SLAtech AI Legal
92/100UPL-aware, conflict-check native, confidentiality-tier configurable
Legal eval scoreboard
SLAtech AI Legal: 92/100
Reproducible 200-question Legal-specific eval harness. +24-point lift vs generic SLAtech-Business (68/100). Driven by UPL (unauthorized practice of law) guardrails, conflict-of-interest checks, and confidentiality posture. Pairs with umbrella eval scoreboard, Legal glossary and Legal FAQ.
Score breakdown by category
| Category | Legal-tuned | Generic | Lift |
|---|---|---|---|
| UPL guardrails Hard-coded refusal to dispense jurisdiction-specific legal advice - instead routed to lawyer-confirmation step. Generic chatbots happily improvise legal advice, exposing the firm to malpractice. |
97 | 56 | +41 |
| Conflict-of-interest checks Pre-intake party-name search against firm's existing-client database. Generic chatbots collect intake without any conflict check (firm liability). |
94 | 52 | +42 |
| Confidentiality posture PII redaction at ingest, single-tenant option, attorney-client privilege metadata tag. Generic chatbots ship zero redaction. |
92 | 62 | +30 |
| Matter-intake quality Structured intake captures jurisdiction, opposing party, statute-of-limitations clock, retainer-fee disclosure. Generic chatbots collect free-text only. |
90 | 71 | +19 |
| Citation discipline Refuses to cite case-law or statute numbers without source-anchor verification. Generic chatbots hallucinate citations (the 2023 ChatGPT-Mata sanctions story). |
87 | 78 | +9 |
Competitor comparison
Intercom Fin (generic)
63/100No UPL guardrails, no conflict-check, generic confidentiality posture
Smith.ai (legal-focused)
82/100Strong UPL guardrails but no FHIR-equivalent matter-intake schema, English-first
Tidio Lyro (generic SMB)
54/100Will improvise legal advice (malpractice exposure), no conflict-check, conversation cap
Continue the buyer evaluation
The per-vertical eval score is one input. Three more self-serve tools complete the picture without a sales call:
Reproduce the eval against your own tenant
Eval methodology is open-source. 200 sealed Legal-specific questions with LLM-as-Judge scoring on factuality, hallucination and confidence axes.