Four risks that are no longer hypothetical.

The 2024 conversation about AI in regulated industries focused on benefits. The 2026 conversation has shifted — because the risks have also been measured, in court, in regulatory filings, and in professional guidance.

Hallucination in work product

Sanctions for AI-fabricated citations now number in the hundreds across U.S. federal and state courts. Hallucination is an inherent property of current AI systems and cannot be eliminated. The firm's responsibility is to adopt tools that make hallucination detectable.

Confidentiality exposure through cloud inference

Cloud-based AI transmits documents to vendor servers for processing. Privacy policies and DPAs describe what the vendor promises. The architectural fact is that at processing time, the vendor has access to your privileged documents.

Regulatory and ethical exposure

The EU AI Act begins enforcement August 2026. Colorado's AI Act begins June 2026. Professional regulators are issuing formal AI guidance. None of it reduces the firm's documentation burden or the duty of care owed when using AI in client work.

Vendor dependency and lock-in

AI for regulated industries is a young market. Vendors have been acquired, pivoted, or absorbed. A firm integrated with a specific vendor's API and cloud is exposed to that vendor's business decisions in a way a firm running on its own hardware is not.

Three deployment patterns. Only one satisfies all four conditions.

A firm evaluating AI is not just choosing a vendor — it is choosing an architectural pattern. Cloud inference, hybrid, or on-premise. Each has a distinct data flow and risk profile.

Architecture comparison: cloud inference vs hybrid vs on-premise deployment patterns
Cloud inference

Large risk surface, controlled by the vendor. Documents transmitted at rest and in use. 10–15 sub-processors typical.

Hybrid

Medium risk surface. Documents stored locally, but queries cross vendor infrastructure at every interaction.

On-premise

Small risk surface, controlled by the firm. No data transmitted. Four scoped outbound connections, each disclosed.

Built around four architectural commitments.

Documents stay where they belong

Installed as a Windows application on your firm's hardware. Every stage — text extraction, indexing, AI inference, response synthesis — runs locally. No transmission of document content to any external system.

Every answer, traceable to a specific source

Source-bound answering. Every claim in a response is directly traceable to a specific passage in a specific document. Hallucinations become detectable rather than invisible.

A product the firm can adapt to itself

FORGE lets the firm train adapters on its own documents — without transmitting training data, and producing a model the firm owns in a standard format.

Built to be reviewed

Per-query audit trail, citation on every answer, EULA transparency statement aligned with EU AI Act Article 13, and full architectural documentation available for IT security review.

Firm and Firm+ tiers.

Ready to review the architecture with your team?

Start with a free trial on your own hardware. Or talk to us about firm-tier deployment.

Start 30-day trial Talk to us about Firm-tier deployment