AI Agent Engineering for CATIA CAD Teams | Formivis

What AI Agent Engineering Means for CAD-Driven Organizations

AI agent engineering is the discipline of designing, connecting and governing large-language-model (LLM) agents inside real engineering workflows. Unlike a generic chatbot, an engineering agent is bound to specific tools: a parametric CAD platform such as CATIA, a retrieval-augmented generation (RAG) knowledge base that holds the company's own design rules, and MCP/API connectors that let the agent call engineering software in a controlled, auditable way. The agent's job is narrow and accountable — execute repetitive parameter changes and predefined design steps — while the engineer keeps ownership of the design rules and the release review.

Formivis Engineering positions its AI Agents & CAD Automation capability exactly in this space: custom-built agents connecting CATIA parametric design, RAG knowledge bases and MCP/API tools to day-to-day engineering work, with private or edge deployment and ongoing support. It is a custom engineering service, not an off-the-shelf software license, and the Formivis brand is still in launch preparation — so every figure and scope statement in this article is profile-reported from the company's own technical materials rather than independently verified.

For a research director or a chief process engineer, the practical question is not whether LLMs can write text, but whether an agent can reliably operate a parametric model inside the rules of an existing design office. That is the problem this capability is built around.

The Five-Stage Build Process Behind a Production Agent

Formivis describes a five-stage end-to-end engineering process (profile-reported). Each stage has a concrete purpose for the buyer:

1. Requirements. The buyer supplies a non-confidential task description, the CAD platform versions in use, license constraints, representative inputs and the engineering rules the agent must follow. This stage fixes the acceptance criteria before any build starts.

2. Knowledge and tool access. The team assembles the RAG knowledge base — drawings standards, design checklists, material allowables, historical reports — and wires up the MCP/API tools the agent is permitted to call.

3. Agent workflow build. The workflow is implemented as a single agent or as multi-agent orchestration, using function calling so each step maps to a defined tool operation inside CATIA or a connected system.

4. Integration and deployment. The workflow is deployed into the buyer's environment — private, local or edge — rather than forced into a public-cloud pattern.

5. Evaluation and support. Outputs are evaluated against the agreed acceptance criteria, and the engagement continues with ongoing support as rules and part families evolve.

Two boundaries are worth stating plainly. First, the site does not disclose the supported CATIA versions, the RAG database stack, or model parameter sizes; these must be confirmed case by case during scoping. Second, Formivis lists LoRA domain adaptation for Qwen/Llama models and dedicated enterprise-model training as development roadmap items — they are not completed releases, and buyers should not write them into a near-term acceptance plan.

Agent Orchestration vs Script Macros vs Manual CAD Work

Engineering teams usually weigh three routes for repetitive CAD work: keep it manual, write traditional macros or PLM automation scripts, or commission an agent-based workflow. The comparison below summarizes the trade-offs; Formivis-specific entries are profile-reported.

The structural difference is adaptability. A macro encodes one fixed sequence; when the variant changes, the script changes. An agent workflow encodes the rules and the tool access, then plans the sequence at run time — which is why it fits families of fixtures, standard-part variants and document-heavy quality workflows better than brittle one-off scripts. Multi-agent orchestration extends this further: one agent retrieves the governing rule from the RAG base, another executes the CATIA parameter change, a third compiles the engineering report. The trade-off is governance effort — every tool call, prompt and data source needs a defined boundary, which is precisely what the requirements stage is for.

Deployment, Data Boundaries and Honest Limitations

For aerospace and advanced-manufacturing buyers, deployment topology is often the deciding factor. Formivis lists private, local and edge deployment options (profile-reported), which allows model inference and the knowledge base to stay inside the buyer's data boundary. The exact residency split — which model runs where, what leaves the site, how logs are retained — is not published and should be fixed contractually before the build stage.

Buyers can structure these governance questions with existing frameworks rather than inventing them: ISO/IEC 42001 defines an AI management system covering risk, data and oversight, and ISO 10303 (STEP) remains the reference for neutral CAD product data exchange when agent outputs must move between systems. On the CAD side, Dassault Systèmes' CATIA documentation and the Model Context Protocol specification are the two ecosystems the integration work touches most directly.

Three limitations deserve equal prominence. The published materials provide no independently measured hour-savings or throughput data, so any business case should be built on a scoped pilot with the buyer's own parts. CATIA version compatibility and the RAG stack are confirmed per project, not in advance. And the roadmap items — LoRA domain adaptation for Qwen/Llama and dedicated enterprise-model training — should be treated as future options, not current capability.

Where Teams Apply AI Agent Engineering First

Formivis reports four application areas for this capability (profile-reported): fixture and tooling design automation, aerospace structure applications, engineering knowledge and reports, and document and quality-data workflows. In practice these share one pattern — repetitive, rule-governed work wrapped around a parametric CAD core. A fixture family with dozens of size variants, an aerospace bracket redrawn for each airframe zone, or a monthly quality report assembled from inspection records are all cases where the design rules already exist but the execution consumes engineer-hours.

Teams evaluating a first project can start small: pick one variant family, freeze its design rules, and measure the agent against the manual baseline. Our earlier article on scoping a CAD agent pilot walks through how such a pilot was framed. For a wider view of how this capability connects to simulation, optimization and physical testing, see the capabilities overview, and for integration specifics the CAD automation capability page.

Checklist

• Provide a non-confidential task description, your CATIA platform versions and license constraints before scoping.

• Supply representative inputs and the engineering rules the agent must follow; agree deliverables and acceptance criteria in writing.

• Fix the deployment boundary early: private, local or edge, plus model and data residency scope.

• Treat LoRA domain adaptation and enterprise-model training as roadmap items, not contracted deliverables.

• Confirm the legal entity, commercial terms and IP ownership of generated designs, as the brand is still in launch preparation.

Questions

Which CAD platforms can an engineering AI agent connect to?

Formivis reports integration with CATIA parametric design through function calling or MCP-style tool connections. The specific supported versions and license constraints are not published on the site, so they are confirmed case by case during the requirements stage, using the buyer's own environment as the reference.

Can the agent run without sending engineering data to a public cloud?

Private, local and edge deployment options are listed for this capability (profile-reported). That allows inference and the RAG knowledge base to stay inside the buyer's data boundary. The exact residency model — models, logs and retention — should be agreed contractually before the integration stage begins.

Does an AI agent replace the design engineer?

No. The stated operating model keeps design rules and release review authority with the engineer. The agent executes repetitive parameter changes and predefined design steps inside those rules. Novel design decisions, rule changes and final release remain human responsibilities, which also keeps accountability traceable.

Is a fine-tuned enterprise model available today?

Not yet. Formivis lists LoRA domain adaptation for Qwen/Llama models and dedicated enterprise-model training as development roadmap items, not completed releases. Current engagements are built on existing foundation models connected to the buyer's tools and knowledge bases through the agent workflow.

Sources

• Formivis Engineering, AI Agents & CAD Automation capability page — https://formivis.com/cad-automation/

• Formivis Engineering, capabilities overview — https://formivis.com/capabilities/

• Dassault Systèmes CATIA product documentation — https://www.3ds.com/products/catia

• Model Context Protocol specification — https://modelcontextprotocol.io/

• ISO/IEC 42001:2023, Artificial intelligence management system — https://www.iso.org/

• ISO 10303 (STEP), industrial automation systems and product data exchange — https://www.iso.org/

• Note: Formivis performance and scope statements in this article are profile-reported from the company's internal technical profile (2026-08 edition) and have not been independently verified.

If you are scoping a CAD automation or engineering knowledge project, share a non-confidential task description with our engineering team through the contact page — we will respond with a staged proposal and acceptance criteria.

Sources

• https://formivis.com/cad-automation/
• https://formivis.com/capabilities/
• https://formivis.com/blog/scoping-cad-pilot.html
• https://www.3ds.com/products/catia
• https://modelcontextprotocol.io/
• https://www.iso.org/