AI in CUSP
How we build responsibly
3Si applies AI to clearly defined problems where its role is bounded, transparent, and reviewable. This reflects our commitments to responsible design, pre-deployment testing, ongoing monitoring, data protection, privacy, accountability, and human judgment. We also seek specific approval from our clients prior to deploying any form of AI within client deployments. How this looks in practice: every 3Si client is provided a separate, dedicated, and secure cloud environment. No information is shared between clients, and no information is sent to public AI services or unauthorized external systems. All AI interaction with client data is explicitly approved by the client, and prompts and results are never used to train any model.
AI in Action: Natural Language Data Querying
CUSP allows authorized state users to ask questions about their data in everyday language and get answers drawn straight from their own database. Here's how the process works:
AI grounded in your CUSP context
- A tailored AI model maps the definitions, logical relationships, and rules across your specific CUSP deployment without interacting with the actual data. This gives the AI precise context for every question.
From natural language to validated SQL
- Users ask the AI model specific questions, and the AI model converts these questions into a SQL query. The SQL query is then used to generate the answer, not the AI model. Your data team can inspect, validate, refine, and reuse the SQL query and its results as desired.
No hallucinations, by design.
- Because the model never produces the answer — only a query your data answers — there is no possibility of hallucination. If a question cannot map to a valid, safe query, the system says so.
Optional, and off unless you turn it on.
- Teams that prefer pure SQL can continue working exclusively in an AI-free SQL workspace over the same data.
The result: broader access to CUSP data for program and policy leaders and frontline decision-makers — with the same rigor and traceability analysts expect, and every number backed by a query anyone on the team can verify.