Two announcements from Crexendo, seven days apart, that most of the market will read as a channel deal.
On 18 August, native vCon support at the platform level in NetSapiens — the open conversation container wired into the core of a platform licensed by more than 250 service providers supporting over eight million end users. On 25 August, a partnership with Tresic putting AI-native conversation intelligence in front of those same licensees to sell under their own brand, with customers already running it in production.
Substrate, then application. Read as a channel deal, it is unremarkable. Read as an architecture decision, it is the first serious attempt in our ecosystem to fix the thing that has been quietly stalling voice AI deployments for two years.
Those service providers did not just get an AI feature. They got somewhere to put what the AI worked out.
The pilot works. The second conversation is where it dies.
The pattern is familiar to anyone who has sat through a deployment review. The pilot performs — our AI Voice research found early deployments delivering operational time savings of thirty percent or more, and in some AI-assisted interactions booking and conversion improvements of up to threefold. Those numbers are real, and they are why budget gets released.
Then it goes wider, and something goes flat. The agent handles each call competently and remembers none of them. A customer who explained their problem on Tuesday explains it again on Thursday. A fraud signal detected on one call has no bearing on the next. Every conversation starts from zero, and the organisation ends up with ten thousand small, competent, disconnected performances.
The instinct at that point is to buy a better model. It is almost always the wrong instinct. The system is not failing to understand the conversation in front of it. It is failing to carry anything out of it.
The diagnosis was made on our own stage a year ago
This is not a new observation, and it did not originate with the AI wave. It is the reason vCon exists.
At CASA25, Thomas Howe of Vconic described the problem that produced the standard, drawn from running a business that spoke to a quarter of a million customers a month: every time a customer got in touch, it was like a new day. The customer remembered the previous conversation. The company did not. The internal project built to fix it was called Papa in a Box — named for the founder’s father, who had personally coached every sales call until the business outgrew what one person could hold in their head.
That is worth sitting with. vCon was not designed as an AI format. It was designed as a memory format, by people solving an institutional forgetting problem that predates large language models entirely. The AI use case arrived later and inherited the property that mattered.
Howe also set out the sequence from that stage: compliance and governance first, then AI and service creation. Twelve months on, that is precisely the order in which it has shipped — the container into the platform core, the intelligence layer on top of it.
A context window is not a memory
This distinction gets lost, so it is worth stating plainly.
A context window is short-term working space. It holds what is in front of the model for a single exchange, and it is discarded. Every generation makes it larger, and every generation of larger windows encourages the belief that the memory problem is being solved at the model layer. It is not. A bigger buffer is still a buffer.
Memory is an architectural property of the system around the model: a durable record of what was said and what was concluded, structured so another system can read it, indexed so the relevant part can be found, and retrieved before the next conversation begins rather than analysed after it ends.
Almost none of that lives in the model. All of it lives in the data layer the industry spent three years treating as plumbing — which is why a container standard, of all unglamorous things, turns out to be on the critical path for AI.
What memory actually requires
Three properties, none of them exotic, all of them missing from most stacks today.
A record that carries its own analysis. Not an audio file plus a summary in a CRM field nobody opens. A single record holding the parties, the dialog and the derived intelligence — transcript, tags, sentiment, extracted outcomes — so what the system concluded travels with what was said. That is precisely what vCon was designed to be, and it is why the Crexendo decision to put it in the platform core rather than offer it as an export option is the part worth noticing. An export format is a checkbox. The core is memory infrastructure for everyone on the platform, whether they asked for it or not.
Indexing by participant, not by call. A record filed under a call identifier answers questions about that call. A record indexed by who was on it answers questions about a relationship. Simwood, another partner building on the container, does exactly this — its Memory service surfaces prior contacts, unresolved issues and previously detected signals before the new conversation’s first word. That is a different product category from transcription, and the entire difference is in the indexing.
Portability, so the history outlives the platform. Memory that only works inside one provider’s system is a hostage, and it depreciates the moment a contract is renegotiated. Simwood writes its vCons to the customer’s own object storage, readable without Simwood in the picture at all. A carrier sits in the one position where conversation data could most easily be held captive, and has chosen a format that works without it.
The value moves to the accumulation
Here is the commercial consequence, and it is the part the market has not priced.
Transcription is finished as a differentiator. Accuracy is table stakes and the cost curve is heading where cost curves head. Analysis of a single conversation is close behind. What does not commoditise is the accumulated, structured, permissioned history of every conversation an organisation has had with a customer — because it cannot be bought, cannot be replicated by a competitor, and gets more valuable with every interaction added to it.
That is the asset. Not the model, and not any individual conversation. The compounding record.
Which is what those 250 service providers actually acquired this month, and probably not what they think they bought. The AI feature is the visible part. The durable part is that from now on, the conversations moving through their platforms are being written down in a form that will still be readable when the intelligence layer above them is on its third generation.
Everyone else in this ecosystem should ask the same question of their own infrastructure. Not whether to add AI features — that decision has been made everywhere. Whether today’s conversations are being recorded in a way that will still be worth something in five years, or thrown away in a slightly more sophisticated fashion than before.
Which makes trust load-bearing, not optional
The moment memory persists across conversations, the trust questions stop being compliance overhead and become the thing the product depends on.
A record that survives one conversation to inform the next contains everything sensitive an organisation holds about a customer: what they disclosed, what was inferred about them, which signals fired. Consent has to travel with it. Provenance has to be provable. Redaction has to work. Lineage has to show which conversations informed which model. A memory system without those properties is not an asset — it is a liability accruing quietly, which will present itself all at once.
The standards work reflects this. The vCon container itself is still a working-group document, but the provenance machinery underneath it — the SCITT architecture the lifecycle work proposes using to record consent and prove a record has not been altered — was published as a standards-track RFC in June. The trust layer is further along than the container it will secure. That is an unusual order of arrival, and a telling one.
This is why we keep saying Trust and Intelligence are not separate agendas. Persistent memory is where they become the same engineering problem.
Who will run the conversation
The question our March research put on the table was who ends up running the conversation — not who transcribes it or who models it, but who governs how it is executed and where the value accrues.
Six months of deployment evidence later, an answer is taking shape, and it is less about intelligence than the industry expected.
Whoever remembers the last one.
We are taking this into a closed room at CASA26
CASA26 runs in Amsterdam, 21–23 September. One session is given over to vCon and the conversation-data layer, with Crexendo and Tresic speaking. The agenda is the commercial model rather than the schema: what conversation data is actually worth, who captures that value, and what has to be true about trust before any of it holds. The open questions in this piece are the ones on that agenda.
CASA26 is open to CPaaSAA members and sponsors rather than general admission. If your organisation is building on this layer and is not yet in the room, talk to us about membership before September.
Governance note: Kevin Nethercott, CEO of Tresic, serves as non-executive Chairman of CPaaSAA and recuses himself from Alliance decisions involving Tresic or its closest competitors. Robert Galop, Tresic’s CPO, is a shareholder in CPaaSAA and holds no operational, editorial or governance role. Editorial decisions sit with the CPaaSAA executive team.
I've spent my career in IT and telecoms building bridges, from the early days of cloud communications to today's reset around APIs, CPaaS and AI.
As Founding Partner of The Next Cloud and CPaaSAA, I help telcos, CPaaS players and AI-native scaleups work out where they fit in the AI era, through strategy, ecosystem building and events.
My view: the value is moving from platforms to outcomes, and real innovation comes from ecosystems, not incumbents acting alone.
Based in Amsterdam, usually somewhere else.
Comments are closed