Legal AI Agents, with Inference
Inside Intel TDX.
Contract analyst, due diligence, legal research. EU company, EU-hosted customer database. Inference runs in Intel TDX. We don't store your prompts.
Three AI agents, each an AI model following a review checklist, with a native RGPD Article 28 DPA. Outputs are not legal advice.
The problem
Client documents need more than a contract clause.
Every contract a lawyer reviews, every M&A data room a deal team opens, every legal memo an associate drafts is confidential client material. Without hardware isolation, the protection an AI tool offers is contractual: the infrastructure that runs the model can technically read prompts in memory, and retention and training terms decide the rest. Bar guidance on generative AI, in the United States and in Europe, asks lawyers to understand how a tool handles client information before using it.
VoltageGPU adds a technical control. Inference runs in Intel TDX: memory is encrypted by the CPU and the host's hypervisor cannot read enclave memory. We don't store your prompts. This does not replace your own confidentiality or privilege analysis.
The three agents
One TEE model family. Three legal workflows.
Contract Review
An AI model following an IACCM / WorldCC clause checklist risk-flags NDA, SaaS, M&A and employment contracts.
- IACCM clause checklists, jurisdiction detection
- CRITICAL / HIGH / MED / LOW risk grading
- Multi-language: EN, FR, DE, ES, IT
- Live tool ready at /contract
Due Diligence
Batch-review deal documents, up to 20 files per request, and get structured risk flags for each one.
- Upload PDF, DOCX, TXT, CSV; OCR for image files only
- Risk flags with clause location and recommendation
- Parties, amounts and dates extracted, two-pass check
- Output: structured results, Markdown export of chats
Legal Research
Research help on French and EU law. For legal questions, EUR-Lex and Légifrance search results are added to the prompt.
- Live EUR-Lex lookup of matching EU acts
- Légifrance search link for French law
- Memo, brief and pleading templates
- No automatic citation check: verify every citation
EACH AGENT IS BUILT ON FOUR LAYERS
Installable playbooks at /marketplace, NDA review, SPA review, due-diligence checklist, brief drafting. Free gets 5, Plus 50, Starter 100, Pro 200, Enterprise unlimited.
Workspaces at /workspaces, one Matter per case, deal or client. Filed documents are stored as notes, encrypted at rest in the EU; the assistant does not read them yet. Free 1, Plus 3, Starter 10, Pro 25, Enterprise unlimited.
A custom prompt run on a cron at /scheduled, result by email. Scheduled runs check no live source and read no files. Starter 1, Pro 3, Enterprise unlimited.
OAuth integrations at /connectors, Gmail, Google Drive, Slack, GitHub, Microsoft 365. Plus 1, Starter 3, Pro 5, Enterprise unlimited.
How Intel TDX isolates inference
The technical argument, in plain English.
Intel TDX (Trust Domain Extensions) is a hardware capability of recent Xeon Scalable processors. When a workload boots inside a Trust Domain, its memory is encrypted by the CPU. The hypervisor, the host operating system and the cloud operator sit outside the trust boundary: they can schedule the workload, but the host's hypervisor cannot read enclave memory. Your prompt reaches our API over TLS; we relay it, without storing it, to a third-party TEE inference network, where each request runs on one attested TEE instance.
Chats in your browser
Inference runs in Intel TDX. Our servers see prompts in clear to relay them, and chats stay in your browser.
Attested instances
Each request runs on one attested Intel TDX instance of a third-party TEE inference network.
Not kept by default, Matter on opt-in
Documents pass through our servers to a TDX enclave and are not kept. A document you deliberately file into a Matter (Beta) at /workspaces is stored, encrypted at rest, in the EU.
EU controller, named sub-processors
VOLTAGE EI is the French controller. The GPU sub-processor is named in the DPA and sits outside the trust domain.
Sanctions and bar guidance
What regulators and courts have said so far.
Mata v. Avianca (US, 2023)
Lawyers sanctioned for submitting six fictitious case citations generated by ChatGPT. The court imposed a $5,000 penalty and ordered letters to the judges falsely named as authors. The textbook AI sanctions case for the legal profession.
CNB guidance (France)
The Conseil National des Barreaux published guidance on generative AI that reminds lawyers of professional secrecy and of the need to check how a tool handles client data.
CCBE guidance (EU)
The Council of Bars and Law Societies of Europe published guidance on AI tools for lawyers that stresses the duty of confidentiality.
EU AI Act (in force since 2024)
AI systems that assist judicial authorities in researching and applying the law are listed as high-risk in Annex III. Other legal tools mainly face transparency duties. Hardware isolation does not by itself make a system compliant.
Comparison
VoltageGPU vs Harvey, Spellbook, Luminance, Lexis+ AI.
Sources: vendor-published documentation as of 2026. "None published" means we found no public description of hardware-isolated (TEE) inference.
Pricing
Solo to mid-market. No BigLaw lock-out.
Solo practitioners
- Qwen3.5-397B-TEE, 256K context
- Contract Analyst tool
- 2,000 messages / month
- EU company, RGPD Art. 28
Small teams (3 seats)
- 2,000 requests / month
- Bulk document analysis
- 10 matters (Beta)
- 3 connectors
Small firms (≤10 seats)
- 5,000 requests / month
- DeepSeek-V3.2-TEE reasoning
- OpenAI-compatible API
- 5 connectors (Gmail, Drive, Slack…)
Mid-market firms + GCs
- Unlimited requests and seats
- SCIM 2.0 provisioning
- RBAC admin console
- Admin audit log
FAQ
Frequently asked questions.
Can I trust an AI vendor with privileged client documents?
That is a judgment under your own professional rules, and we make no privilege promise. Here is what the technology does: your prompt reaches our API over TLS; our servers see it in clear and relay it to our inference provider, where the model runs inside Intel TDX. Each request runs on one attested TEE instance. Enclave memory is encrypted by the CPU, and the host's hypervisor cannot read it. Chats stay in your browser.
How do I demonstrate compliance to my bar association?
We cannot demonstrate it for you, but we can give you the facts for your own assessment: the data flow (TLS to our API, relay to a TDX enclave, no prompt storage), our sub-processor list and our RGPD Article 28 DPA. We hold no SOC 2 or ISO 27001 certification.
Where exactly is the data processed?
Inference runs on a third-party TEE inference network, inside an Intel TDX trust domain. The controller is VOLTAGE EI (France, SIREN 943 808 824) and the customer database is EU-hosted (Frankfurt). The GPU host is a listed sub-processor in the DPA and is US-operated, so we do not claim the compute stays in the EU. The protection is the enclave: memory encrypted by the CPU, with the host operator outside the trust domain. Documents pass through our servers to the enclave and are not kept.
What about hallucinations on case law?
For legal questions, the Sovereign Legal AI agent can pull EUR-Lex and Légifrance search results into the prompt; check every citation yourself. There is no automatic citation check, and no LLM removes hallucination risk.
Can the agents work in French only?
Agents accept and produce content in EN, FR, DE, ES and IT. The legal-source lookup covers EUR-Lex and Légifrance only; for US or other national law, the model answers from its training alone.
Read how it works. Then bring your hardest contract.
See how inference is isolated, and what is and is not stored, before you upload anything. The Contract Analyst is open on the Plus plan for $20/month.