← Industries/LegalTech
Promethean for LegalTech
Lawyers using AI face overlapping pressures: bar-association rules on competence + supervision (US Model Rule 5.3; EU equivalents); EU AI Act Annex III §8 for AI in administration of justice; professional-liability insurance carriers asking for audit evidence. Promethean's evidence layer is the per-decision record that makes 'the AI's draft was reviewed and the reasoning is preserved' structurally true rather than asserted.
Who this is for
Compliance + product teams in legaltech SMBs (contract analysis · e-discovery · legal research · document automation). Typically:
- Seed–Series B legaltech with 5–60 engineers; 1–4 AI features in production.
- Customers: solo + small-firm attorneys (heavy ABA/bar-rule scrutiny), larger firms (procurement-heavy), in-house legal teams.
- Decision volume: 100–50k AI-touched decisions per attorney per month.
- Professional-liability insurers (CNA, AON Affinity, ALAS, state-bar mutuals) increasingly ask: 'show us your AI audit trail' as a coverage requirement.
- Selling into law firms means surviving the most cynical procurement questions on the planet.
The regulatory pressure
EU AI Act Annex III §8(a) — administration of justice (scope-dependent)
AI systems used by (or on behalf of) a judicial authority to assist in researching, interpreting facts and law, or applying law to concrete facts — or used similarly in ADR — are high-risk. Most private-sector legaltech (contract review, e-discovery) sits outside §8(a) because the deployer is not a judicial authority; legaltech sold into courts or ADR providers is squarely in scope.
Deep-dive →ABA Model Rule 1.1 + 5.3 — competence + non-lawyer assistance
Lawyers must competently use technology; supervise non-lawyer assistance (which now extends to AI tools by recent bar opinions). Audit trail of how AI was used helps demonstrate competent supervision.
Confidentiality / attorney-client privilege
Client data in AI prompts can risk privilege waiver depending on the LLM provider's data-retention terms. PII-safe audit trail (inputHash only) demonstrates that the privileged content didn't leak into a third-party log.
GDPR Art. 22 — automated decisions affecting legal rights
When legaltech AI surfaces conclusions that materially affect a client's case (e.g. settlement-value estimates), the underlying processing may engage Art. 22 if the human review is rubber-stamp.
Deep-dive →Where LLMs typically run in LegalTech
Contract clause extraction
Extractor pulling structured clauses + risk flags from contracts. Schema-bound output; reviewer-gate on schema-violations + high-risk-clause flags.
E-discovery responsiveness classification
Classifier producing responsive / non-responsive / privileged per document. Always-human reviewer on privilege flags; chain shows reviewer engagement rate for adversarial discovery disputes.
Legal-research summarisation
Summariser producing case-law digests from search results. Citation-verification reviewer gate (hallucination risk is real in legal research); fallback when sources can't be verified.
Document drafting (motions, contracts)
Drafter producing first-draft documents. Always-human reviewer gate (no document goes to the client unreviewed); reviewer-verdict captured per draft.
How the substrate maps to your audit
| Your LLM feature | What the regulator asks | Promethean evidence |
|---|---|---|
| Contract extraction | Bar rule competence — what did the AI extract; did a lawyer review? | L12 entry with reviewerVerdict + specHash per document. |
| Privilege-flag review | Was every privilege determination reviewed by a lawyer? | Always-human gate + reviewer identifier per flag. |
| Citation hallucination check | Did the AI invent case citations? Did a human verify? | schemaValid: false + fallbackTriggered: true on unverifiable citations. |
| Client confidentiality | Did prompts containing privileged content get retained by the LLM provider? | inputHash only in the L12 chain; raw client content never leaves operator's control via Promethean. (Operator's LLM-vendor contract governs vendor-side retention.) |
| Settlement-estimate appeal | GDPR Art. 22 + ABA competence — meaningful human review on material outputs. | Reviewer gate + chain shows override rate for material outputs. |
Which Promethean tier fits
Recommended for typical SMBs in LegalTech
Team€79 / spec / month
Unlimited specs; €79 per active spec per month. 100k entries per spec. Daily OTS anchoring. Single-tenant.
Most legaltech SMBs start with 2–5 active specs (extraction + e-discovery + research + drafting + occasional fifth). Team tier at €79/spec/month is the natural starting point. Upgrade to Production (€499 flat) when you cross 6+ specs or sell into larger firms demanding multi-tenant primitives (one platform serving multiple law firms, each with their own audit needs).
Larger LegalTech operators with multi-tenant or framework-template needs upgrade to Production (€499 / month flat).
What this looks like in practice
Hypothetical: a contract-AI SMB in a malpractice-insurance procurement review
A contract-extraction SMB applies for professional-liability coverage. The insurer's underwriter requires: 'demonstrate that every contract your AI processes is reviewed by a human attorney; demonstrate hallucinated-clause risk is bounded; demonstrate client confidentiality is preserved.' Without Promethean: written assertions + sample documents. With Promethean: an audit-bundle export showing 100% reviewer-engagement rate on schema-violating outputs, model-identity per decision, inputHash-only audit-trail confirming raw client text never enters the chain. The underwriter runs verify.mjs against the sample. Coverage approved at standard rates instead of high-risk-tier rates.
Frequently asked
Won't our LLM-vendor relationship retain client data anyway? How does Promethean help?
Your LLM-vendor's data-retention terms are between you and the vendor; Promethean doesn't proxy LLM calls and doesn't see prompts. Where Promethean helps is the audit-trail side: the L12 chain records the inputHash (SHA-256) of every prompt, not the prompt itself. If a privileged-content question arises, you can prove that PROMETHEAN's audit trail did not retain the privileged content. The vendor question is separate (and best resolved with a Business Associate Agreement / DPA + zero-retention clauses with your vendor).
Does Promethean reduce hallucination risk in legal research?
Reduces visibility of hallucinations and forces fallback handling. Closed-enum output schemas + citation-verification gates mean output that doesn't pass validation triggers the deterministic fallback (e.g. 'return verified-only citations'). The L12 entry records schemaValid + fallbackTriggered per query, so a partner reviewing the AI's research history sees exactly how often hallucination protection fired. Doesn't make the LLM smarter; makes the failure mode visible.
Bar associations are issuing guidance on AI use almost monthly. Does Promethean help me track that?
Indirectly. The substrate provides the per-decision evidence layer; mapping that evidence to your specific bar's monthly opinion is your compliance work. But the evidence layer is sector-agnostic: when your bar issues new guidance on (say) AI-supervision standards, you can demonstrate compliance with new guidance against the same chain, retrospectively. That's the leverage — the audit trail is reusable across multiple regulatory regimes.
Our firm-customers want their own audit trail — separate from ours. How do we segment?
R4 multi-tenancy. Each L12 entry can carry a tenantId field (typically firm-id or matter-id). Production tier supports this natively. You ship each firm a per-tenant audit-bundle filtered to their tenantId; they verify with verify.mjs without seeing other firms' data. Particularly useful when firms ask the chain be exportable to their own e-discovery / records-management system.
What about US bar rules — Model Rule 5.3 supervision standards?
ABA Model Rule 5.3 and parallel state rules require lawyers to supervise non-lawyer assistance; Model Rule 1.1 requires competence with technology. ABA Formal Opinion 512 (July 2024) is the leading authority on attorneys' generative-AI duties — competence, confidentiality, supervision, candour to tribunal, and fees. Many state bars (CA, FL, NY, DC) have parallel opinions. The reviewer-verdict per L12 entry is the concrete evidence of supervision; the override rate over time demonstrates the supervision is meaningful (i.e. not rubber-stamping) — which matters whether the scrutiny is bar-rule competence or GDPR Art. 22 'meaningful human intervention' for EU-data matters.
Definitions used on this page
The substrate primitives referenced above (L12 receipt chain, spec hash, reviewer gate, fallback behaviour, OpenTimestamps anchor, tenant ID) all have canonical definitions in the glossary: