The best line of this week’s excellent inaugural Cavell CX Summit in London wasn’t a prediction. It was an accounting question.

Sam Wilson, CEO of 8×8, opened his keynote by refusing to give the talk everyone expected. “This isn’t a talk about AI demos,” he said. “It’s about who captures the margin.” He had fifteen minutes, and he closed on the same slide he started with — mind the gap, that’s where the margin gets won. Everything in between was an argument about where the money goes when AI stops being a feature and starts being the product. Most speakers took half an hour to say less.

I’ve been making a version of that argument for two years. It was instructive to watch a public-company CEO make it on stage, to a room of his own competitors, and to watch the room agree.

So let me take it further than he could.

The interaction was never the asset

For thirty years, communications was a volume business. You carried the call, the message, the session — and you billed for carrying it. The platform that carried more, cheaper, more reliably, won. CPaaS, UCaaS, CCaaS: different acronyms for the same logic. Move the interaction, take a margin on the movement.

AI breaks that logic, because the interaction is no longer the scarce thing. Carrying a conversation is becoming free. Understanding it — deciding what it means, what to do next, and doing it — is where the value now sits. The profit pool is migrating from carrying interactions to orchestrating the intelligence that runs them.

Wilson said it in customer language, and it was the sharpest slide of the day: customers don’t buy categories. They struck UC, CC and CPaaS through on screen and replaced them with the four things people actually pay for — resolution, revenue, trust, and lower cost to serve. Nobody, he pointed out, ever bought “a CCaaS.” They buy a problem getting solved. The categories were invented to organise sales teams and fill analyst quadrants.

This is the thing the industry keeps avoiding: we sell platforms, and customers buy outcomes. AI is going to make everyone confront the gap between those two sentences.

Horizontal reach, vertical margin

Here’s where it gets interesting, because two of the strongest talks of the day appeared to contradict each other — and the contradiction is the whole story.

Wilson: “Outcomes are vertical.” Stop talking horizontal platforms; go deep where you have an unfair domain advantage — mortgage origination, hospital intake, auto dealerships — because specialisation compounds and a generic agent loses to a specialised one.

Zoom, the same morning — Chris Morrissey, who runs CX go-to-market there: CX is becoming a horizontal strategy, not a business unit. It stops being the contact centre’s job and becomes something that spans product, marketing and sales — one intelligence layer across every channel.

Both are right, and saying so is not a fudge. The operating model goes horizontal — the intelligence has to flow across the whole organisation, not sit locked in a contact-centre silo. But the margin stays vertical, because the outcomes worth paying for are industry-specific and that’s where specialisation defends a price. Horizontal is how the data moves. Vertical is where the money is made.

And between the two sits the layer nobody has won yet: orchestration. The control plane that takes a horizontal flow of conversation and turns it into a vertical, completed outcome. Zoom drew it without naming it — conversation, to decision intelligence, to agentic execution. That middle box is the prize. Whoever owns where intelligence gets deployed, governed and monetised owns the margin. Everyone else is back to carrying interactions for a thinning fee.

Per-seat is dying, and nothing has cleanly replaced it

There’s a commercial problem hiding under all of this, and the summit circled it all day without quite landing it: if the interaction is becoming free, how do you bill for it?

The old model was per-seat — predictable, easy to budget, easy to renew. AI breaks it. Consumption pricing is the obvious successor — pay for what you use — but it imports the thing customers hate most: token costs are volatile, cloud bills swing month to month, and “use it and pay” becomes “we can’t forecast next quarter.” Cavell’s own research put cost unpredictability among the top barriers to AI adoption in the contact centre. People will pay more for a number they can plan around than less for one they can’t.

Which is why outcome-based pricing keeps surfacing as the destination — charge for the resolved ticket, the booked appointment, the closed sale, not the minutes or the tokens. It’s the right idea and the hard one, because someone has to define the outcome, measure it, and own it when the AI gets it wrong. And here’s the tell: outcomes are easiest to define inside a vertical. “A resolved mortgage application” is a priceable thing. “A unit of CX” is not. The pricing model and the vertical strategy turn out to be the same move — sell a defined outcome and, as more than one speaker put it, the customer stops arguing about price. The margin problem and the pricing problem have the same answer.

The tell: the work is human, not model

The most useful number at the summit came from Cavell’s own research, not a vendor deck. In a survey of more than 800 contact-centre decision-makers across five markets, almost none — barely over one percent — said AI had failed to deliver clear value. The technology works. That argument is over.

So why does it still feel hard? Because the same research puts the real barriers in plain sight: integration with legacy systems, data readiness, governance and compliance. Not the model. The plumbing around the model.

You’ll have seen the counter-headline — the MIT figure that “95% of enterprise AI pilots fail.” Ignore it, or rather, understand what it actually measures. Most of those pilots were never scoped to reach production. They were experiments designed to kill bad ideas cheaply and scale the one that works. A trial ending is the process succeeding, not failing. The number counts learning as loss.

Put the two findings together and the picture is clear. Capability is proven. Deployment is where it stalls. And the gap between “the AI works” and “the outcome ships” is filled with integration, data engineering and trust — human and organisational work, not model work.

That gap is the margin. And it is, overwhelmingly, partner work.

Five9 was the most honest about this. George Wilson coined a useful term on stage — humantic data: the record of human judgment and outcomes that becomes the standard for agentic processes. Strip the branding and it’s the argument I’ve made about voice for a year — the value isn’t the conversation, it’s the captured intelligence the conversation leaves behind. Then Five9 did the commercially revealing thing: they launched a partner enablement programme, AI-only, invite-only, built entirely around helping partners do the integration and deployment work that the platform alone can’t. When a vendor productises the partner channel, they’re telling you where they think the hard, valuable, defensible work lives.

The middle ground is gone

If outcomes are vertical, and the margin sits in orchestration, and the real work is integration and trust — then the strategic position of the horizontal platform gets squeezed from both sides. And Wilson named the threat precisely: it isn’t the vendor across the hall. It’s the company that didn’t exist eighteen months ago, the one that didn’t start horizontal — it started with a single workflow, in a single vertical, where the margin actually is.

That leaves incumbents two honest options. Find the capital to acquire your way into vertical depth and orchestration, and stay relevant. Or dress up for a sale while the multiple still holds. There is no comfortable middle, because trust evaporates the moment you’re visibly a step behind — and in this market, being a step behind is now measured in quarters, not years.

Compliance and sovereignty, incidentally, stopped being the boring panel. Wilson filed them under trust — “what a regulated buyer can’t outsource to a demo.” That’s not a cost centre. In regulated verticals it’s a moat, and it’s exactly the kind of vertical, defensible outcome the margin is migrating toward.

So: mind the gap. Not the gap in the demo — the demos are fine. The gap between what AI can do and what actually gets deployed, governed and paid for. That gap is where the next decade of margin in this industry will be won or lost.

The platforms that understand it will move up into orchestration and out into verticals. The ones that don’t will keep optimising the price of carrying an interaction nobody is willing to pay much for anymore.

The conversation is becoming free. The intelligence that runs it is not.