Private deal intelligence

Why a box

Your pipeline is the most sensitive text your company writes.

Prices, concessions, who is unhappy, which customer is about to leave. Most AI sales tools send all of it to a shared cloud model. Legal, security and procurement teams often say no to that, and the reps go back to guessing.

The box

A desktop-sized NVIDIA Spark-class AI computer with 128 GB of unified memory. Enough to run our model, keep your database, and retrain overnight, on one quiet box under a desk or in your rack.

included power, network, done

The model

Clossara's own 14-billion-parameter model, built for revenue conversations. It reads the whole thread, names the root cause of a stall, forecasts the outcome, and cites the exact message behind every claim.

fixed never trained on your data; updated by package, each release gated

The portal

The same Clossara Portal: Deal Health board, diagnosis with receipts, forecasts with a stated confidence, grounded drafts. Served from the box over your LAN, to your people only.

same product no feature cut for going private

How it works

Four steps. None of them involve the internet.

STEP 1

We ship the box

It arrives with the Portal, the database and Clossara's base model already installed. Your IT team gives it power, a network cable and an internal address.

STEP 2

You connect, read-only

Mailboxes and CRM are connected inside your network, read-only. Names and addresses are replaced by labels before any text reaches the model, the same redaction we run everywhere.

STEP 3

The model reads your deals

The same release we ship everywhere reads your threads on the box's own GPU and writes the diagnosis, the forecast and the drafts. It never trains on your data, on the box or anywhere else.

STEP 4

Your team works

Reps open the Portal on Monday and see the board: which deals are at risk, why, and what to send. New conversations are read as they arrive. Nothing is uploaded, ever.

Software updates arrive as packages your IT team installs. Support is remote only when you open the door, and never carries your conversations.

Measured

The model we put in the box is the one we measure.

Every Clossara release is judged on held-out sales threads the model has never seen, finding by finding, against the full conversation. These are the current numbers.

86%

of the findings a rep sees were judged true against the thread, on 176 held-out conversations. The model hides what it cannot support rather than showing it.

100%

of citations point at a real message. A finding without a receipt does not ship. You can open the message behind every claim.

~30s

for a full diagnosis of a long thread, on the box, with no network round trip. Threads with no written reply from the customer are diagnosed from the facts, in under a second.

Evaluation of the November 2026 release on our held-out set. The method, the bar and the numbers are in the evaluation report, four pages, sent on request.

Growing

Need more? Add a box.

There is no bigger plan to migrate to. One box serves one team. When a second team, a second office or a bigger pipeline comes, you order a second box. It ships pre-installed, your IT plugs it in, and it is working for that team the same day.

1 boxYour first team. Sales, from pipeline to forecast.

2 boxesA second team or office, with its own data and its own model.

3 boxes and upA region, a business unit, or a pipeline that outgrew one box.

What one box carries

In our own measurement a box re-reads and re-diagnoses about 2,000 deal threads in a working day on a real pipeline mix, where six threads in ten need no model call; long threads alone run at about a thousand. For most sales teams that is one box for a long time.

When to add one

A new team that should not share data with the first. A second site. Or the day the box no longer finishes the day's re-diagnoses inside the day. Each box is its own Clossara: no cluster to build, no shared database to migrate, no outage on the first box while the second arrives.

What the hardware allows

Every box has two 200 Gb network ports. Two boxes link with one direct cable and pool their memory, which NVIDIA rates for models up to 405 billion parameters. More boxes connect through a 200 Gb switch, one cable each; NVIDIA documents four on one switch and the same setup for more. Clossara's model is 14 billion, so one box runs it with room to spare, and a pool of boxes is there when a pipeline needs more.

Priced per box, hardware and installation included; seats are added per person. Ask for a quote for two or more.

Who it is for

Built for the teams that had to say no.

If any of these sound like your company, the box is the version of Clossara you can actually run.

Cloud AI is restricted by policy, as it often is in finance, legal, insurance, defence, healthcare and the public sector.
Data residency is non-negotiable. The data never leaves the room, so the question of which country does not arise.
Customer contracts forbid subprocessors. There are none. Clossara is not one either, at runtime.
Your deals are the company's secret. Pricing, terms and relationships that must not train anyone else's model. They train only yours.
You already have a security team. They get a box they can inspect, isolate and switch off, rather than a vendor's word.
You want a model that never trains on you. Not on the box, not anywhere. Your data is read, never learned from.

Three ways to run it

The box is Clossara with the last door closed.

	Clossara Box	Clossara cloud	Shared-cloud AI tools
Where your conversations live	On a box in your building	Your dedicated single-tenant instance	A shared multi-tenant cloud
The model at runtime	Clossara's model, on your box	Clossara's model, operated by Clossara	A shared third-party LLM
Trained on your data	Never	Never	Your data may train their model
Data leaving your network	None	Redacted text, encrypted, to Clossara's model	The full text
Claims in the analysis	Cited to the source message	Cited to the source message	Generated text, no citation
Who can switch it off	You, at the wall	You, in the Portal	The vendor, on its terms

FAQ

What people ask before they say yes.

What exactly is in the box?

A desktop-sized NVIDIA Spark-class AI computer with 128 GB of unified memory, running Linux. On it: the Clossara Portal, its database, the redaction service, and Clossara's deal-intelligence model served locally. It needs power and one network cable.

Do we need machine-learning staff to run it?

No. Nothing is trained on the box. Model updates arrive as packages your IT team installs, each release gated on our held-out set before it ships. Your IT team treats the box like any other server on the network.

Does Clossara ever see our data?

No. The box makes no outbound connection to Clossara. Software updates are packages your team downloads and installs. If you want remote support, you open a session and close it; a support session never carries conversation text.

How is this different from the Enterprise plan?

Enterprise is a dedicated single-tenant instance that Clossara operates, where the model only ever sees redacted text. The box moves the whole thing, model included, into your building. Everything else, the Portal, the citations, the forecasts, is the same product.

How many people can one box serve?

A sales team rather than a call centre. A single box diagnoses a long thread in about half a minute and re-diagnoses about 2,000 threads in a working day on a real pipeline mix, keeping every deal's board current. When a team outgrows it, or a second team needs its own, you add a box; more boxes rather than a cloud.

What does it cost?

A price per box, which includes the hardware and installation, plus seats for the people using the Portal. We quote after a short briefing about your team size and your systems.

Deal intelligence that
never leaves the building.

Deal Health