VCONify came from a pretty simple realization: businesses are spending enormous amounts of money on AI, but much of the conversational data they want that AI to learn from is still fragmented across call recordings, transcripts, emails, chats, CRM notes, and different vendors.

My co-founder Jeffrey Elletson and I saw the opportunity to build the infrastructure layer around the emerging IETF Virtualized Conversation (vCon) standard. A vCon turns a conversation into a structured, portable record that can include the interaction itself, the participants, chronology, metadata, attachments, analysis, and provenance.

What makes us different is that we’re not trying to be another call recorder, transcription service, analytics dashboard, or AI agent. We sit underneath those applications. Our job is to make conversations durable, portable, governed, and usable by whatever humans, systems, or AI models need them next.

I sometimes describe it as the difference between having a pile of documents and having PDFs. We’re trying to give conversations that same kind of standardized, reusable existence.

Today, most of it comes the old-fashioned way: relationships, referrals, partners, and conversations that start because somebody in our network has a problem and realizes what we’re building might solve it.

LinkedIn and the thought leadership I’ve been writing around vCons and conversation infrastructure also create conversations for us, but we’re still early enough that education is a big part of selling. Very few buyers wake up in the morning saying, “I need conversation infrastructure.”

Where I’d like more of our business to come from is people discovering the category because they’ve already recognized the underlying problem. They’ve got five systems containing five different versions of the same customer conversation, or they’re trying to give an AI system trustworthy organizational memory, and they start looking for a better architecture.

That’s a much better (and faster) conversation than convincing someone a problem exists.

Pattern-finding is something I’m very comfortable handing to AI.

Give it enough conversational data and it can surface commitments, recurring objections, unanswered questions, follow-up items, inconsistencies, and risk language far faster than a person could review the same volume manually. That’s exactly the kind of work machines should be doing.

What I don’t want AI doing by itself is deciding what those things mean or what should happen next.

We’ve become more conservative about that boundary, not less. AI can tell me, “Here are the twelve conversations where somebody appears to have made a promise.” I still want a human deciding whether it really was a promise, whether it matters, and what action should follow.

The more consequential the decision, the less interested I am in removing the human judgment from it.

Absolutely on the first part. Prospects are much more educated before the first conversation than they were even a year ago.

The questions have moved from “What can AI do?” toward things like, “What data is it using?”, “Where did that answer come from?”, “Can I move that context between systems?”, and “Can I reconstruct why this decision was made six months from now?” That’s a very healthy change for us because those are infrastructure questions.

As for people finding VCONify directly through ChatGPT or another LLM, I don’t yet have enough attributable data to claim that as a meaningful acquisition channel. I’m sure AI-assisted research is happening, because that’s increasingly how all of us research unfamiliar categories, but “I’m sure” and “I can prove it” are different things.

I expect that distinction to disappear pretty quickly. Being understandable to AI systems is becoming another form of being discoverable.

I’d tell them to spend less time chasing the newest model and more time fixing the data and context, the “robot food,” those models are being asked to work with.

Models will keep getting better and cheaper. That’s the part you can almost take for granted. Your proprietary advantage is more likely to be the quality of the information you can give them: complete records, useful context, permissions, provenance, history, and the ability to know where something came from.

I’d also start preserving important AI interactions themselves. If an AI recommendation influences a customer, employee, financial, or operational decision, “the model said so” is going to become an increasingly inadequate answer.

What would I ignore? Most predictions about which model, agent framework, or AI company is going to “win.”

Twelve months from now, that leaderboard will look different again. Your architecture shouldn’t have to.