London brought the heat in June, and so did the conversation.

We hosted an intimate CXO lunch with leaders from Hush, Bella+Duke, Charlotte Tilbury Beauty, DigitalGenius and others. The room was small on purpose, so the conversation stayed practical: what AI adoption and transformation actually look like inside retail and ecommerce businesses today, and where the real friction is.

We ran it as a three course dialogue: stories of the past, stories of now, and stories of what comes next. Here is where each course took us.

Starter: stories of the past

Every transformative technology has been misread. So we started by looking back at the last big shift each industry went through, and what the people in charge got most wrong about where it was heading.

The pattern that kept surfacing: leaders underestimated behaviour change and overestimated the timeline. Mobile was treated as a smaller website. Social was treated as a channel to broadcast into. In both cases the technology arrived roughly on schedule and the organisational rewiring took years longer, because the hard part was never the tooling.

The second question was sharper. The internet succeeded partly because it ran on open protocols nobody owned: HTTP, SMTP, TCP/IP. As AI becomes infrastructure, does it matter whether it is proprietary or open, and is that a question tech teams should be raising with the business? The room's view was yes, but framed commercially rather than ideologically: the risk is not philosophical, it is switching cost, pricing power, and who owns your customer data three years from now.

Main course: stories of now

The bottleneck has shifted. It is no longer writing code, it is knowing what to build and why.

When we asked what had actually become the real constraint, almost nobody said engineering capacity. The answers were decision-making, data quality, and the internal alignment needed to say yes to one thing and no to nine others. Building faster only helps if the organisation can decide faster, and most cannot.

The plumbing came up here too. Data that lives in six places, systems that do not share state, and workflows that still depend on a human copying a reference number between tabs. If the data is consistent, an agent can act. If it is not, the agent automates confusion at scale.

We also asked what the most revealing question someone senior had asked about AI in the last eighteen months, the one that showed how wide the gap really is. The answers were funny and a little uncomfortable, and they landed on the same point: the fluency gap between the board and the builders is now the single biggest tax on progress.

"[Insert quote from Frances on hosting the lunch]" - Frances, CX Collective

Dessert: stories of what comes next

There was a shared impatience with using AI simply to make existing processes cheaper. The bigger opportunity is redesigning the experience itself: anticipating a problem before it becomes a ticket, surfacing the right information before the customer has to ask, making the handoff between channels feel like one continuous conversation.

As automation absorbs the repeatable, the remaining human touchpoints get higher stakes. A conversation with support stops being a cost to resolve and becomes the moment a brand either earns trust or loses it.

Teams are evolving faster than org charts to match. The lines between CX, engineering, data and product are blurring, and the best progress in the room was coming from small cross-functional groups shipping together, not steering committees reviewing roadmaps.

What to take back to your board

Three small bets almost any team can start now. First, audit the top ten reasons customers contact you and separate decisions that need judgment from problems that only need information. Information problems are the low-hanging fruit for agents.

Second, get explicit about your dependency on proprietary AI infrastructure. Know what it would cost to move, and make that a board-level conversation before it becomes an emergency.

Third, pick one integration that currently hurts customers and fix the data model behind it. It will not be the headline project, but it will teach you more than any proof-of-concept demo.

"[Insert quote from a guest leader]" - A leader in the room