Deterministic AI takes the stand: LawFairy challenges probabilistic tools in regulated legal work
Generative AI in law just met a hard stop from inside the profession. Raj Panasar, founder of AI-driven firm LawFairy and former Hogan Lovells partner, called probabilistic AI "fundamentally unsuitable for regulated legal work, where an outcome is either right or wrong."
His point is simple: legal outcomes need certainty and an audit trail. Guesswork, even when it sounds confident, doesn't pass regulatory muster.
Deterministic vs probabilistic: why it matters
Probabilistic systems (like large language models) predict the next likely word. That's useful for drafting, not for decisions with strict statutory thresholds.
LawFairy says its core engine is deterministic-rule-based logic where the same inputs always produce the same outputs, and every step can be shown. Think expert systems, not black boxes.
A helpful analogy: commercial flight simulators. The logic is fixed, auditable, and repeatable. That's the bar for high-stakes work.
LawFairy's model and early focus
LawFairy has been authorised as a law firm by the Solicitors Regulation Authority (SRA). It uses generative AI only for a clearer interface; legal outcomes come from pre-validated rules embedded in structured logic.
Panasar on the choice to start with immigration: "Immigration decisions aren't about sounding right - they're about being right, and being able to prove it. When decisions affect someone's right to work, remain, or settle, confidence without accountability isn't good enough. We've built LawFairy so every outcome can show its working."
Practical takeaways for law firm leaders and in-house teams
- Use generative AI for speed and polish: first drafts, tone, summaries, brainstorming-always with human review.
- Use deterministic systems for regulated outcomes: eligibility checks, statutory thresholds, rule-based triage, and where you must evidence every step.
- Keep auditability non-negotiable: full decision logs, versioned rules, and reproducible outputs.
- Treat model "confidence" as cosmetic unless it's tied to verifiable logic and citations.
Procurement checklist: questions to ask vendors now
- Determinism: Given the same inputs, do you guarantee the same outputs?
- Rules and testing: Are legal rules explicit, versioned, and regression-tested before release?
- Audit: Can I see every rule fired, data used, and reasoning path for each decision?
- Controls: How are updates approved, rolled back, and monitored for drift or defects?
- Liability and supervision: How does the system support SRA expectations on competence and oversight?
- Data: Where is data stored, how is PII handled, and what third parties are involved?
Governance moves to implement this quarter
- Map use cases by risk: draft-level vs decision-level. Lock decision-level behind deterministic logic.
- Set policy for human-in-the-loop review thresholds and sampling rates.
- Create a rule-change board: legal, risk, and tech sign-off before any rules go live.
- Establish incident response for AI errors: pause, remediate, notify, and document.
- Train teams on limitations: what genAI can and cannot do under your policy.
Why immigration makes sense as a starting point
Eligibility often turns on structured criteria, evidence matrices, and timelines. That's fertile ground for deterministic logic with clear audit trails.
Firms serving this area can gain consistency, reduce rework, and support file defensibility-without pretending a probabilistic output is "good enough."
Market context
LawFairy's approval lands as LawtechUK reported a 35% rise in startup investment to nearly £200m in 2025. Expect more tools-but also sharper questions about auditability and accountability.
Helpful resources
Your membership also unlocks:
