What does the future of customer insights look like in an AI-driven world?

CX Network's All Access: The Future of Customer Insights 2026 brought together leaders across insights, CX, journey management, research, product management and design to address a major question: what does the future of insights look like in an AI-driven world?

In short, artificial intelligence (AI) is only as trustworthy as the human oversight and cross-checked data it runs on.
Across sessions, certain themes appeared time and again:

Polished AI output can nurture a false sense of confidence, a risk noted by several speakers.
Because of this, our faculty noted that human oversight remains critical and validation is essential.
Combining data sources is part of this, something that was raised in almost all sessions.
Several speakers also noted that one of the most effective ways of proving value, of either AI or the insights or CX function itself, is to point out the consequences of not using the research, insights and CX teams in projects.

If you didn't catch the sessions live, you can watch them on-demand on CX Plus here. Alternatively, this article rounds up the key themes and top takeaways from the event, session by session, including:

Panel discussion: How insights teams can own AI internally – and where the real value of synthetic data lies

Takeaway one: Match the level of due diligence to the risk of the use case

When asked how insights teams can navigate the "speed versus due diligence" challenges faced by insights teams who can be perceived as blockers when it comes to implementing AI and synthetic data, Bill Staikos, founder and chief customer officer of Be Customer Led, said: "I think due diligence becomes an issue when you apply the same level of due diligence to every use case… And then that just slows you down, frankly."

Instead, he suggested a three-tiered model for applying due diligence to insights:

1	Low risk	Use pre-approved processes
2	Medium risk	Use controlled pilots and compare synthetic with real data
3	High risk	Requires human approval, monitoring and deep governance

Jack Austin, commercial insight lead for consumer strategy and insight at Virgin Media O2, added that insights teams should act as internal consultants, deciding which tiers to place requests in. He said that, for low-risk requests, insights teams should be able to consult their back catalogues and perhaps AI, and provide an answer on the same day.

Takeaway two: Acknowledge the fear of AI and frame efficiency savings as reinvestment opportunities rather than staff cuts

Many insights teams feel threatened by AI, which can lead to reluctance to adopt. This, ultimately, will leave teams behind. Staikos pointed to an example of efficiency savings being reinvested into engineers and salespeople, changing how staff perceived the AI-induced savings.

"The company that takes a saving, but doesn't fill it up, hold back up, is the company that has run out of ideas," he warned.

Austin reframed AI efficiency gains as an opportunity for insights teams to focus on more interesting and challenging work – that can't be outsourced to AI. "If an AI tool can help us answer 70 percent of the question, we can now really focus on that very interesting, new, shiny 30 percent that's going to make all the difference in what we share back."

Takeaway three: Insights teams should claim ownership of AI tools and show clearly what happens when they are not consulted

Austin warned that, in organizations where business intelligence (BI), research and data are separate functions, other teams may race ahead with AI using their raw data. He recommended staying ahead by embracing innovation, saying: "Get your pieces on the board. If you have to go and knock down the door, do it. If that isn't possible with structure, with governance, with accountability, then buy something in. Talk to an agency, talk to others in the industry that are doing this."

Staikos reminded the audience that anyone in any team can now use AI to produce research summaries, so insights professionals must be visibility accountable for research rigor.

"Don't be afraid to show examples where your skill set wasn't used, and where it was used, and the difference in those outcomes," he advised.

Takeaway four: Synthetic data is for early-stage experimentation and ideation but should be centrally owned by insights

Synthetic data and its pairing with LLMs has been touted as a huge step forward in customer insights as it can democratize and ease access, and facilitate experimentation at a scale previously unreachable. It can also plug the gaps in a company's own data, increasing representation of typically hard-to-reach demographics.

While Staikos is a self-proclaimed "huge proponent of synthetic data", he acknowledged its limitations, warning that it can "over-represent dominant groups [and] flatten differences between groups" and can "sound more reliable than it is in many cases". He noted that synthetic data "should not independently establish market size or customer prevalence".

Austin noted that stakeholder access is acceptable for ideation as long as the synthetic data tools are "still owned and disseminated from a central Insights function, so we know what ingredients they're cooking with".

Staikos said "I actually think… synthetics are going to save the CX team. And give them a seat at the table in a very different way than they have been in the past."

Five9: Where to start with QA automation

Takeaway one: Decide the outcome you want from QA automation first

Gautam Mourya, senior product manager and Suraj Khanna, lead product manager, both at Five9, advised teams to decide what they want from quality assurance (QA) automation before moving forward with implementation. Mourya gave a few examples: agent coaching, evaluating all calls against certain criteria and reducing agent disputes.

Takeaway two: When moving away from spreadsheets, pilot by automating only the most important quality questions to prove initial ROI

When an audience member asked about moving away from legacy spreadsheet scorecards, Khanna advised that "the first step would be to identify the most significant quality questions that you have, which are very important for your organization." Some examples he gave were compliance, revenue and upsell. Once this first automated use case has proven ROI, teams can scale automation into others.

Takeaway three: Use full interaction coverage to find the oft-missed good calls as well as bad

Mourya noted that with the sampling approach used by most organizations currently, both excellent and problematic interactions are missed. The full coverage provided by automation means that both can now be surfaced. "If an agent is doing something exceptionally well, that interaction can become a coaching example for others," Mourya pointed out.

Heineken: The best way to implement LLMs into the insights function

Takeaway one: Vet vendors on security and don't dismiss smaller firms

Narek Garit, global commercial analytics manager at Heineken, recommended that audience members in the market for AI should ask vendors about their data architecture, security measures, lead and attack testing reports and compliance certificates. However, he cautioned against dismissing smaller companies in favor of larger, more established ones, "I do not want to say you should always go to very well-known agencies, because that undermines the beauty and the power of many small companies."

Takeaway two: Catch bad data before it goes in and test outputs

Garit said that, while LLMs are good at summarizing text, using them to extract numbers from unstructured reports can be risky. Whatever an LLM produced will be polished, so he warned "the first challenge here is not only the "rubbish in, rubbish out" rule but identifying that the LLM's output is rubbish, because the output will always be beautifully written."

Takeaway three: Begin with your business objective

Garit recommended beginning with the business objective and feeding this objective to your LLM to build use-case specific tools. He also advised that VOC work should be judged against business objectives and brand KPIs rather than a forced sales ROI. He noted that social media data has random peaks and troughs according to buzz moments, which can make ROI look low.

Hyatt Hotels: Proving the value of research and UX

Takeaway one: Use research to check that you're building the right thing and put designs in front of real users before building.

Paul Weaver, vice president of global digital product design at Hyatt Hotels, gave the audience a poignant example of what can happen when research isn't included in design processes.

The example was from his time designing Disney's Epic Mickey video game, in which a 21-second puzzle he and his team designed ended up taking children players 43 minutes to complete. His later example, a fictional product he made with AI, called "Cardiopath", met every requirement on paper, but "still shipped untested assumptions", leading to an 18-step login process combined with a 15-minute timeout, for example.

The point of these examples was to warn the audience what happens when design plows ahead without consulting research first.

Takeaway two: Make the business case for research by pricing reworks in sprints

Weaver presented his "user experience dartboard", showing that each sprint – with six engineers, a product manager and quality assurance included – costs around US$10,000 per person, per month, which totals roughly $40,000 per sprint.

"If you do a feature, and you get it right first time, that feature costs you $40,000. Great, job done, move on," he said. However, "if you have to go back and fix something, that's going to take at least one more sprint. So that's another $40,000."

Including research in sprints makes the initial sprint more costly to begin with, but significantly reduces the need for the far more costly fixes. "Not understanding your user and having to go back and fix things after the fact gets really expensive really quickly", Weaver warned.

Takeaway three: Translate UX improvements into operational results

Translating UX improvements into operational impact can help to proselytize the discipline in an organization. Doing tasks faster means more productivity, fewer errors means fewer support requests and less friction means less training and improved staff retention.

Lloyd's Banking Group: Customer journey maps as live insights

Takeaway one: Map journeys as a hierarchy but start at top levels or with one part of the business

Louise Williams, customer lifecycle management lead at Lloyd's Banking Group, recommended mapping customer journeys as a hierarchy:

Level 1	Goals
Level 2	Lifecycles
Level 3	Journeys
Level 4	Journey sections

She advised the audience to begin with the top one or two levels, or to begin with just one part of the business. "My advice would be to perhaps go level one, and then level two… your level threes will then appear as you start to understand more about your level two journeys," she said.

Takeaway two: Give each cross-silo journey one senior owner who can make decisions

"The customer doesn't care that one director owns one part of the journey, one director owns the other. The customer just wants a seamless experience," Williams said. To achieve this, she recommended that each department-spanning journey be assigned a senior owner. This person should be chosen based on "who's got the most skin in the game" and has a "strong, influential voice". Although it's not necessary to do this for every journey immediately, it is wise to aim for it.

Williams gave an example from her previous role at TUI, the world's largest multinational leisure, travel and tourism company. Flight time changes were handled cross-departmentally through airline commercial, operations, comms and customer services. To ensure this worked smoothly, the commercial director was made accountable for every step's KPIs, because it was the commercial business that benefited from flexible flight networks.

Takeaway three: Treat journey maps as data models

Experience, behavior and value (EBV) metrics should be added to journey maps and updated regularly, Williams advised. She recommended showing trends alongside when changes are rolled out. Lloyds uses EBV metrics in a journey tool with around nine months of data and markers to show when new products and features are released.

"Most organizations have loads of insight and loads of data. If you don't put it against something like a journey, it's just data," she said.

Takeaway four: Set up regular journey reviews

Williams recommended setting up (ideally) monthly journey reviews, creating either a dedicated meeting or securing a slot in an existing leaders' meeting. At TUI, retail, airline and contact center leaders were meeting monthly to review performance and agree next steps, she shared.

While at Argos, she had a regular slot in an existing board meeting. "Whether it's a dedicated meeting, or whether you try and get into a meeting that already exists, it's really important that people understand performance and agree what will happen as a result," she said.

Takeaway five: Work with your organization's existing structure rather than trying to reorganize around journeys

At Lloyds, Williams needed to influence around 500 experience designers and 1500 others involved in product management. At TUI, teams were aligned through the "Makers of Happy" customer centricity program. "Don't try and reinvent the wheel. Don't try and change the business massively. Just try and think, well, how is the organization working now, and how can we work with it rather than against it?" For accountability there should be top-down OKRs: "You can't just show thousands of frontline colleagues a journey map. It doesn't work."

Uber: The questions to ask before implementing AI into customer insights

Takeaway one: The three checks to make before trusting AI-generated insights

Ekaterina Mironova, CX program manager at Uber, recommended checking three things before treating AI-generated insights as factual:

Can you trace it back to the original feedback?
Do you know who gave the feedback and in what context?
Does this come up across more than one source?

The example sources she gave for cross-checking were surveys, support conversations, repeat contacts, product behavior and search logs. She also discussed this in detail, in this pre-event interview.

Takeaway two: The three questions that decide how much human oversight is needed in each AI task

Mironova said that practitioners should ask three questions to understand how much human review an AI task needs:

What are we asking the AI to do?
What will it cost if it's wrong?
How easy is it to reverse the decision once the AI has made it?

"The main goal," Mironova said, "is not to put a human inside every single step, but to use human judgment at the moments where context, responsibility, or consequences matter most."

Organizing information, for example, needs only light human review. Explaining meaning or recommending decision with irreversible or hard-to-reverse consequences requires stronger oversight.

Takeaway three: The four steps to prepare your knowledge base for AI

Mironova outlined four steps to get a knowledge base in order for AI:

Inventory it: usage, last review date, owner
Prioritize articles by contact volume and sensitivity
Test a representative sample with the AI you're planning to implement
Use genuine failures to direct data clean-up

She outlined a scenario in which 400 articles have been written – for humans. AI can struggle with this due to inconsistent terms, missing regional differences and scattered information.

"The very first thing I would do is create a basic inventory," she advised. Then, "the idea is to improve the most important content first, and use real AI failures to guide the next stage."