Legal AI needs human feedback or it drifts off course
The RegTech market is full of AI claims. New features, bigger models, more advisors. Yet a core problem remains: tools trained on what the law says, not how it actually gets applied.
That gap is where outcomes break. Zeidler Group is calling it out and they're right - without practitioner feedback, legal AI is just a smarter search bar.
Law vs. practice: the gap that trips teams up
Blackletter law sets the floor. Real compliance lives in interpretation, risk appetite, and process - and it varies by firm while staying within the rules.
This "market practice" isn't in statutes or guidance. It sits with the people who make calls under pressure and defend them in audits.
Why static data isn't enough
Feed a model only with statutes, policies, and PDFs and you'll get tidy answers that don't survive contact with business reality. Garbage in, garbage out.
Context, thresholds, and exceptions are learned in the field. If your training data ignores that, your outputs will too.
Zeidler Group's stance
Treat AI as a living product. Pull in continuous feedback from active practitioners. Update the system so it reflects day-to-day compliance - not just the letter of the law.
That means curating qualitative insight from clients and industry participants and folding it back into models, prompts, and rules on a regular cadence.
What useful feedback looks like
- Edge cases: where policy meets judgment (e.g., thresholds, exceptions, materiality).
- Risk-based rationale: why a decision was acceptable, not just what was chosen.
- Jurisdiction nuance: different supervisory expectations for the same rule.
- Process metadata: approvers, evidence standards, escalation paths.
- Outcome data: overrides, false positives/negatives, and audit outcomes.
An operating model you can deploy now
- Define a decision catalog: list recurring determinations, linked to policy and controls.
- Put humans in the loop: assign SMEs to review AI suggestions with clear SLAs.
- Capture the "why": require structured rationale and cite sources for every override.
- Close the loop: retrain, re-prompt, or update rules weekly or monthly from captured feedback.
- Governance: version models/prompts, log changes, and keep an exportable audit trail.
Controls and metrics legal leaders should demand
- Precision/recall by use case, override rates, and time-to-decision.
- Drift monitoring: shifts in data, outputs, and reviewer disagreement.
- Access, segregation of duties, and change control for prompts, rules, and models.
- Citations-first UX: every answer ties back to sources, with clear confidence signals.
- Data boundaries: client data isolation, residency options, and retention controls.
Vendor due diligence checklist
- Source of "market practice" data: how it's collected, validated, and refreshed.
- Evidence of practitioner input: named SMEs, review process, and measurable impact.
- Customization: can your risk appetite be encoded and audited?
- Safety: hallucination controls, human oversight defaults, and regulator-ready logs.
- Compliance posture: privacy, confidentiality, and clear training data boundaries.
Aligned with regulator expectations
Human oversight, traceability, and risk controls are now table stakes. Frameworks like the NIST AI Risk Management Framework emphasize feedback loops, measurement, and governance across the AI lifecycle.
What this means in practice
AI can scale good judgment, not replace it. Without practitioner feedback, you get pretty outputs that don't stand up in audits. With it, you get decisions that reflect policy, risk appetite, and supervisory expectations.
Zeidler Group's approach - continuous feedback from clients and industry participants, curated into retraining and rules updates - is the model to copy.
Bottom line
The best legal AI teams act like product teams. They ship, measure, learn, and update. The firms that build tight feedback loops will set the standard; everyone else will be stuck explaining why the tool sounded confident and still got it wrong.
Further learning: AI for Legal
AI Learning Path for Regulatory Affairs Specialists
Your membership also unlocks:
