1 of 14

motivations and limitations of adopting voice ai systems for legal intake

riya ranjan, pierce kelaita, hazel he, megan ma

2 of 14

roadmap

current landscape for ai-driven legal intake

voice ai tooling and prototypes

pilot study with mills legal clinic

3 of 14

roadmap

current landscape for ai-driven legal intake

voice ai tooling and prototypes

pilot study with mills legal clinic

4 of 14

the average legal intake takes

60

minutes of attorney Q&A

Use of AI tools cuts this time in

½

5 of 14

for legal aid groups, impact is even higher

  • 90% of legal aid professionals believe AI can address the access to justice gap
  • Legal Services Corporation-funded legal aid organizations must turn away 1 in 2 requests at intake due to limited resources
  • 74% of legal aid organizations already use AI (double the rate of the wider legal profession)

6 of 14

roadmap

current landscape for ai-driven legal intake

voice ai tooling and prototypes

pilot study with mills legal clinic

7 of 14

voice ai tech stack

MODELS

S2S, Cascaded (Eleven Labs, Cartesia)

ORCHESTRATION PLATFORMS

Vapi, Bland, Retell

END TO END AGENTS

Hello Megan AI, CaseGen

8 of 14

voice ai for legal intake

ACCESSIBILITY

only a phone number required

multilingual support

DATA

call recordings, transcripts, sentiment signals

​

​

OPERATIONAL CONTINUITY

fits into existing pipelines seamlessly

9 of 14

roadmap

current landscape for ai-driven legal intake

voice ai tooling and prototypes

pilot study with mills legal clinic

10 of 14

TASK: class action refugee immigration case; intake required to assess eligibility for potential settlement

COST

AGENCY

ACCESS

LANGUAGE

PRIVACY

11 of 14

workflow

agent debugging and redesign

testing with community members

feedback session with clinic

12 of 14

demo v1 (vapi + custom docs)

​

13 of 14

fundamental post pilot insights

  • Context management remains a technical challenge (interruptions, misspellings, etc.)
  • Legal aid organizations require customization beyond what classic orchestrator platforms offer
  • S2S models are too expensive for access to justice right now
  • Voice realism is not a silver bullet for driving interaction quality

14 of 14

special thanks to

​