many are eager to access MBSE tools and training, yet the default seems stuck in legacy, document-based workflows. We currently can use AI tools to interact with SysML v1 with limited success. However, there are no auto - mated loops to use AI reliably to in - terface with the model. Existing capa - bilities are one-way. AI can be used to generate files in a CSV (Comma-Sep - arated Values) or XMI (XML Metadata Interchange) format and manually load them into the model. This has capacity and quality limitations, par - ticularly with XMI, where file uploads often fail or are incomplete. The other side of this is export - ing model data for AI analysis. These methods may include copying images of the SysML diagrams or exported artifacts such as XMI. However, this again requires manual involvement to create the file and import it into the AI tool. An additional limitation is that, though AI can be used to query the model artifact, there is no reliable way to modify the information and import it back into the model. AI tools are being used with SysML v1, but they have significant limita - tions due to the architecture. SysML v1 lacks standardized Application Programming Interfaces (APIs), leav - ing AI agents without a reliable pro - grammatic pathway into a SysML v1 model. The authoritative data is in the model. However, it is buried within levels of complexity, and the data connections are not capable of real-time AI augmentation. Phase 3. Data on Demand (SysML v2 and AI synergy). The SysML v2 stan- dard was released in July 2025 as a ground-up architectural redesign which directly addresses the data in - terface challenges. The new SysML v2 standard now includes text-based model formats and a standardized API that enable connecting DoW data into a true digital thread. First, the text-based model format now appears and behaves like code. Models can leverage our robust soft - ware development processes, such as
residing in configuration-controlled Git repositories and being edited with any text-editing tool. Further - more, the text-based format is per - fectly suited for use with AI tools and agents. AI has revolutionized software code generation, and we can now le - verage the same model to develop systems in a code-like, text-based format. The standardized API in SysML v2 is equally significant. This stan - dardized API now opens a defined data port into the model, enabling integration with AI agents and other digital tools. This will allow devel - opment of AI agents that can query, create, update, or delete model ele - ments through the common inter - face. This feature eliminates vendor lock-in, enables interoperability in the digital thread, and transforms the sys - tem model from a locked, proprietary database into a common, accessible data resource. A third consequence of this archi - tecture is model federation, which addresses one of the most persistent structural problems in DoW acquisi - tion. Under SysML v1, program models were often isolated silos. The DoW has a vision for evolving developmen - tal system models into digital twins to supercharge Operations and Sus - tainment; however, tool compatibil - ity limitations have caused friction, especially when connecting models across complementary domains,
such as Product Lifecycle Manage - ment and physics. SysML v2’s stan - dardized API completely changes this, enabling compliant models from dif - ferent organizations and tools to be linked across the digital thread. This seamless connection of models will create a federation in which data are connected and propagated up and down without manual intervention. AI agents can query the live model and state directly, generate content that a formal compiler validates be - fore it touches the program baseline, and operate within a governance framework controlled by the program office. The vision of combining SysML v2 with AI agents is powerful, but we must acknowledge and address the challenges it presents. AI hallucinations. When using AI tools for warfighting acquisition pro - grams, hallucinations are a real and critical concern. Generative AI has the potential to generate answers based on its training data and to infer plausi - ble answers rather than reliably com - municate the authoritative informa - tion in the model. DoW systems have mission-criti - cal, life-and-death implications, so all information must be treated as pre- decisional until verified by a human expert. AI will not replace the need for a skilled and responsible acquisi - Navigating the Real Hurdles (Caveats)
SEPTEMBER –OCTOBER 2026 | DEFENSE ACQUISITION MAGAZINE 45
Made with FlippingBook - Online Brochure Maker