ARETGroup

Let's take a closer look at the important role of private AI systems as a driver for innovation in public safety, justice and interagency human services collaboration. While cloud-based AI models foster innovation, transparency, and broader participation, unrestricted openness can introduce substantive risks.

Especially when employing public safety records, vulnerable populations data or data associated with crisis events, further inspect use of any AI. Where cloud-based AI systems are not an option, think of what you can do with a private AI tool.

What is a Private AI Sandbox? 

As defined, a private or ‘closed’ AI system is an artificial intelligence platform or feature operating within a restricted, proprietary, or non-transparent framework. These systems are typically developed and controlled in a way that limits user access to underlying code, training data, or model outputs. Here, research and evidence-based practice teams introduce sensitive information (e.g. agency notes, case records and forms) to artificial intelligence models for ethical experimentation, pilot studies, or trials before full-scale deployment— without ever exposing raw data outside secure boundaries. 

General Use Cases

Controlled and Experimental Environments. Closed or private AI sandboxes offer a secure, policy-compliant setting for experimentation. Within this environment, models can be exposed to real-world data. This structure reduces common human and institutional risks that limit the scope of experimental study in human services. 

Workflow Optimization. AI sandbox models can be deployed to test and streamline professional workflows. Human experts—such as social workers, dispatchers, clinicians, case managers, nurses, or officers—are able to review and interpret each output. This collaborative loop enables private AI systems to target and reduce redundant, repetitive, or low-value administrative work. For example, think of using AI to process hundreds of pages of case information and notes, avoiding redundant and error-prone work. 

Continuous and Iterative Learning. Each human interaction within the sandbox—whether a data input, review or decision—feeds into the ongoing refinement of the AI system. The sandbox thus becomes a dynamic learning ecosystem.

Commercialization and Scalable Impact. Validated sandbox results provide an evidence base for broader testing and/or market adoption. Successful pilots demonstrate technical feasibility, measurable outcomes, and ethical readiness for broader go-to-market planning. 

 Case Study:  AI for Accelerated Crisis De-Escalation and Deflection Response

Across the nation, community justice and public safety providers face mounting pressure to provide a timely and adequate response to behavioral health crises. For many communities, crisis intervention, mobile crisis, and alternative response teams offer an effective, legally appropriate, and efficient solution. Traditional data systems, however, have failed to keep pace with the demands of emerging social work and community response teams. Teams often lack information systems capacity for a coordinated response. Commonly this leads to incomplete records and missed opportunities for coordinated care. 

Commonly, teams are looking for better tools for collection and coordination of field contact and referral notes. Real-time, field notes have the potential to address a major gap in information management. When consistently maintained for each case, field contact notes contain a longitudinal record of evolving individual risks, needs and outcomes reported through first-hand observation rather than self-reported surveys. Social work teams embedded in frontline service delivery not only need these notes to be maintained and retrievable, they are looking for updated technology features that streamline workflow and offer quick information. 

Inspired by the notion of private AI sandboxes for a justice and mental health collaboration, AI-Insight  technology helps to streamline workflow for notes processing and offers quick information to teams in the field. Data from multi-state trails are currently yielding important clues for crisis response teams. In this quick video, ARETGroup AI-Insight interacts with authentic field notes, provided by a behavioral health response team. AI-Insight is used for instant case summary and open search capacity to retrieve information across the case history.  

Our Results

• Faster Referrals: AI-Insight has potential to reduced processing time from field contact to behavioral health referral

• Improved Interagency Coordination: Summary reports generated can be retrieved and shared;

• Ethical Oversight: AI-generated recommendations are reviewed by human professionals, preserving accountability and decision-making authority.

• Actionable Insights: Aggregated data from repeat visits and contacts all-in-on place helped to identify service gaps and recurring crisis types, informing future training and deployment strategies.

Conclusion 

AI adoption is advancing rapidly among public agencies. This article outlines the value of the private AI sandbox as a model for further use and exploration. We feature an innovative approach to managing community response information for improved crisis outcomes. In closing, keep in mind the need for robust governance models for any research or exploratory project. These should include:

• Data Use Agreements (DUAs): Clear rules on how private data is collected, stored, and used inside the sandbox.

• Oversight boards: An ethics or community advisory board reviews pilot progress, ensuring community voices are heard.

• Transparency measures: Even if raw data cannot be shared, agencies can publish plain-language summaries of what the model does, its limits, and its risks.

• Incident protocols: If the AI produces harmful or biased results, the sandbox team can roll back the model immediately.

• Human-in-the-loop requirements: Experimental AI outputs should never used without human review, especially in early pilots.

• Use de-identified datasets whenever possible.

• Apply strict access controls (role-based, need-to-know).

• Require data encryption and use agreements as well as participant consent when needed.