from tech, finance, retail and healthcare to discuss these and similar topics to get to the heart of these trends.
Our speakers discussed the future of agentic orchestration and what this will mean for customers; the quickly evolving vendor market and how to navigate it; turning customer data into actions; and proving ROI of AI investments.
If you missed the sessions live, you can catch them on CX+ here. Alternatively, read on…
Privacy and security are getting closer to CX
For the first time in the 11 years that CX Network has published an annual state of CX report, consumer privacy appeared as a top-10 trend. Melanie Mingas, CX Network's editor-in-chief, tied this to a broader trust crisis in CX, especially with regards to AI use.
Her practical recommendation was to prioritize transparency – stating clearly what data is being collected and how AI is using it to make decisions – as well as offering customers the choice to interact with a human. All three of these measures will be mandated under the EU AI Act from August 2026.
In a similar vein, and also for the first time in 11 years of research, data security has entered the top-5 CX challenges list. Mingas highlighted three main threats specific to AI, drawing on an interview with Assaf Keren, CSO at Qualtrics and former CISO for PayPal.
- AI agents making mistakes at scale, at high speed and volume.
- Manipulated or corrupted data feeding AI systems – long-termed "rubbish in, rubbish out".
- Malicious actors and prompt injection that causes agents to bypass their own guardrails or leak data.
"Be very, very careful sharing closed service tickets with your LLMs when training them. If it goes in, there's a chance of it coming out again", she cautioned.
Agentic orchestration can bridge the gap between loyalty and efficiency
An audience poll asked: "Which of these CX priorities is currently getting the most executive attention in your organization?" The highest percentage of respondents (35 percent) answered "increasing operational efficiency".
Traditionally, operational efficiency and customer loyalty have been seen as mutually exclusive, but what if this isn't the case anymore? This is exactly the question that Tim Friebel, director of AI and business value and Nathalie DeChellis, senior director of product marketing at Genesys discussed. Agentic orchestration, they argued, is the way to bypass this supposed dichotomy.
Many organizations, they pointed out, have accumulated AI deployment across departments that operate in isolation with different data sources, access and contexts. This "AI sprawl", DeChellis explained, is "scaling fragmentation, and your customers are going to start to feel it every day through inconsistency".
In these scenarios, agentic orchestration emerges as the solution. An underlying orchestration layer that governs how AI agents share context and guardrails across all channels and systems.
"It's a system that operates across the entire enterprise stack, coordinating the moments that matter for customers, employees and your business," explained DeChellis.
Tim Friebel laid out a framework for organizations beginning or scaling AI deployments:
- Start with low-hanging fruit, such as self-service automation for high-volume enquiries and agent co-pilot for summarization and knowledge surfacing.
- Run a production pilot rather than a proof of concept to understand how systems work when actually dealing with live traffic.
- Scale incrementally rather than trying to automate everything at once.
Agentic AI addresses the disparity between AI investments and business outcomes
Brian Mistretta, product marketing director and Karen Inbar, industry marketing director, both at NiCE flagged that, while most organizations focus on measuring and improving customer interactions, the real CX pain points often happen later on in fulfilment and back-office processes.
For example, Mistretta said: "A customer calls because an order is wrong. The agent handles the conversation perfectly and it's a great experience in the moment. But then the fulfilment process breaks down, the shipment still goes to the wrong address, and now the customer is frustrated all over again. The interaction succeeded, but the outcome failed."
This is down to disconnected AI.
Inbar pointed to three industry examples to demonstrate the difference between connected and disconnected AI. In retail, a loyal "platinum tier" customer has a delayed order before a big game. A contact center agent apologizes and cancels the order, not taking into account that the customer is a platinum member who has a store pick-up option. This is the fault of a disconnected system.
With connected AI, the system would cross-reference the customer's loyalty status and check all resolution options across systems proactively so that a human agent can call the customer with solutions to hand.
In healthcare, Mistretta and Inbar said claims rework costs around US$20 billion per year. Current claims rework processes take an average of 36.7 minutes per claim. When an AI agent takes over, that falls to 6.7 minutes. The AI agent can have natural language conversations with claimants and resolve 75 percent of cases autonomously. In this scenario, per-interaction cost drops from between $6-12 if a human handles it to $0.25-0.50 if an AI agent handles it.
As many as 64 percent of financial services customers who fail to get resolution through digital channels will go into a branch, having to re-start the conversation all over again. Agentic AI that connects policy with claims in a single interaction can eliminate hand-off complications such as this.
When Inbar and Mistretta asked the audience "What's the biggest barrier preventing end-to-end AI orchestration in your organization?", the greatest percentage at 41 percent answered "technology – our current stack can't connect these workflows".
Where teams can prove ROI from AI investments – and where they fall short
Linking CX initiatives to ROI remains a persistent challenge for CX practitioners, as uncovered in CX Network's CX Horizons: The Global State of CX 2026 report.
Monika Aufdermauer, VP of user success at Koho, shared the fintech's implementation story in conversation with Caitlin Coen from Kustomer.
Koho went live with AI-assisted support and was deflecting 30 percent of conversations within 35 days of launch. Aufdermauer explained that this was possible due to a strong pre-existing knowledge base.
She said: "Your AI is only going to be as good as the information it can read. If you don't have a solid knowledge base – something that your humans can refer to – your AI isn't going to be able to do that either."
She also pointed out that, while deflection rate is commonly cited as the key measure of AI success, it can often muddy the waters.
"Without having an understanding of what was true deflection and what was people leaving or people being frustrated, A) you're not solving those people's problems, and B) you're artificially inflating your actual deflection rate," Aufdermauer warned. She recommended measuring CSAT and layering QA to reach true understanding of the effectiveness of artificial intelligence (AI).
Given Koho's growth trajectory, Aufdermauer estimates that the support function would require approximately $4.8 million per year in funding without AI. With AI, however, the efficiency gains have allowed Koho to shift its remaining team from low-cost, offshore, tier one support to a more advanced support model with more highly skilled, better paid agents.
When asked about the impact on jobs, Aufdermauer said: "My opinion on AI is that it's not here to take away jobs. It's here to change jobs."
Discover more from Kustomer in the CX Network report The State of AI: From assistive to agentic
Changes in the vendor market and their implications for CX leaders
Bill Staikos is the founder of the consultancy Be Customer Led and has previously worked and consulted for brands such as Apple, BNY, Uber and Bank of America. Staikos joined the series to discuss the changing vendor market and what it means for CX leaders and strategic decision-making.
In 2025, 74 percent of practitioners surveyed by CX Network said they acquired new AI capabilities through third-party vendors. By 2026, that figure dropped to 50 percent. According to Staikos, this is less about vendors losing relevance and more about buyers becoming more selective.
Staikos argued that "AI vendor" is a broad term – perhaps now too broad. Instead, he pointed to five distinct categories:
- Model providers that are the foundation
- Cloud providers that control infrastructure and compute
- CXM platforms
- Customer intelligence systems, a more recent AI-native category focused on the customer record
- Point solutions that are purpose-built and for specific tasks
The CXO, Staikos said, doesn't need to be super technical but does need to understand which category or layer an investment sits in and the customer problem it solves.
Looking into the future, Staikos said that AI-native "customer intelligence systems" that capture signals continuously and infer meaning before recommending or autonomously taking actions – will be the next major market shift.
The practical implication of this for practitioners is that those who maintain a solely "survey and dashboard" focus will fall behind. Leaders must focus on connecting customer intelligence, AI capabilities, risk and operations with business value.
You can read more insights from Staikos on this topic
in this article he wrote for CX Network.
The future of customer insights is agentic twins
Customer awareness of how AI uses their data ranked as the top customer behavior trend for 2026, according to CX Network's research. CVS is using agentic twins – AI representations of hundreds of thousands of real customers based on consented behavioral data and interviews – to augment VOC data and gain deeper insights, faster, allowing for increased experimentation and faster, less risky CX innovations. Sri Narasimhan, VP and head of enterprise customer experience and insights, is leading the initiative.
This technology is distinct from the more-established synthetic data model as the twins are built on real, consenting customers. CVS created over 100,000 individual twins to represent customer populations.
Narasimhan said: "We've become more customer-centric by creating an army of robots. What we've done is amplified the consumer voice... I used to put an empty chair in meetings and say, 'that's the customer.' I don't need to do that anymore."
He listed four distinct advantages of using agentic twins:
- Speed: Studies that once took weeks now take minutes.
- Fidelity: The "say-do gap" that frequently occurs in customer research whereby customer feedback belies behavioral data is reduced as the twins look at scenarios rather than questions.
- Reach: Underrepresented populations can be more easily consulted via the twins model.
- Scale: CVS has (at time of writing) run around 2,500-3,000 studies – in 2026 alone – compared with the few hundred that would have happened using traditional methods.
When asked about practical deployments, Narasimhan shared that messaging and marketing testing, product launches and capital investment decisions were all utilizing the twins.
We then moved on to discuss validation. CVS runs three forms of validation: in-sample comparisons that check if the twins reproduce what was said in original interviews, comparisons against human panels which typically have an 85-95 percent match rate, and comparing the outcomes of the twins' predictions with real-world results.
"If it doesn't match, you lose all credibility internally. That's what we care about most," Narasimhan said.
Narasimhan's vision for the future of the agentic twins will move beyond research to adopt more of a testing and simulation focus. He plans on using the twins to model entire product launches and service changes before they happen.
"The vision is consumer centricity realized. Not just 'can we get to this panel?' We're bringing the representation of you into every meeting, every product decision, every service decision," he said.
Read more about CVS Health's agentic twins
in this CX Network case study
Bridging the gap between collecting and actioning data
There is a persistent and frustrating gap between having lots of data and actually doing something with it.
In the words of Ekaterina Mamonova, global head of broker proposition at Liberty Specialty Markets and a CX Network Advisory Board member, "adoption has really outpaced integration".
She explained: "We've become very good at running voice of the customer programs, mapping journeys, tracking metrics, but the system of decisions linked to budgets and incentives and operating models still runs through an incredible number of silos. Insight accumulates significantly faster than action."
She emphasized that metrics alone don't tell a compelling story; they also need a list of root causes, potential trade-offs and priorities that leadership can act on.
Patty Soltis, senior CX manager at Upwork, echoed this. She recommended understanding what matters most to the business unit being served then tailoring insight work to serve them.
"My role as a CX leader is to serve my biggest stakeholders, those decision makers that sit out there. The more I come at this from a servant leadership standpoint, the more I can help them find the friction that matters."
Mamonova set up a CX champion network across 25 countries. When asked how she made that work, she pointed to governance: "A champion network only works when governance turns insight into operating discipline. CX can gain visibility without authority, diagnosing problems but lacking control over the decisions that matter." The governance model includes consistent measurement, clear ownership and regular reviews. Soltis added that knowing when and how to celebrate wins keeps up momentum to sustain programs.
Designing equitable AI at scale
Kapil Poreddy, senior software engineering manager at Walmart Global Tec, argued that enthusiasm for personalization can obscure the need for equity in AI products.
"The through line for me is: does the system see the person, or is it just the transaction? What you quickly learn at that scale is that the customer is never homogeneous. How do you build one platform that works equitably well for all of them?"
He added: "I've never been able to build something at scale without asking: who are we accidentally leaving out?"
So, what does inequitable AI at scale look like?
Poreddy used an example from his Castlight Health days, where he built a population health platform serving millions.
"The people who needed the platform most, such as workers with more complex chronic conditions, lower health literacy, non-English speakers, were the ones least likely to engage with it. The cost isn't abstract. It's a preventive care visit that becomes an ER admission. Even a one-to-two percent gap in equitable engagement across millions of covered lives translates to a material health outcome difference." In this case, CX has a clinical consequence.
When asked about measurement of equity in AI, Poreddy introduced the Engagement Equity Index, which compares task completion rates for the lowest access segments versus highest access segments. He emphasized that "model scores are systematically lower for populations underrepresented in your training data".
He said: "This is the most underappreciated signal. I consistently see model performance reported in aggregate where the variance across subgroups would never pass a fairness audit. If you're not disaggregating performance metrics by population, you actually don't know what your system is doing."
Other warning signals are support volume being higher for specific segments – even while aggregate metrics look healthy – and opt-out and disengagement rates increasing for certain languages or regions.
We asked how to design for more equitable AI in CX, and Poreddy argued that the compromise between serving the majority efficiently while accounting for edge cases is false. He said: "Reducing friction for high-friction users improves outcomes for all users. It's like improving tail latencies – it helps everybody. Designing for the hardest case makes the median better. I call this the equity dividend."