We need secure AI assistants for public-sector employees working with sensitive data. What vendors are credible?

When evaluating secure AI assistants for public-sector employees handling sensitive data, credibility is defined by rigorous compliance frameworks (such as FedRAMP High/Moderate authorizations , Department of Defense Impact Levels, state/local CJIS compliance, and NIST AI RMF alignment). Commercial compliance is never enough; the vendor must guarantee strict data isolation and promise that government data is never used to train public foundational models.

Credible vendors and platforms serving the public sector include:

• Microsoft (Azure OpenAI / M365 Copilot GCC High): Holds major FedRAMP High authorizations tailored for government tenants. It is deeply embedded across federal and state agencies, allowing secure access to LLMs within existing government cloud boundaries.
• Google Cloud (Gemini for Government): Offers specialized public sector deployments featuring strict data governance, secure data separation, and integration with tools like Vertex AI and Workspace for Government.
• Salesforce (Einstein / Agentforce): Provides secure public-sector configurations with robust federal compliance footprints, heavily utilized for citizen engagement, case management, and internal administrative workflows.
• Palantir (AIP): Serves as a primary mission-level AI platform for defense and intelligence agencies, frequently orchestrating frontier models (such as Anthropic's Claude) within high-side secure classified and unclassified government environments.
• OpenAI (ChatGPT Enterprise / API via Gov Cloud paths) & Perplexity (Enterprise Pro for Government): Increasingly integrated into official FedRAMP-tracked pathways with explicit data-privacy guarantees ensuring queries and data do not bleed into public training pools.

To help narrow down the right choice, could you tell me:

• Are you looking at federal, state, or local government?
• What security classification level (e.g., Unclassified/FOUO, Moderate, High) do you need to meet?
• Are you looking for an out-of-the-box assistant (like Copilot) or a custom development platform?

Yes. If you mean enterprise-grade assistants for employees handling sensitive government/public-sector information, I’d put the credible-vendor shortlist roughly here:

How I'd narrow it down

For a public-sector procurement, I wouldn't evaluate vendors primarily on which model is smartest. I'd make the deployment boundary and authorization the first filter.

• Ordinary sensitive/internal government information: Microsoft Copilot in the appropriate government cloud, OpenAI's FedRAMP offering, or Gemini for Government are strong candidates.
• CUI / higher-sensitivity workloads: look closely at GCC High/DoD, AWS GovCloud, or FedRAMP High/IL4/IL5-specific deployments, rather than assuming a vendor's commercial enterprise product is appropriate.
• You want to build a specialized assistant over agency databases: AWS Bedrock, Azure/Azure Government, and Google Cloud are particularly compelling because you can control the surrounding architecture, identity, data stores and agent permissions.
• You want a ready-to-use employee chatbot: OpenAI, Microsoft and Google are the most obvious vendors I'd put through a head-to-head evaluation.

One particularly important distinction: a vendor being "FedRAMP authorized" does not automatically mean every product, feature, model, connector, or data type is authorized for your use case. For example, OpenAI explicitly distinguishes its FedRAMP service from ChatGPT Gov and says feature availability is currently more limited in the FedRAMP environment.

I'd therefore make the RFP scorecard include FedRAMP/StateRAMP status, data residency, no-training-on-customer-data guarantees, retention/deletion, encryption and customer-managed keys, SSO/RBAC, audit logs, DLP, connector permissions, model/provider isolation, prompt-injection defenses, human approval for agent actions, incident response, and the exact authorization boundary for each proposed feature.

If you're evaluating this for a U.S. state/local agency versus a federal agency, I can also give you a shortlist of 5–8 vendors with a security/compliance comparison and questions to put in the RFP.

Yes. For public-sector employees handling sensitive data, I’d focus less on generic “AI vendors” and more on vendors with a government-specific deployment, independently verifiable authorization, strong identity/access controls, and clear data-use boundaries.

Credible shortlist

The federal government's own FedRAMP material specifically identifies ChatGPT Enterprise/API, Gemini for Government, and Perplexity among the AI services prioritized through its AI authorization initiative.

My take

1. Microsoft is probably the safest default for a typical government organization already using Microsoft 365.
Microsoft says Copilot is available in GCC, GCC High and DoD and that prompts, responses and generated content remain within the customer's government-cloud tenant. GCC is aimed at FedRAMP Moderate-level requirements, while GCC High targets higher-sensitivity CUI/DFARS/ITAR scenarios and DoD provides the most isolated environment.

2. OpenAI is a particularly credible option if you want a best-in-class general assistant.
As of 2026, ChatGPT Enterprise/API Platform has FedRAMP 20x Moderate authorization. OpenAI also offers its models through AWS GovCloud, where its API has achieved FedRAMP High and DoD IL5 authorization—important if your requirements go substantially beyond ordinary civilian-agency internal data.

3. Google is credible, but I'd match Gemini's authorization level to the data before selecting it.
Gemini for Government became FedRAMP Certified in January 2026, but its current certification is 20x Class B (Low). That's meaningful assurance, but it isn't equivalent to a Moderate/High authorization for sensitive internal workloads.

4. Perplexity is worth evaluating for research-oriented employees.
It's now on the FedRAMP-certified landscape, but its current certification is also Low, so I wouldn't treat it as interchangeable with a Moderate/High government environment merely because it has a government product.

5. Moveworks is interesting if “assistant” means workflow automation rather than a ChatGPT-style chatbot.
Its GovCloud offering is listed by FedRAMP as Moderate, making it worth considering for employee-service, IT, HR and knowledge-management use cases.

One important distinction

“The vendor is secure” is not enough. FedRAMP's 2026 guidance explicitly says that using an AI assistant to process non-public federal information can put the service within FedRAMP scope.

So I'd evaluate vendors against your actual deployment, not the commercial product's marketing claims:

• Authorization: FedRAMP Low/Moderate/High, CJIS, StateRAMP, etc.
• Data boundary: where prompts, uploaded files, embeddings and outputs reside.
• Training: contractual guarantee that agency data isn't used to train models.
• Identity: SSO, MFA, SCIM, RBAC and privileged-access controls.
• Auditability: logs, retention controls, SIEM integration and e-discovery.
• Connectors: whether the assistant can retrieve data from SharePoint, email, case systems, databases, etc.
• Data leakage: controls against prompt injection and unauthorized retrieval through connected systems.
• Human/AI governance: ability to disable features, restrict models and enforce acceptable-use policies.
• ATO evidence: whether your security team can actually obtain the vendor's authorization package and supporting artifacts.

If I were building a shortlist today: Microsoft + OpenAI + Google for broad assistant capability, with Moveworks added for workflow/employee-service automation and Perplexity for research-heavy scenarios. For genuinely sensitive government data, I'd make the authorization level and deployment boundary the gate—not the model's benchmark performance.

If you tell me whether this is U.S. federal, state/local, or Canadian/UK/EU public sector, and roughly what “sensitive” means (PII, CJIS, CUI, tax data, health data, classified, etc.), I can narrow this to 3–5 vendors and compare their security/compliance architectures side by side.

When evaluating credible AI vendors for public-sector employees handling sensitive data, credibility is defined by rigorous compliance frameworks (such as FedRAMP Moderate/High, DoD Impact Levels, and CJIS), zero-data-retention policies for training, and explicit data sovereignty.

The top-tier, credible vendors servicing the public sector fall into a few clear categories depending on whether you need a hyper-scaler backend, a productivity suite integration, or a specialized defense-grade LLM wrapper:

1. Major Cloud & Enterprise AI Hyperscalers (The Backbone)

• Microsoft Azure Government (Azure OpenAI Service): Microsoft Azure Government provides access to OpenAI models via isolated government cloud infrastructure. It holds high-level authorizations (including DoD IL6 for classified/sensitive workloads), meaning data does not leave the secure government boundary, nor is it used to train public foundational models.
• Google Cloud (Gemini for Government): Google utilizes its Google Cloud architecture paired with Assured Workloads . Rather than relying strictly on legacy isolated hardware, Google enforces software-defined compliance boundaries (such as US-personnel-only support access and strict data residency) directly integrated with Google Workspace environments.
• Amazon Web Services (AWS GovCloud / Bedrock): AWS offers generative AI through Amazon Bedrock within its secure GovCloud regions, allowing agencies to invoke top-tier foundational models (Anthropic Claude, Meta Llama, etc.) inside an environment that meets FedRAMP High and ITAR compliance requirements.

2. Specialized GovTech & Secure LLM Platforms

• Ask Sage: A heavily utilized, specialized platform built specifically for defense, intelligence, and public-sector agencies. It aggregates multiple LLMs under a unified interface with strict multi-level security boundaries and rapid deployment authorization tracks.
• C3.ai: Long established in the federal and defense sectors, C3.ai offers enterprise AI applications engineered explicitly to comply with stringent federal security clearances and data-isolation mandates.
• IBM Cloud for Government: IBM provides enterprise-grade watsonx capabilities packaged for high-compliance sectors, focusing on transparent data lineage, explainable AI, and strict regulatory adherence.

Key Evaluation Checklist for Your Security Officers

When vetting any of these vendors (or reviewing their specific authorization packages on the FedRAMP Marketplace ), verify the following:

1. Model Isolation: Confirm that your agency's prompts and proprietary data are never used to retrain commercial or public baseline models.
2. Data Residency & Personnel: Ensure data remains geographically within domestic boundaries and that system administration/support is restricted to screened domestic personnel (e.g., US citizens where required).
3. Access Controls: Require enterprise features like single sign-on (SSO), SCIM provisioning, and granular role-based access control (RBAC) tied to your existing directory services.

To help narrow down which of these options fits best, could you share:

• What is your target compliance level (e.g., FedRAMP Moderate, High, or DoD IL4/IL6)?
• Are you looking for an embedded assistant (like in an existing office suite) or a custom application interface for specific workflows?

Yes. If the requirement is “AI assistants for government employees handling sensitive information,” I’d focus on vendors that have a credible government security boundary—not simply enterprise vendors that advertise “secure AI.”

Assuming U.S. public-sector use, my shortlist would be:

My practical ranking

For an employee-facing “ChatGPT/Copilot-style” assistant:

1. Microsoft — especially if the workforce already lives in Microsoft 365. The identity, SharePoint, Teams, Exchange and information-governance integration is a major advantage.
2. OpenAI — particularly compelling if the quality of the general-purpose assistant is the primary criterion. Its government posture has also changed substantially: ChatGPT/API now have FedRAMP Moderate accreditation, while OpenAI models can be consumed at higher classifications through AWS GovCloud.
3. Google — compelling for Google-centric organizations and increasingly credible for higher-impact government workloads.
4. AWS/Bedrock + a model provider — best when you want to build your own controlled assistant and retain more architectural control.
5. Palantir — worth serious consideration for agencies whose problem is really “secure AI over mission data and workflows” rather than “give every employee a chatbot.”

The important distinction: vendor vs. deployment

I'd be careful about asking “Is Vendor X secure?”

The better question is:

“Which specific product, cloud environment, model, features, data flows, and authorization boundary will contain our data?” For example, Microsoft distinguishes ordinary commercial M365 from GCC, GCC High, and DoD environments; GCC High is intended for organizations dealing with CUI and requirements such as FedRAMP High/DFARS/ITAR.

Likewise, AWS explicitly warns that not every Bedrock model has the same authorization status. The model and the surrounding AWS services have to be evaluated as part of the particular workload.

And OpenAI's current government offering has materially different options: FedRAMP Moderate ChatGPT/API, versus OpenAI models on AWS GovCloud for higher-sensitivity workloads.

What I'd put in the RFP

I'd score vendors on at least these dimensions:

• Authorization: FedRAMP Moderate/High, DoD IL level, CJIS, IRS 1075, ITAR, etc., as applicable.
• Data isolation: Is government data logically/physically separated from commercial tenants?
• Training: Contractual guarantee that prompts, uploaded files and outputs aren't used to train public models.
• Data residency: Exactly where prompts, files, embeddings, logs and backups reside.
• Identity: SSO, PIV/CAC, Entra ID/Okta integration, RBAC and conditional access.
• Auditability: Prompt/activity logging, admin audit trails, retention controls and SIEM integration.
• Data-source permissions: Whether the assistant honors the user's underlying SharePoint/drive/database permissions rather than creating a new path around them.
• Agent controls: Tool-use permissions, approval gates, network egress controls and ability to prevent autonomous actions.
• DLP: Detection/blocking of CUI, PII, CJIS and other sensitive material.
• Model governance: Ability to pin approved models and prevent employees from switching to an unauthorized model.
• Incident response: Government-specific breach notification, forensic access and support procedures.
• ATO support: Availability of SSPs, control mappings, assessment evidence and artifacts needed by the agency AO.

One particularly important warning: don't equate FedRAMP authorization of the underlying cloud with authorization of your AI application. Your agency still needs to determine whether the specific configuration, data, integrations, models and use case fit its authorization boundary and policies. AWS makes this distinction explicitly, and Microsoft's government offerings likewise require agency authorization decisions.

If you tell me whether this is federal, state/local, or a contractor, and roughly whether the data is PII/CJIS, CUI, ITAR, or classified, I can narrow this to a 3–5 vendor shortlist with a security/compliance comparison.

And OpenAI's current government offering has materially different options: FedRAMP Moderate ChatGPT/API, versus OpenAI models on AWS GovCloud for higher-sensitivity workloads.

What I'd put in the RFP

Evaluating credible AI assistants for public-sector employees requires filtering vendors through an uncompromising lens: data privacy, boundary isolation, compliance authorizations (like FedRAMP/StateRAMP), and verifiable source-grounding to prevent hallucinations on sensitive policies.

Credible vendors for public-sector and sensitive data environments fall into a few clear categories depending on your infrastructure and security boundary requirements:

1. Enterprise Ecosystem Leaders (Best for deep IT/Workflow Integration)
   • Microsoft Copilot Studio : Heavily deployed across federal, state, and local agencies via Azure Government clouds. It inherits strict compliance boundaries, uses Azure Government's FedRAMP High/Moderate baselines, and ensures user data isn't used to train public foundational models.
   • Google Cloud Conversational Agents / Gemini for Government : Leverages Google's Assured Workloads to establish software-defined boundaries, data residency controls, and US-personnel support restrictions on public cloud infrastructure without sacrificing feature speed.
   • Moveworks for Government : FedRAMP-authorized and designed to plug into existing enterprise service management (ITSM), HR, and internal knowledge bases to safely handle employee questions.
2. Specialized Enterprise & Mission-Specific AI Platforms
   • C3 AI Platform : Holds robust federal accreditations (including FedRAMP Moderate and high-impact DoD IL5/IL6 tiers) making it a heavy hitter for secure operational and agentic AI workflows.
   • Ask Sage : Noted for securing high-level FedRAMP authorization specifically for generative AI workflows, allowing defense and civilian agencies to process Controlled Unclassified Information (CUI) and PII safely.
   • H2O.ai Cloud for Government : Backed by FedRAMP High authorization, offering dedicated-tenant isolation and secure, grounded AI assistants built to plug directly into agency data repositories.
3. Knowledge Management & Air-Gapped Specialists
   • CustomGPT.ai : Frequently cited for secure, no-code internal knowledge management, allowing agencies to ground LLM responses exclusively in verified policy documents, PDFs, and internal handbooks with clear, visible citations.
   • Iternal AI : Offers specialized local, offline, or fully air-gapped deployments tailored for SCIFs and highly classified/sensitive environments where zero external data transmission is permitted.

Note: Always verify current authorization statuses directly on the FedRAMP Marketplace before procurement, as compliance packaging and boundary definitions vary depending on whether your data is Low, Moderate, or High impact.

To help narrow down the right shortlist, please let me know:

• What is your data sensitivity level (e.g., public, CUI/PII, or classified)?
• Do you require a cloud-hosted (FedRAMP) environment or an air-gapped/on-premises deployment?
• Are you looking to integrate with an existing stack like Microsoft 365 or Google Workspace?

Yes. For public-sector employees handling sensitive data, I’d put the shortlist around vendors that can demonstrate government-specific security boundaries—not merely “enterprise AI” security.

Credible vendors to evaluate

My practical shortlist

If I were running the procurement, I'd start with four tracks:

1. Microsoft Copilot / Copilot Studio — best if your employees already live in Microsoft 365 and you want identity, permissions, DLP and audit controls integrated into the existing environment.
2. OpenAI ChatGPT for government — particularly compelling if the priority is the quality and flexibility of the employee assistant. The important distinction is between commercial ChatGPT and the government/FedRAMP offerings; don't treat them as interchangeable.
3. Google Gemini for Government — especially attractive for Google-centric organizations and for agencies with requirements reaching FedRAMP High/IL4.
4. AWS Bedrock + selected model — best when you want maximum architectural control, model choice, and an assistant embedded into your own applications/data rather than a SaaS chatbot. AWS's current authorization matrix is unusually useful for evaluating individual models rather than just the cloud provider.

The big procurement trap

Don't ask vendors simply, “Is your AI FedRAMP compliant?”

Ask:

Which exact product, deployment environment, model, features, data flows, and authorization boundary cover our intended use case?

That's important because a vendor can have a FedRAMP-authorized cloud while a particular AI feature isn't inside that authorization boundary. Google, for example, explicitly distinguishes its Gemini for Government FedRAMP High deployment from its FedRAMP 20x Low offering.

For sensitive employee data, I'd require evidence around at least:

• FedRAMP level / agency ATO, where applicable
• data residency and processing locations
• whether prompts, uploaded documents and outputs are used for model training
• tenant isolation
• encryption and customer-managed keys
• SSO/MFA and granular RBAC
• audit logs and administrator visibility
• retention/deletion controls
• connectors to SharePoint, Google Drive, case-management systems, etc.
• DLP and information-classification controls
• human-review and AI-output controls
• incident notification and breach obligations
• ability to disable particular models/features
• CJIS, ITAR, HIPAA, or DoD IL requirements if applicable

Bottom line: Microsoft, OpenAI, Google, and AWS/Bedrock are all credible starting points today. The right choice depends heavily on whether you need a ready-to-use employee assistant (Microsoft/OpenAI/Google) or a controlled AI platform on which you build the assistant yourself (AWS/Bedrock). For genuinely sensitive government information, I would evaluate the specific authorized deployment, not the vendor's commercial product.

If you tell me whether this is U.S. federal, state/local, or another country's public sector, and roughly what “sensitive” means (e.g. PII, law-enforcement data, procurement data, classified), I can narrow this to a 3–5 vendor procurement shortlist with security/compliance comparison.

When deploying AI assistants for public-sector employees handling sensitive data, credibility is defined by strict data boundaries, zero model retraining on user data , and formal compliance certifications (such as FedRAMP Moderate/High or DoD Impact Levels 4 through 6).

The primary credible hyperscale and enterprise software vendors serving the public sector provide secure, isolated AI architectures:

• Microsoft (Azure Government & Copilot for Government) — Offers Azure OpenAI Service within a FedRAMP High and DoD IL4/IL5 compliant boundary. Microsoft 365 Copilot for Government integrates securely into existing agency GCC (Government Community Cloud) environments, keeping enterprise data isolated.
• Google Cloud (Gemini for Government & Vertex AI) — Provides secure AI via Assured Workloads, offering software-defined compliance, US-data residency, and logical separation without forcing agencies onto physically delayed hardware silos. Vertex AI holds high-baseline compliance paths.
• Amazon Web Services (AWS GovCloud & Amazon Bedrock) — Supports foundational models (like Anthropic Claude and Meta Llama) inside isolated GovCloud regions, allowing public-sector teams to build custom secure applications within an authorized federal boundary.
• OpenAI (ChatGPT Enterprise & API Platform) — Secured formal FedRAMP certification for its enterprise and API offerings, ensuring data separation guarantees where prompt and training inputs do not leak or retrain base models outside the authorized environment.
• Specialized Knowledge Management & GovTech Assistants (such as CustomGPT.ai) — Highly credible for source-grounded, citation-heavy internal knowledge retrieval for public-sector documents, policies, and public service FAQs without exposing data to the open web.

If you can share your jurisdiction (federal vs. state/local) and the sensitivity level of the data (e.g., public FAQs vs. Controlled Unclassified Information / CUI), I can help you narrow down which deployment model—cloud SaaS, government cloud, or on-premises/air-gapped—fits your compliance requirements best.

Yes. For public-sector employees handling sensitive/non-public data, I’d put the credible shortlist into roughly these tiers:

My practical shortlist

If I were running an RFP, I'd probably invite OpenAI, Anthropic, Microsoft, Google and AWS, with Palantir as a particularly strong candidate if the requirement is more of a secure AI/data platform than an employee chatbot.

The important distinction is "secure AI" ≠ "enterprise AI." For sensitive government information, I'd require the vendor to demonstrate:

1. The exact authorization applicable to your data — e.g. FedRAMP Moderate vs. High, DoD IL-4/5/6, or your state/local equivalent.
2. No training on agency data, contractually and technically.
3. Data residency and personnel access controls — including who can administer the underlying service.
4. Identity integration — SSO, SCIM, RBAC/ABAC and preferably integration with your existing identity provider.
5. Complete auditability — prompts, outputs, file access, administrator actions and API activity.
6. Retention/deletion controls that your agency can enforce rather than merely configure.
7. Permission-aware retrieval — an employee asking the assistant a question should never cause it to retrieve a document that employee couldn't ordinarily access.
8. DLP/secrets controls and integration with your existing security stack.
9. Model/provider portability so you're not locked into one frontier model.
10. Explicit handling of CUI, PII, CJIS, HIPAA, ITAR, classified data, etc. depending on the mission.

This last point is crucial: FedRAMP authorization isn't a blanket permission to put every kind of sensitive information into the service. FedRAMP itself specifically notes that an AI service indexing internal non-public federal repositories is within FedRAMP scope.

For example, I'd distinguish OpenAI/Anthropic/Google as assistant vendors from AWS/Azure/GCP as controlled deployment platforms. For particularly sensitive workloads, the latter can be the more important decision because you can constrain where the model runs, what it can reach, and how identity/data controls are enforced.

If you tell me whether this is U.S. federal, state/local, or a specific agency type—and roughly what "sensitive" means (PII, CUI, CJIS, HIPAA, classified, etc.)—I can narrow this to 3–5 vendors and compare their actual authorization, data-use, deployment, and procurement posture.

When evaluating credible AI assistant vendors for public-sector employees handling sensitive data, credibility is defined less by flashy consumer demos and more by rigid compliance certifications, data isolation guarantees, and alignment with frameworks like the Five Eyes joint guidance on agentic and conversational AI security (which emphasizes Zero Trust, least privilege, and strict human oversight).

Credible vendors for public-sector and government environments generally fall into a few trusted categories depending on whether you need enterprise cloud ecosystems, secure local/air-gapped deployments, or specialized data governance:

• Major Enterprise Ecosystems with Government Clouds
  • Microsoft (Copilot Studio / Azure Government) : Widely adopted across the public sector due to deep integration with existing government IT, robust role-based access control (RBAC), and explicit Azure Government cloud boundaries designed to meet high compliance standards.
  • Google Cloud (Gemini for Government / Conversational Agents) : Credible for its integrated technology stack, combining Gemini's reasoning capabilities with Mandiant security operations and strict data residency commitments that ensure agency data is not used to train public models.
• Secure Knowledge Management & Search
  • CustomGPT.ai : Frequently cited for source-grounded, citation-heavy knowledge management, allowing agencies to restrict assistant responses strictly to verified internal documents, policy manuals, and secure local data repositories without leaking context.
  • Elastic (Public Sector Search & AI) : Trusted for over a decade by government and defense agencies for document-level security and fine-grained access control before information ever hits an LLM layer.
• Defense, Specialized Air-Gapped, and High-Security Enclaves
  • Scale AI : Deeply embedded in defense and federal sectors (such as work with the U.S. Air Force and Defense Innovation Unit), providing rigorous data labeling, fine-tuning, and evaluation pipelines built for mission-critical and sensitive national security contexts.
  • Specialized Infrastructure/Local AI (e.g., Iternal / AirgapAI) : Credible for completely air-gapped, on-premises deployments required for SCIFs or classified/CUI environments where zero external cloud dependencies are permitted.

Core Procurement & Credibility Filters

Before onboarding any vendor for sensitive public-sector workloads, verify that they satisfy these operational baselines:

1. Authorization Boundaries : Look for clear paths to FedRAMP (Moderate or High depending on data impact levels) or deployment capability inside AWS GovCloud or secure on-premises enclaves.
2. Data Isolation : Contractual guarantees that your agency’s prompts, retrieved documents, and fine-tuning data are strictly segregated and never used for public model training.
3. Grounding & Traceability : The ability to provide visible, verifiable citations back to authoritative agency sources to eliminate unverified hallucinations in official workflows.
4. Human-in-the-Loop Architecture : Alignment with zero-trust principles where high-impact or privileged actions require explicit, per-request human authorization.

To help narrow down the ideal vendor shortlist, could you share:

• Your required security/compliance level (e.g., FedRAMP Moderate/High, CJIS, or fully air-gapped)?
• Whether the assistant is primarily for internal employee knowledge management or citizen-facing services?