In the first five working days of November, four companies bought an AI agent.

Salesforce signed a definitive agreement for Cresta, at a reported $2.7 billion. PartnerOne bought 3CLogic. Dialpad and Ada agreed to combine. And Five9 agreed to buy Level AI for about $325 million in cash. I got the Five9 news by email on the Thursday evening, and my first reaction was not about Five9 at all. It was: that is four.

Then I looked at what happened one layer down in the same weeks, and the picture got a lot more interesting.

Four deals, one signal

Strip out the press release language and three of these deals are the same move. A company that owns the customer, a CRM or a contact centre, buys the AI that was running on top of it. The fourth, PartnerOne, is a long-term software holder, and it tells you something too: AI voice is now an asset worth owning outright. NICE did the first move with Cognigy. Genesys did it with Pinkfish. By CMSWire’s count, Cresta takes Salesforce past $15 billion in acquisitions this year alone. This month the pace simply picked up.

The reason is not mysterious. Contact-centre and cloud communications companies grow slowly; AI is where their growth comes from. Five9 reported AI revenue up 91 percent, to roughly 17 percent of subscription revenue, in the same week it announced the deal. When AI is your growth engine, you do not want to rent it from a partner who might be bought by your competitor next quarter. You buy it.

And you buy rather than build, because building up the stack is much harder than it looks. Salesforce has had Agentforce since 2024, and it still paid billions for Cresta to get voice agents that already work inside live calls.

In the same weeks, the hyperscalers handed back the phone line

Here is the part that did not make the deal headlines. Microsoft is retiring most of Azure Communication Services: it stopped taking new customers on 23 October, and most of the standalone services, SMS and email among them, end on 30 September 2028. What survives, calling and recording among it, has to run alongside Teams. AWS put Chime SDK PSTN Audio into maintenance mode at the end of September and closed it to new customers a month later, pointing everyone to Amazon Connect. And Salesforce, which has never run its own contact-centre telephony, is buying the agent and leaving the phone line to partners.

For Microsoft this is a pattern, not a one-off. It bought Metaswitch in 2020 and sold it to Alianza less than five years later.

Put the two lists next to each other. The companies with the deepest pockets in technology just told us which half of the call they want. They want the agent and the customer. They do not want to run numbers, carriers, caller ID, porting and compliance in a hundred countries.

Diagram: the agent was bought (Salesforce and Cresta, Five9 and Level AI, Dialpad and Ada, PartnerOne and 3CLogic) while the phone line, meaning numbers, carriers, caller ID and compliance country by country, is handed back by Microsoft, AWS and Salesforce.

I understand why. As we kept hearing at CASA26, this layer is messy. It is regulated, it is different in every country and sometimes in every state, and it has very little in common with the agent that runs on top of it. That does not make it less valuable. It makes it a different business, and the one thing an AI company cannot build in an afternoon. Tim Banting, one of our advisory partners at CPaaSAA, made the point after Infobip’s analyst event in July: the plumbing matters a lot more than most suppliers like to let on.

I called Five9 the org chart bet. It was step one.

I should be honest about my own record here. In July I wrote that NICE and Genesys were betting the balance sheet while Five9 was betting the org chart: a new leadership team and a new story, rather than a cheque. That was accurate on the day. It was also incomplete. The piece ended with a line about selective acquisitions, and I treated it as a closing thought. It turned out to be the next move. Org chart first, balance sheet second.

The detail I find most telling is in Five9’s own announcement. Level AI has worked with Five9 since 2022 and sat in the Five9 CX Marketplace. This was not a platform scanning the market and buying a pitch deck. It bought something its own customers had used, on its own marketplace, for years, and it spent roughly half its cash to do it. Tim Banting reads it the same way: a distribution and margin move more than a capability move. The marketplace was the diligence.

Every AI company on a platform marketplace should read that twice. Integration builds distribution and trust. It also shows the platform exactly what you are worth. The partners who integrate deepest are, whether they plan it or not, auditioning. And every other AI company in that marketplace now competes with the marketplace owner.

Neutrality is the new test, and nobody has passed it yet

Cresta runs inside live calls on the major contact-centre platforms. Salesforce says it will keep supporting the platforms Cresta customers already use. Five9 plans to sell Level AI standalone, including to customers on other cloud platforms. Everyone is promising openness.

Maybe they will deliver it. But an enterprise on a rival platform is now buying its AI from a competitor’s subsidiary, and it will judge the promise on three things: where its conversation data goes, whose roadmap the product follows, and how fast the integration with its own platform gets fixed when it breaks. Whether an intelligence layer can stay open while it is owned is the question this industry will spend 2027 answering. I do not know the answer. Neither, I suspect, do the buyers.

Consolidation at the top makes the open layer below worth more

This is where I think the market is reading the month the wrong way round.

The headlines say the value is in the agent, and the buyers are voting with billions. Fine. But software is getting cheaper by the month, and an agent that is worth $2.7 billion today will have twenty credible imitators next year. AI still handles only a small share of the world’s voice calls. The rest will not be won by four acquirers. They will be won by hundreds of AI voice companies, most of which nobody has heard of yet, building for one vertical, one language or one country.

Every one of those companies needs the same thing the hyperscalers just walked away from: a route to the phone network that is not owned by a competitor. Numbers. Carrier access. Trusted caller ID. Compliance, country by country. A platform it can get on without negotiating for a year. The more the top consolidates, the scarcer and more valuable that neutral route becomes.

The layer underneath has three bets, and has to pick one

So what should the operators and communications platforms in that layer do? I see three bets.

Buy your way up. That is what Five9, NICE and Genesys have done. It works when you already own a large customer base to sell the agent into, and when you buy rather than build.

Become the open marketplace. Be the place every independent AI company wants to plug into, and let them take you into the verticals. A big contact centre is a horizontal product. Bakeries, restaurant chains and clinics are not, and that is where most of the next AI voice companies will be built. In Europe it goes further. At CASA26, Dstny’s Daan de Wever pointed out that every country has its own systems in a sector like healthcare, so expect micro-verticals.

Be the best horizontal layer. Numbers, carrier access, trust and compliance, done better than anyone else, for everyone. Not glamorous. Possibly the most defensible of the three.

The CASA26 panel we publish this Tuesday argues exactly this. Sam Wilson of 8×8 says go deep into verticals or be chewed up by a hundred startup ants. Sinch’s Jesper Larsson wants to be the trusted platform others build their vertical answers on. Vonage’s Kunal Shukla asks whether we are over-indexing on verticals and not talking enough about value. Each of them is right for someone. The mistake is not choosing.

Whichever bet they make, the hyperscalers just handed that layer a business. Somebody should pick it up on purpose. It is also why, at CPaaSAA, we keep putting AI voice innovators, operators and platforms in the same room. That was the premise of CASA26 in September. This month made the case better than any of us could on stage.

What I would do this quarter

If you run an operator, stop waiting for the AI companies to come to you through an aggregator. Package numbers, voice, caller trust and compliance for them directly, and price it for a company with ten customers, not ten thousand.

If you run a platform, decide whether your marketplace is a shop window or a hunting ground, and say so. AI companies will read the Level AI deal as a signal either way. The platforms that win the independents will be the ones that can credibly promise not to compete with them. And if you are building your own agents, look hard at what Salesforce just did before you spend another year on it.

If you build an AI voice company, map your dependencies now. Which of your routes to the phone network could be owned by your competitor in twelve months? Have a second one before you need it.

If you are one of the buyers, the neutrality promise is now part of the product. Publish how it works, for data and for roadmap, before your customers ask.

In June I wrote that there is no third door: every midsize player has to buy its way up the stack or make itself worth buying. This month showed the doors in use. What it also showed is that there is a floor underneath all of them, and the biggest companies in the world just decided they would rather not own it.

Everybody is buying the agent. Somebody still has to carry the call.