We spent nearly a million dollars so the answer could be “nowhere.”
Somewhere in your office, in the last twelve months, somebody turned on an AI feature. A transcription add-on. Smart summaries. A note-taker that quietly joins the call and emails everyone afterwards.
It probably worked. It probably saved someone twenty minutes.
Nobody asked where the recording went.
That question is where a private AI phone system starts: not with what the AI can do, but with where the conversation goes when the AI starts doing it.
That changes the question from “what can this feature do?” to “where does this conversation go while the feature does it?”
That is not carelessness. The question simply does not occur to you. You bought a feature, the feature did the thing, and the plumbing behind it is somebody else’s problem. Which is very much how the plumbing prefers it.
But if you run a collections floor, a medical practice, a law firm or a dispatch desk, the plumbing is your problem. A call recording in your world is not a convenience. It is a discoverable record, an FDCPA exhibit, or protected health information. And for most AI features sold by most communications vendors, the honest answer is this: your consumer’s conversation left your building, left your vendor’s building, and was processed on hardware owned by a company you have never spoken to.
That path is not background plumbing. It is part of the product.
“Every call a business has is full of intent, sentiment, and signals that used to disappear the moment someone hung up.” — Greg Schreiber, CEO and Co-Founder
The default nobody actually chose
Here is how a typical “AI-powered” phone feature works, in plain English.
Your call is recorded. The audio file is uploaded to a third-party AI provider. It is transcribed and analysed on their hardware, in their data centre, under their retention schedule. The summary comes back a few seconds later, looking like magic.
The alternative is designed around a different default: keeping that processing path inside infrastructure you can actually account for.
Somewhere in that round trip, a consumer’s account number, a patient’s chief complaint, a debtor’s payment arrangement or a privileged conversation with a client became a file on infrastructure you do not control. The terms governing it live in a data processing addendum, in an appendix, behind a link. You have not read it. Your compliance officer has not read it either.
This is not a scandal.
It is the default, which is arguably worse, because nobody had to decide it.
Most AI risk works this way. It is not a black swan swooping in from nowhere. It is a flock of entirely predictable grey ones. A transcript sitting in a jurisdiction you never approved. A retention window nobody set. A subprocessor added to a list you were notified about by email in March. An examiner asking where protected health information is stored, and three companies standing between you and the answer.
This approach does not make those questions disappear. It makes the answers easier to trace.
So we bought the hardware. Then we wrote the software.
The alternative to sending your conversations somewhere else is not sending them anywhere. That sounds obvious. It is also expensive, which is why almost nobody does it.
Building a private AI phone system means owning more than the part customers can see. It means thinking about the hardware, the voice network, the recording layer, the transcription engine, and the systems that turn the conversation into something useful.
Vaspian has put nearly a million dollars, about $930,000 if you want the invoice-level truth, into private GPU infrastructure dedicated to AI workloads. Machines we own, in a building we own, in Buffalo.
That hardware is one of the reasons a private AI phone system can keep more of the processing path under direct control.
Hardware on its own would only get you halfway.
Plenty of companies own servers and still hand the interesting part to somebody else.
So the rest of it is ours too. We rebuilt the voice network: the Kamailio session border controllers every call passes through, the RTP media handling, the FreeSWITCH recording layer, all running as a virtualised stack our own engineers operate. We wrote the softphone for desktop, mobile and WebRTC. We wrote the portal. We wrote the REST API and the webhooks that give your developers programmatic access to call events, recordings, transcripts and summaries. And we replaced the speech-to-text engine we used to license from an outside vendor with an ASR stack we built ourselves.
A private AI phone system only works as privately as the layers underneath it. Owning one layer and outsourcing the rest does not answer the question this article started with.
On top of that sits BlueMesh, built with Netstratum rather than bought from them. Their orchestration platform, our voice data, and two engineering teams working on it daily to make it survive contact with regulated industries.
Which means there is no seam. From the moment a call hits our session border controller to the moment a summary appears in your portal, every layer in between belongs to somebody who answers a Vaspian phone. No integration tax. No finger-pointing between the voice vendor and the AI vendor when a transcript goes missing. No third-party model API call touching your customer conversations, no new vendor to onboard, and no new business associate agreement for your counsel to review.
That continuity is what makes the private AI phone system argument meaningful instead of cosmetic.
Most companies selling AI for phone calls own one layer of that stack and rent the rest. It is a perfectly reasonable way to build a business. It is just not a way to promise anybody that their data stays put.
It is ultimately a stack question, not just a feature label.
How it actually works
This is the part most vendors skip, so here it is properly. Four stages.
A private AI phone system only matters if the privacy argument survives every stage of that process. The important question is not simply where the final summary appears. It is what happens to the conversation from the moment the call begins.
That means the system has to be understandable from the first packet of audio to the final structured insight.
1. The call happens
A consumer calls your collections floor. A patient calls to reschedule. A shipper calls dispatch about a missed pickup. An agent answers. Nothing about the conversation changes because AI is involved, which is rather the point.
2. Identifying information comes out first
Before anything is analysed, the recording passes through a webhook that strips personally identifiable information. Card numbers, Social Security numbers, dates of birth, account numbers, protected health information. Redaction happens at the front of the pipeline rather than being bolted on at the end. A model cannot leak what it was never given.
For a private AI phone system, redaction at the front of the pipeline keeps sensitive information out of later stages that do not need it.
3. Transcription runs on our GPUs
The audio is transcribed by our own ASR on the hardware described above. Stereo recording keeps the agent on one channel and the consumer on the other, which is why the transcript can attribute who said what rather than producing an undifferentiated wall of text. That distinction matters enormously the day you need to prove the agent delivered the mini-Miranda before the consumer started talking.
Running transcription this way keeps the private AI phone system tied to infrastructure Vaspian operates rather than handing the audio to another model provider.
4. BlueMesh turns the conversation into data
BlueMesh is the orchestration layer. It decides which models run, in what order, with what guardrails, and where a human needs to sign off. Netstratum built it as a private AI platform: deploy agents and workflows on private, on-premises or hybrid infrastructure, with local models and human-in-the-loop control rather than a black box that answers and hopes.
Inside a private AI phone system, that orchestration layer determines how the conversation moves from audio to structured business information.
It is assembled from a handful of specialised engines. Echo AI handles voice and text conversation. Privy AI is the private search layer, so your own material can be queried without any of it leaving for a public model. Vision AI turns documents into structured data. Lingo AI covers translation and multilingual calls.
What comes out the other side is structured. Who called and why. The disposition. What was promised and by when. The next action. Sentiment across the call rather than one blunt score at the end. And whether anything on it needs a compliance officer’s attention. A right-party contact becomes a flagged record. A promise to pay becomes a field with a date on it. A conversation goes in, a row of usable business data comes out.
Side by side
- The recording — Uploaded to a third-party cloud — Stays on the Vaspian PBX
- PII and PHI — Sent as-is, redacted downstream if at all — Redacted at the webhook, before analysis
- Transcription — Someone else’s model, someone else’s hardware — Our ASR, our GPUs
- Storage and retention — Their cloud, their retention schedule — Vaspian and Netstratum private cloud
- Paperwork it creates — A new subprocessor, a new DPA, possibly a new BAA — None. It runs under the agreement you already signed.
- Who else benefits — Possibly the model. Possibly its other customers. — You. That is the whole list.
Responsible AI is mostly a list of things you decided not to do
Every company with an AI product now has a page about responsible AI. Most of them are adjectives. Here are four decisions instead, including the ones that cost us something.
For Vaspian, a private AI phone system is part architecture and part restraint. What the system does not send, share, automate, or decide matters just as much as what it can produce.
We do not train on your conversations.
Customer data is never pooled and never used to train shared models. Your calls make your business smarter. They do not make our other customers’ products smarter, and they certainly do not improve somebody else’s model.
That boundary is part of what a private AI phone system is supposed to preserve: your conversations should serve your business, not become training material for somebody else.
Redaction happens before analysis, not after.
Sequence is a policy, not an implementation detail. Anything else is cleaning up after a spill you chose to have.
In a private AI phone system, the sequence matters because privacy decisions have to happen before unnecessary exposure, not after it.
Post-call before real-time.
We could ship live in-ear agent assist and demo it beautifully. We are not going to, until it is accurate enough that a compliance officer would stake their job on it. An AI that whispers a mini-Miranda reminder after the agent has already skipped it has not prevented a violation. It has documented one.
A human stays in the loop.
BlueMesh flags, ranks, summarises and surfaces. It does not close your tickets, score an agent into a corrective action, or decide whether a dispute is valid. Human-in-the-loop control is built into the platform because judgement is still a human job.
A private AI phone system can surface the information, but the system does not need to pretend judgement has stopped being a human responsibility.
“We are doing that in a sequence we can stand behind, not racing to ship something we can’t support.” — Greg Schreiber, CEO and Co-Founder
Nothing on that list is thrilling. That is rather the test.
If you are already a customer, this is the interesting part
Some of you have been running your calls through Vaspian for a decade. Every one of them carried, connected and, if you are on recording, kept.
For an existing customer, a private AI phone system starts with conversations that are already part of the phone environment rather than requiring an entirely separate communications stack.
That is not an archive.
It is an asset nobody has opened yet.
Switched on, it becomes this. Every call searchable by account number, matter, claim, client name, keyword or phrase. Summaries instead of listening. Sentiment and trend across a whole floor instead of a supervisor sampling one call in a hundred and hoping the other ninety-nine were clean. Plain-language questions, of the “what did the Harrison account actually agree to in March” variety, answered in seconds rather than over an afternoon.
That is where a private AI phone system stops being an infrastructure story and becomes an operational one.
On what it costs, we would rather be specific than vague:
- Call Recording, $10 per seat per month. Capture everything, transcribe on demand, search it later.
- Call Recording Plus, $20 per seat per month. Everything transcribed automatically, webhooks, three-year storage.
- Conversational Intelligence, $50 per seat per month. Live dashboards, action-item extraction, natural-language search, full developer API.
- BlueMesh Custom, for teams that want their own agents and workflows built on the private stack.
The question was never whether fifty dollars a seat is cheap. It is whether one afternoon of a supervisor listening to recordings, or one FDCPA claim, or one client relationship lost because nobody could find what was actually said, is worth more than fifty dollars a seat.
Most months, it is. Some months, spectacularly.
Setup usually takes about a day. If you are already on Vaspian there is no hardware, no migration and no new vendor. It switches on over the phone system you already have.
That makes the private AI phone system an extension of the communications environment already in place rather than another disconnected AI tool.
Where this series goes next
This post is the foundation. Where your calls go, who owns each layer of the stack, and what we decided not to build. What it does not cover is what any of that means on your floor specifically, because the answer is genuinely different depending on whose floor it is.
The same private AI phone system can support very different workflows because the calls, risks, and operational questions are different in each industry.
So the next four go deep, one industry at a time.
- Collections. FDCPA and Reg F exposure, mini-Miranda verification, right-party contact detection, promise-to-pay extraction, and what changes when QA covers every call instead of one in a hundred.
- Healthcare. Patient intake, no-show recovery, protected health information handling, and how many patients a practice loses each month to a phone number written down wrong.
- IT support and MSPs. The gap between what your ticketing system records and what the caller actually said, and what that gap costs you in repeat escalations.
- Logistics and dispatch. Wrong addresses, missed pickups, accessorial disputes, and tracing an exception back to the call that caused it.
Same platform underneath. Four very different conversations.
We still own the building
Vaspian has been in Buffalo since 2007. Independent, unacquired, and still answering our own phones in an average of nine seconds, which we mention rather a lot because we had to work for it.
The reason we can promise that your conversations do not leave the building is unromantic.
We still have a building. The servers are in it. Most of this industry was bought, merged and outsourced somewhere over the last twenty years. We were not. Your recordings are here, and so are we.
That physical ownership is part of the private AI phone system story because infrastructure has to exist somewhere, even when the interface makes it feel invisible.
We do not sell minutes. We sell what minutes mean.
Ask us what your calls have been hiding. A real person picks up.
FAQ
This section answers common questions about Vaspian’s private AI approach and what happens to customer conversations during analysis.
What makes this a private AI phone system?
A private AI phone system is defined here by where the conversation is processed and who owns the layers involved. Vaspian’s approach keeps the voice, transcription, AI orchestration, and customer context inside the controlled stack described in this article rather than sending the conversation to a third-party model provider.
Where does Vaspian process customer call data?
Vaspian processes the call through its own voice and private AI infrastructure as described in this article. The goal of the private AI phone system is to keep the conversation inside the controlled stack rather than sending it to a third-party model provider.
What happens before a recording is analysed?
Identifying information is removed at the front of the pipeline before analysis. In the private AI phone system described here, redaction happens before the model receives the information rather than after processing.
Does Vaspian train shared models on customer conversations?
No. Customer data is not pooled or used to train shared models. The private AI phone system uses the calls for the customer’s own business context rather than to improve another customer’s product.
What does BlueMesh do?
BlueMesh is the orchestration layer that determines which models run, in what order, with what guardrails, and where a human needs to sign off. Within the private AI phone system, it also supports the private search, voice, document, translation, and workflow components described in the article.
Why does Vaspian keep a human in the loop?
BlueMesh can flag, rank, summarise and surface information, but the article keeps judgement with people. The private AI phone system can support the decision process without pretending that decisions such as whether a dispute is valid or whether an employee requires corrective action no longer need human judgement.