UPDATE, 19 SEPTEMBER 2026

On-Premise AI Contract Review For Law Firms And Confidential Deals

A first-year associate can summarise a 300 page share purchase agreement in four minutes with a language model, or in an afternoon by hand. In a July 2026 survey of 500 employed United States adults commissioned by Kolmogorov Law, 11.4% said they had entered contracts or legal documents into a personal AI account their employer does not control. ARSA Sovereign exists so that the fast version of that work can happen inside the firm’s own rack, from $26,900.

THE PROBLEM

Why Unredacted Contracts End Up In Public Chatbots

Staff paste client documents into public chatbots because the manual alternative costs them an evening, and an IT policy that bans the tool changes where the behaviour happens rather than whether it happens. The arithmetic is what drives it, so the fix has to change the arithmetic.

The exposure is specific in legal work. A draft in a public tool may sit in a vendor’s retention window, may be reviewable by human raters under some terms of service, and may be subject to a preservation order in unrelated litigation. Once a third party holds the document, the firm has to explain to a client, and possibly to a court, why privilege survived that transfer. That is a conversation worth avoiding before it starts.

The profession has already written the rule. ABA Formal Opinion 512, issued on 29 July 2024, warns that self-learning generative AI tools raise the risk that one client’s information is disclosed improperly, and it treats informed client consent as a prerequisite before confidential material goes into such a tool. The opinion also says boilerplate consent buried in an engagement letter does not carry that weight. A tool that keeps the document inside the firm’s own network removes the question from the engagement letter entirely.

Clients are starting to write the same rule into their own paperwork. Research by Fulkerson Advisors found that only 2% of 1,054 outside counsel guidelines mention AI at all, and among the ones that do, the common demand is that matter data stays out of tools that train on inputs or retain them. A firm that can answer that demand with an appliance in its own server room answers it in one sentence.

THE CAPABILITY

What ARSA Sovereign Does With A Contract

ARSA Sovereign is an appliance that reads documents, reasons over them with a local vision-language model, and writes structured results into the systems the firm already runs, with no cloud API and no outbound connection required at inference time. Document text extraction, embedding, indexing and reasoning all happen on encrypted local storage. No document, query or result reaches ARSA.

For a deal team, the practical shape of that is three things.

Reading What Arrives In The Data Room

Scanned PDFs, dense multi page layouts, handwritten annotations on an executed original, schematics attached to a construction contract. Full page layout parsing runs on the appliance, so a thirty year old scanned lease is handled the same way a clean Word export is.

Producing Structured Output Instead Of A Conversation

Stating a goal in plain language generates a step by step form for that goal, with the fields, tables, validation rules and approval steps the task needs. An associate auditing a vendor contract against the firm’s standard NDA terms gets a populated review sheet, and can change any field by asking in a side panel.

Running Overnight On Volume

Multi hour batch jobs run in the background. A thousand document audit started at 7pm produces a structured report and a notification when it finishes, which fits the rhythm of a due diligence window better than a session that has to be watched.

Where the output has to land in a business system, connectors for Odoo, SAP and Zoho are available over Model Context Protocol, and every connector is read only by default. Every write is staged for explicit human approval before it commits. There is no autonomous write mode to enable.

THE NUMBERS

The Two Systems, And What Separates Them

Both systems run the same operating system, the same workflows and the same support commitment. The difference is model precision, document throughput, and how much context the system holds at once.

Three years of support, model updates and fleet management are included with a perpetual purchase. Entry is also available at $1,090 per month. A software only licence for the operating system is $14,900 per node per year, or $39,000 perpetual plus 18% maintenance, and it requires a validated hardware configuration with a 48 GB accelerator minimum. All figures are international list price in USD and exclude import duty, VAT, customs clearance and rack installation.

For a firm running long context work across a full data room, the 96 is the one that matters, because document processing never competes with reasoning for memory. For a practice starting with contract review and internal workflows, the 48 does the job at $18,000 less.

THE BOUNDARY

What This System Will Not Do For You

This runs a 27B-class open-weight vision-language model under a permissive licence. On the hardest reasoning tasks it is behind a frontier-scale hosted model, and ARSA says so rather than letting a buyer discover it in week three. Where the requirement is maximum raw capability and the data is genuinely allowed to leave the building, a cloud model is the right answer.

Three other cases fall outside it. Workloads that are bursty and occasional are cheaper in the cloud, because an appliance sitting idle still costs what it cost. A buyer who needs one specific proprietary model available only as a hosted API cannot get it here. And a firm with no rack space, no power budget and no appetite for on-premise hardware should solve that first, although the chassis converts between 4U rack mount and a standalone tower with acoustic damping and is designed to run in an office.

The system also does not practise law. It extracts, summarises, structures and drafts. A lawyer reviews the output, and under Opinion 512 that review is the lawyer’s own competence obligation rather than a setting in the software.

THE ENVIRONMENT

Built For A Firm That Gets Audited

The platform runs a hardened Linux host with kernel livepatching, so security updates land without reboots, and it ships with FIPS 140-3 validated cryptographic modules and CIS and DISA-STIG hardening profiles applied at build. Workloads are isolated in containers with dedicated accelerator binding.

The fleet management channel is encrypted, and it can be disabled entirely. The appliance can be air-gapped, in which case model updates and licence activation arrive as signed offline packages transferred by the firm’s own approved media process. For a practice handling M&A, IP portfolios and client-confidential documents under HIPAA, GDPR or PDPL exposure on the client side, that is the difference between a policy statement and a network topology.

ARSA Sovereign is in pilot. ARSA has been building and operating AI systems in production since 2018 across defence, law enforcement and industrial deployments, and the same engineers build the language model systems described here.

Frequently Asked Questions

Does Any Document We Process Reach ARSA Or A Third Party?

No. Text extraction, embedding, indexing and reasoning all run on the appliance, against a vector index and document vault on encrypted local storage. No outbound connection is required at inference time, and the management channel can be switched off.

Can It Run Completely Air-Gapped For A Sensitive Deal?

Yes. No outbound connection is needed for inference. Model updates and licence activation are delivered as signed offline packages, moved across by the firm’s own approved media process.

Which Model Does It Run, And Can We Tell A Client?

A 27B-class open-weight vision-language model under a permissive licence, kept current under the support term. ARSA does not fix a named version in documentation because the open-weight field moves faster than product literature, so the appliance ships with the model that performs best for the workload at that time. What a firm can tell a client is the part that matters to them: the model runs on hardware the firm owns and trains on nothing.

Can The System Change A Record In Our Practice Management Or ERP Without A Person Approving It?

No. Connectors are read only by default and every write is staged for explicit human approval before commit. There is no autonomous write mode to switch on, which is a design decision rather than a configuration flag.

Do We Need A Data Centre To Run This?

No. The chassis converts between 4U rack mount and a standalone tower with acoustic damping, and it is designed for an office environment. Lead time is three to four weeks from order.

How Do We Find Out Whether This Fits Our Document Volumes?

Through a paid feasibility assessment, from $4,500 over two weeks, which produces an operational diagnosis, a technical feasibility judgement against the firm’s actual data and infrastructure, a cost model, and a go or no-go recommendation. The fee is deducted from the project fee if the firm contracts within 90 days. The scope of that and every other engagement is published on ARSA’s services page.

Put The Review Inside Your Own Building

If associates are already using AI on client documents, the open question is where those documents sit tonight. ARSA Sovereign moves that work onto an appliance in your rack, with a human approval gate on every write and nothing leaving your network. Read the full specification and pricing on the ARSA Sovereign product page, or tell us your document volumes and we will tell you which of the two systems fits.

Sources:
– Nearly 2 in 5 US workers have put company information into personal AI accounts (Kolmogorov Law / Pollfish survey, July 2026)
– ABA issues first ethics guidance on a lawyer’s use of AI tools (Formal Opinion 512)
– 2% of 1,054 Outside-Counsel Guidelines Mention AI, Fulkerson Advisors