AI Capabilities & Limitations

This page documents the intended use, deployment contexts, oversight expectations, and known limitations of the government-facing software offering built on the Trunnion AI multi-agent orchestration framework. It exists so that buyers evaluating the offering do not have to infer scope from marketing copy.

Intended use

The offering is intended for workflow orchestration, document and data processing, procurement and operations automation, and analytical decision support in government, defense, and regulated enterprise environments, always operating under the governance controls described below. It is designed to accelerate and structure work performed by qualified people. It is not intended to make, and this public site does not provide, eligibility, employment, procurement-award, security-clearance, credit, housing, health, legal, or other consequential decisions without human review.

Deployment contexts

The architecture is designed for deployment on commercial cloud, AWS GovCloud, private cloud, on-premise, or fully air-gapped environments with locally hosted models, matched to the security posture a program requires. These are design capabilities of the architecture; the accreditation, authorization, and classification posture of any specific deployment is established per engagement through the applicable official process, not by this website. Formal CMMC certification and FedRAMP authorization are not currently held; see the Security & Trust page for authoritative status.

Human review expectation

Human-in-the-loop approval gates are a required design element of the platform, not an optional feature. AI outputs can be inaccurate, incomplete, biased, outdated, or unsuitable for a given purpose, and require qualified human review before reliance. Deployments are expected to route consequential actions through named human approvers, and audit trails record both agent actions and human approvals.

Known limitations

Like all systems built on large language models, the platform can produce confident but incorrect output, can be sensitive to prompt and context quality, and depends on the quality and currency of the data sources connected to it. Model behavior varies across the underlying LLMs a deployment selects. Performance claims about specific deployments require deployment-specific evaluation; capability descriptions on this website describe what the architecture is designed to support, not a certified test result for your environment. Evaluation and acceptance testing are scoped per engagement.

Governance controls

The platform is designed around governed execution: human-in-the-loop approval gates, tamper-evident audit logging, role- and attribute-based access control, and classification-aware workflow design for regulated environments. Documentation of these controls for a specific procurement is available to contracting officers and evaluators on request at govops@viceroynm.com.

See also: Software & Applied AI · Trunnion AI platform · Website Terms · Security & Trust