People seeking legal help should describe their problem, not diagnose its legal category before they can ask for help.
Collections: Human still decides · Queue eliminated
Executive brief
Justice Connect moved legal classification out of the applicant's form and into the intake system, while keeping a non-AI route and an expert loop for exceptions and retraining.
AI value · 'Something else' selections during intake
Company-reported9% among users offered the classifier
JusticeBench reports the operator's staged comparison. The retained non-AI path strengthens the comparison, but no independent statistical evaluation was published.
Before
Selects a legal category from a fixed list, often choosing the wrong area or 'Something else'. → Manually reclassifies uncategorized and misclassified applications. → Reviews the referral and redirects matters that do not match its expertise or mandate.
After
Describes the legal problem in ordinary language without selecting an area of law. → Maps the free-text description to likely legal categories and confidence scores for routing. → Reviews edge cases and works from cleaner, pre-categorized applications. → Compare model classifications with expert judgments and feed…
Human boundary
The model proposes intake categories and routing. People may use the non-AI path, and humans retain review of edge cases and all legal-service decisions.
Why it matters
Move classification burden off the requester and into the intake system; keep experts on exceptions and model governance.
This case is company-reported. Use it for the operating-model shift; do not treat the numbers as independently measured.
Before
How the work ran before the change.
Help-seeker
Selects a legal category from a fixed list, often choosing the wrong area or 'Something else'.
ControlThe applicant must understand the legal nature of the problem.
Intake staff
Manually reclassifies uncategorized and misclassified applications.
ControlHuman triage and service eligibility rules.
Referral partner
Reviews the referral and redirects matters that do not match its expertise or mandate.
ControlPartner-specific service criteria.
What changed
People seeking legal help should describe the problem, not diagnose its legal category before they can ask for help.
Decision rightAI classifies; experts govern exceptions
After
How the same work runs now.
Help-seeker
Describes the legal problem in ordinary language without selecting an area of law.
ControlDisclosure, opt-out, and a retained non-AI path.
Legal-problem classifier
Maps the free-text description to likely legal categories and confidence scores for routing.
ControlThe model classifies and routes; it does not provide legal advice.
Intake staff
Reviews edge cases and works from cleaner, pre-categorized applications.
ControlHuman review remains available for uncertain or consequential cases.
Pro bono legal experts
Compare model classifications with expert judgments and feed disagreements into refinement and retraining.
ControlStanding expert-review loop.
Exception path
Users can opt out to a non-AI route. Low-confidence or consequential edge cases receive human intake review, while expert disagreement is recorded for model refinement.
Decision authority
The model proposes intake categories and routing. People may use the non-AI path, and humans retain review of edge cases and all legal-service decisions.
| # | Actor | Action | Control |
|---|---|---|---|
| 01 | Help-seeker | Selects a legal category from a fixed list, often choosing the wrong area or 'Something else'. | The applicant must understand the legal nature of the problem. |
| 02 | Intake staff | Manually reclassifies uncategorized and misclassified applications. | Human triage and service eligibility rules. |
| 03 | Referral partner | Reviews the referral and redirects matters that do not match its expertise or mandate. | Partner-specific service criteria. |
| # | Actor | Action | Control |
|---|---|---|---|
| 01 | Help-seeker | Describes the legal problem in ordinary language without selecting an area of law. | Disclosure, opt-out, and a retained non-AI path. |
| 02 | Legal-problem classifier | Maps the free-text description to likely legal categories and confidence scores for routing. | The model classifies and routes; it does not provide legal advice. |
| 03 | Intake staff | Reviews edge cases and works from cleaner, pre-categorized applications. | Human review remains available for uncertain or consequential cases. |
| 04 | Pro bono legal experts | Compare model classifications with expert judgments and feed disagreements into refinement and retraining. | Standing expert-review loop. |
Work that left the path
Human role before
Applicants had to self-diagnose their legal category, while intake staff manually repaired misclassification and referral partners absorbed mismatched referrals.
Human role after
Applicants describe the problem in plain language; intake staff handle exceptions; legal experts govern model quality through ongoing adjudication.
AI role
Classifies a free-text description into legal-issue categories with confidence scores for routing and resource suggestions; it does not generate advice or determine entitlement.
23% among users on the retained non-AI path9% among users offered the classifier
Staged rollout, initially to about 30% and then about 50% of users · Justice Connect's production online Intake Tool
JusticeBench reports the operator's staged comparison. The retained non-AI path strengthens the comparison, but no independent statistical evaluation was published.
Move classification burden off the requester and into the intake system; keep experts on exceptions and model governance.
Anti-pattern
Automating intake while forcing every user onto the new path removes the comparison group and hides whether routing actually improved.
Questions
Portability conditions
Reputation risk
low