AI Knowledge Base · Private AI Data Centres

Private AI for law and accounting firms

Client confidentiality is not a setting you can switch on in someone else's cloud. This guide is for law firms and accountancy practices that want the productivity of AI while keeping privileged and financial data under their own control: on their premises, in a private data centre, or in a cloud locked down so tightly that it behaves like one.

Law firms
Accountancy practices
On-premises
Private data centre
पूरी तरह सुरक्षित क्लाउड

Why cloud may not be an option

Professional firms hold the data that clients trust them with most. Before any of it reaches an AI model, the firm has to be able to answer: where is it processed, who can see it, how long is it kept, and can we prove it?

Professional duties

Solicitors must keep clients' affairs confidential under the SRA Code of Conduct, and legal professional privilege must not be put at risk. Confidentiality is a fundamental principle of the ICAEW and ACCA codes of ethics for accountants.

Data protection law

UK GDPR requires appropriate security, a lawful basis, and safeguards for any transfer outside the UK. AI processing of client data usually needs a data protection impact assessment (DPIA).

Client and insurer demands

Outside-counsel guidelines, panel questionnaires, audit clients and cyber insurers increasingly ask exactly where documents are processed and whether they train any model.

Financial and AML records

Accountants hold tax, payroll and anti-money-laundering records that carry their own retention and confidentiality obligations.

A favoured target

Law and accountancy firms are attractive to ransomware and fraud groups because they hold valuable data and work to deadlines. The NCSC has published threat guidance specifically for the legal sector.

Information barriers

Ethical walls between matters and clients must survive inside the AI too: a firm-wide index that ignores them can create a conflict.

Deployment options, from most to least isolated

There is a spectrum between a machine with no network connection and a public AI service. Most firms end up with two tiers: a private platform for client data, and tightly controlled cloud for everything else.

The five options, from most isolated at the top to least isolated at the bottom. Each rung shows its internet access and where the data lives.

Option	Where data lives	Internet	इनके लिए सर्वोत्तम	Trade-offs
Air-gapped on-premises	Your building, on hardware you own	None; updates by signed offline bundles	Highest-sensitivity matters, litigation, disclosure	Manual update process; capacity fixed by hardware
Offline-first on-premises	Your building	Outbound blocked; controlled update window	Most firms' private AI platform	Needs a patching routine and local support
Private data centre or colocation	Dedicated racks you own in a UK data centre	Private links only	Multi-office firms, firm-wide AI, shared services	Contracts, cross-connects and remote hands
Sovereign or dedicated cloud	Dedicated tenancy in a UK region, your keys	Private endpoints only	Burst capacity, firms without server rooms	Provider trust and exit planning
Managed cloud AI, fully secured	Provider's UK or EU region	Private endpoints, no data retention	Non-privileged work, marketing, internal knowledge	Provider processes the data, even if briefly

Use cases for law firms

Where a law firm's data stays. Matter files go to the private AI platform inside the firm, which checks matter-level permissions and logs every request. The lawyer reviews the answer. Nothing leaves the network.

Disclosure and document review

Classify, summarise and prioritise large disclosure sets, with every answer linked back to the source document.

Contract analysis

Extract clauses, compare against the firm's standard positions and flag unusual terms across hundreds of agreements.

Due diligence

Work through data rooms quickly, building issue lists and red-flag reports with citations.

Know-how and research

Ask questions of the firm's own precedents, opinions and practice notes, respecting matter-level permissions.

First drafts

Draft letters, attendance notes and standard documents from the firm's templates for a lawyer to review.

Dictation and transcription

Local speech-to-text for dictation, interviews and hearings, with no audio leaving the building.

Matter summaries

Chronologies, case summaries and handover notes generated from the matter file.

Time narratives

Suggest time-entry narratives from activity, reviewed by the fee earner before submission.

Use cases for accounting firms

The same idea for an accountancy practice. Client records and documents go to the private platform for extraction and checks, every decision is logged, and an accountant approves. Client data stays inside the practice.

Document capture

Read invoices, receipts, bank statements and payslips (OCR plus AI extraction) straight into the ledger workflow.

Categorisation and reconciliation

Suggest nominal codes and match transactions, learning each client's patterns, with an accountant approving.

Audit analytics

Test whole populations instead of samples, flag anomalies, duplicates and unusual journals for review.

Tax research and preparation

Answer questions from the firm's tax library and prepare working papers for review, never filing on its own.

Client queries

Draft replies to routine client questions from their own records, held inside the practice.

Accounts and notes

Draft management commentary and disclosure notes from trial balances and prior-year files.

AML and KYC checks

Read and check identity and source-of-funds documents, with every decision logged.

Forecasting

Cash-flow and scenario models built from the client's own data, run privately.

What a professional-grade private AI platform includes

These are the capabilities we build into private AI for law and accounting firms. Each one exists because a partner, a client, a regulator or an insurer has asked for it.

The five feature groups below are layers of one platform, numbered as in the cards.

Security and privacy

No outbound internet: default-deny egress, proven by a 24-hour packet-logging test before handover
Full-disk encryption (LUKS2) sealed to the TPM, with unattended restart and a sealed recovery key
UEFI Secure Boot with signed drivers; firmware passwords and locked boot order
Confidential computing on supported GPUs and CPUs (NVIDIA confidential computing, AMD SEV-SNP, Intel TDX) where data must be protected even from administrators
Mutual TLS between services; keys in an HSM or vault
Own network segment: staff reach one HTTPS port, nothing reaches back
Immutable, append-only audit logs of every prompt, document and answer, exportable to your SIEM
PII and secret redaction before anything is logged or indexed

Access and information barriers

Single sign-on with MFA through your identity provider
Role- and attribute-based access, mirrored from the document management system
Matter- and client-level permissions enforced at query time, so ethical walls hold inside the AI
Per-user and per-matter usage reports for partners and compliance

AI capabilities

Retrieval over your document management, practice and accounting systems through their APIs, with source citations on every answer
Several models side by side, routed by task: small and fast for classification, large for drafting and reasoning
Fine-tuning on the firm's own precedents and style (LoRA), kept on your hardware
Local speech-to-text, OCR for scanned bundles and translation
Agent workflows with human approval before anything leaves a draft state
Grounding checks and 'I don't know' behaviour instead of invented answers

Governance

An AI usage policy and DPIA prepared with you
Model register: which models, versions and licences are in use, and for what
Evaluation sets built from the firm's own work, re-run before every model change
Retention aligned with matter retention and legal hold
Reports ready for client audits, panel reviews and insurers

Operations

Signed, verified offline update bundles for the OS, containers and models
Monitoring of GPU health, latency and usage, alerting your IT team or ours
Encrypted, offline backups (3-2-1-1-0) and tested bare-metal recovery
72-hour burn-in and a handover pack with every test result
Support and patching agreements sized to the firm

When cloud is acceptable: fully secured options from the major providers

For some work, a properly locked-down cloud service is acceptable, and it adds capacity on demand. These are the controls we insist on, by provider. Your DPIA and client terms decide what may go where.

Two tiers. Client data stays on the private platform in the firm. Other work your DPIA allows goes through a private gateway, then a private endpoint, to the provider's model in a UK or EU region with your own keys.

Microsoft Azure (Azure OpenAI, AI Foundry)

Customer data not used to train models
Private endpoints and network isolation
Customer-managed keys
UK South and UK West regions; EU Data Boundary for EU offices
Modified abuse monitoring for eligible customers, so prompts are not retained for review
Entra ID with Conditional Access

Amazon Web Services (Amazon Bedrock)

Prompts and outputs not used to train base models or shared with model providers
PrivateLink VPC endpoints, no internet path
KMS customer-managed keys
London region
Guardrails for PII redaction and content control
CloudTrail logging of every call

Google Cloud (Vertex AI)

Customer data not used to train models
Zero data retention configuration for supported models
VPC Service Controls perimeter and CMEK
London region and data residency controls
Access Transparency and Access Approval for provider access
Model Armor for prompt and response screening

Oracle Cloud (OCI Generative AI)

Dedicated AI clusters for isolated model hosting
Customer-managed keys in OCI Vault
London and Newport (Wales) regions
Private service access

We pair any cloud use with a private gateway that enforces the firm's permissions, redacts identifiers and logs every request before it leaves your network.

On-premises and private data centre AI platforms

We are vendor-neutral and build on the platforms that suit each firm, from a single workstation to a private AI data centre.

Where an on-premises platform lives, and what runs on it: the systems, the private AI platform on top, and the firm's apps it serves.

Systems

NVIDIA DGX Spark and DGX Station for desk-side and small-team AI
Workstation towers and 4U servers with NVIDIA RTX PRO 6000 Blackwell (96 GB per GPU)
NVIDIA HGX B200 and B300 8-GPU servers for firm-wide platforms
OEM AI servers from Dell (PowerEdge XE), HPE, Lenovo (ThinkSystem) and Supermicro

Private AI platforms

NVIDIA AI Enterprise and NIM microservices, licensed for on-premises use
Dell AI Factory with NVIDIA and HPE Private Cloud AI as turnkey stacks
Red Hat OpenShift AI and SUSE AI on Kubernetes
VMware Private AI Foundation with NVIDIA for VMware estates
Nutanix Enterprise AI for hyperconverged environments
Open-source stacks: vLLM, Open WebUI, pgvector and Qdrant, on Ubuntu or RHEL

Where it lives

A server room or comms cupboard, with the right power and cooling
A dedicated rack in a UK colocation data centre
A private AI data centre for groups and networks of firms

Sizing by firm size

Indicative starting points. Every proposal is confirmed with your own documents, models and user numbers.

Firm	Concurrent users	Typical platform	Models
Sole practitioner or small practice	Up to ~10	Workstation with a 24–32 GB RTX PRO GPU	8–14B models
Small to mid-sized firm	~25	Workstation or server with a 48 GB GPU	Up to ~32B models
Mid-sized firm, busy department	25–50	Server with a 96 GB RTX PRO 6000	70B models at 4-bit
Large firm, multiple offices	Hundreds	Rack of 8-GPU servers plus all-NVMe storage	Several large models at once
Firm networks and groups	Thousands	Private AI data centre with HGX B200/B300	Frontier-scale and fine-tuned models

Size it in the AI data centre planner →

Compliance mapping

Technology doesn't make a firm compliant, but it should make compliance easy to show. This is how the platform supports the frameworks firms ask about.

UK GDPR and the Data Protection Act 2018

Data stays in the UK, processing is documented, access is logged, and we help prepare the DPIA.

SRA Standards and Regulations

Confidentiality and information barriers enforced in software, with evidence for supervision and audits.

ICAEW and ACCA codes of ethics

Confidentiality of client information, with records of who accessed what and why.

ISO/IEC 27001

Controls for access, cryptography, operations, logging and supplier management map directly to the platform.

Cyber Essentials Plus

Firewalls, secure configuration, patching, access control and malware protection built in.

EU AI Act

For EU offices: transparency, human oversight and record-keeping supported by design.

How to start

The four steps, in order: Discovery, Pilot, Prove it, Scale.

1
Discovery
A confidential conversation under NDA about your work, data and risk appetite.
2
Pilot
A desk-side system or a small server running your own documents in your building.
3
Prove it
Measured results on real matters or client files, plus the security tests and DPIA.
4
Scale
Firm-wide platform, private data centre or hybrid, with support and training.

Book a confidential consultation →Ask the AI Sales Engineer ↗

Microsoft Azure, Amazon Web Services, Google Cloud and Oracle Cloud names and logos are trademarks of their respective owners, shown only to identify their products.