Skip to content
AwakastLet's talk

VAIOT, legal AI platform

Legal cases at VAIOT now take 14 hours instead of 48

We replaced a single AI assistant with a coordinated team of legal AI agents. They draft, review and explain legal work, and every answer can be traced back to its source.

Growing a product

faster legal case processing, from 48 to 14 hours
70%
lower LLM token cost
38%
users in the first week of the live demo
500+
from concept to production-ready
12 weeks

Where VAIOT started

VAIOT is a European LegalTech company. Its product relied on one AI model that had to draft, review and explain complex legal work across several areas of law. One assistant doing everything meant long turnaround times, and it was hard to see where an answer came from. In legal work that matters: a lawyer has to be able to check the source behind every statement. Before its public launch, VAIOT wanted a setup that could take on more work, show where each answer came from and pass uncertain cases to a person.

What we built

We designed a team of AI agents led by a supervisor agent. The supervisor reads each task and routes it to agents that specialize in different areas of law. Answers draw on a legal knowledge base built for VAIOT, so each one cites its sources. Every answer gets a confidence score, and anything below the threshold goes to a person. Every step is logged: the source, the score and any hand-off. The system runs on Azure behind private endpoints, with no data going out to the public internet.

Deliverables

  • Supervisor agent that routes work to specialist legal agents

  • Legal knowledge base with source citations (RAG)

  • Confidence scoring, with a hand-off to a person when the score is low

  • Event logs and documentation for every generated answer

How it works

  1. The right specialist for each task.

    The supervisor agent splits the work and sends each part to the agent for that area of law, based on the task type and the confidence score.

  2. Every answer shows its source.

    Agents work from VAIOT's legal knowledge base and cite the documents they used, so a lawyer can check them.

  3. When the system is unsure, a person decides.

    Each answer carries a confidence score. Below the agreed threshold, it is escalated to a person for review.

  4. Private, and logged step by step.

    It runs on Azure Kubernetes behind private endpoints, and each step is recorded with its source link, confidence score and any escalation.

Results

faster case processing
70%
lower LLM token cost
38%
users in the first week of the live demo
500+
from concept to production-ready
12 weeks
  • Average turnaround48 before, 14 after

  • Legal reasoning is split across specialist agents that work together

  • Faster turnaround, with control and traceability kept

  • Clear rules for when a person reviews the output, suited to legal work where mistakes are costly

12 weeks from concept to production-ready

  1. Weeks 1–2

    Discovery and risk mapping

    Mapped legal workflows, constraints and the checkpoints where a person reviews the output.

  2. Weeks 3–6

    Agent architecture

    Built the supervisor routing, the legal agents and their grounding in the knowledge base.

  3. Weeks 7–10

    Validation and controls

    Set confidence thresholds, escalation rules and quality tests.

  4. Weeks 11–12

    Go-live readiness

    Hardened the deployment and delivered the documentation.

Technology

  • Python
  • AutoGen
  • Azure OpenAI
  • Azure Kubernetes
  • FastAPI
  • PostgreSQL
  • RAG

Show us the workflow your lawyers still do by hand.

Tell us what you're building