Service · discovery from $900 · implementation from $3,000

AI without data leaks

AI automation for businesses where confidentiality isn't a preference but an obligation. Sensitive data never leaves your infrastructure — and answer quality stays at frontier-model level.

A clinic administrator, a law firm partner, an insurance agency owner — they all want the same thing everyone else does: less manual work. And they all hit the same question, which most vendors answer poorly: where exactly does the client data go?

Because when a staff member pastes a medical summary or a signed agreement into a chat window, that data leaves your perimeter. Legally, that's disclosure to a third party. For protected health information, privileged communications or regulated financial records, that isn't a technicality.

The market offers two bad options: cloud AI with the exposure, or a local model — safe, but noticeably weaker than frontier models, and carrying hardware costs.

We build a third. Processing happens inside your infrastructure, and what goes out — if anything does — is a de-identified fragment that can't be traced back to a person. The obligation is met and the quality holds.

How it works

Five parts of the design

Data flow mapping

We start with what actually leaves your perimeter today and which of it is sensitive. Often it turns out that 80% of the tasks don't need identifying data at all — it just travels along for the ride.

De-identification layer

A service inside your network strips everything that identifies a person: names, phone numbers, addresses, record and policy numbers. What goes out is a de-identified fragment; the answer comes back and real values are re-inserted on your side.

Local processing where de-identification isn't enough

If a task genuinely requires full context, processing stays inside: a local model or conventional logic on your own server, with no outbound call at all.

Hybrid routing

Rules that decide what stays in, what can go out de-identified, and what isn't sensitive in the first place. We write them with you, not for you.

Logging and documentation

You can see what data went where and when. That's what you show an auditor or your privacy officer — instead of "the vendor assured us it was safe."

Who needs this

Seven fields — recognizing one is enough

Medical, dental and mental-health practices, labs — protected health information
Law firms and solo practitioners — privileged client communications and case files
Accounting, tax and bookkeeping firms — client financial records
Insurance agencies and brokers — personal data and claim histories
Financial services and lenders — regulated customer records
HR and recruiting — candidate files, background checks
Companies with proprietary IP — designs, source code, research data

Why not just run a local model

Four reasons

Local models are weaker. On realistic hardware they fall noticeably short of frontier models on complex work. A lawyer used to good drafts will get worse ones — and stop using the tool.
Not every task needs identifying data. Classifying an inquiry, drafting a reply, summarizing a document — all of that works on de-identified text. The real values go back in on your side.
Hardware is cost and maintenance. A GPU server takes money, space, power and attention. We recommend one when the task genuinely requires it, not by default.
The right architecture costs less than the wrong one. A hybrid setup is usually several times cheaper than going fully local — and performs better.

What we don't promise

Where our responsibility ends

We don't issue compliance certifications, and we don't replace your counsel or your privacy officer. Our part is technical: building the system so sensitive data physically doesn't leave your perimeter, and documenting exactly how that works.

The legal assessment is theirs to make — but they make it based on a clear map of data flows rather than a vendor's assurances.

How we work

Four steps

01
Free 30-minute call

We look at what you want to automate and which data those tasks involve.
02
Discovery & audit

From $900. What leaves the perimeter today, what's sensitive, what can be de-identified, what has to stay inside. You get a data-flow map and a plan.
03
Implementation

From $3,000. De-identification layer, routing, local processing where needed, logging.
04
Ongoing support

From $500/mo. Monitoring, changes as tasks evolve, model updates.

The prompt is ready

See how we'd help your business — on autopilot

We've prepared a ready-made prompt for you. Add a few words about your business, and AI will show which Proplat services would best remove your busywork and free up time for growth.

Click any AI button — the ready-made prompt is copied automatically.

FAQ

What people ask before the first call

Can we use AI at all if we handle regulated data?

Yes — the question isn't whether, it's what leaves your perimeter. If what goes out is de-identified text that can't be traced to an individual, the third-party disclosure exposure goes away. If a task genuinely requires full context, we keep the processing inside your infrastructure.

How is this different from just deploying a local model?

A local model is one tool, not the answer by itself. It's weaker than frontier models and it needs hardware. We use it where nothing else works and solve the rest through de-identification — which is both cheaper and better.

Who is responsible for compliance?

Your counsel or privacy officer. Our part is technical: building the system so sensitive data doesn't leave, and providing clear data-flow documentation for them to assess.

What does it cost?

The first 30-minute call is free. Discovery and audit start at $900, implementation at $3,000, ongoing support at $500 a month. Projects requiring local processing run higher than standard automation work.

Do we need to buy a server?

Usually not. The de-identification layer is an ordinary service running on infrastructure you already have. Dedicated hardware is only needed when a task genuinely requires a local model — and we'll say so plainly if that's your case.

How reliable is the de-identification?

We design it around the specific data types in your field and test it on real examples before launch. Plus the log: you can see exactly what went out and audit any case. There are no absolutes in security — but there's a controlled, verifiable design instead of an opaque one.

We already use cloud AI. What now?

Start with the data-flow audit: usually it turns out most workflows are fine and only two or three are the problem. We rebuild those without breaking what already works.