AI Integrations
The AI Modeling Workbench is a native Sparx EA feature that lets you configure LLM providers within the EA client itself: separate from EA GraphLink, the MCP server, or any third-party extension.
This distinction matters because the Sparx EA AI ecosystem has multiple components that can easily be conflated. The AI Modeling Workbench is the EA client’s built-in AI configuration panel. It allows architects to connect the EA client directly to AI providers and use those models for tasks within EA. It is not EA GraphLink (which provides outbound connectivity from the repository to external tools). It is not Kernaro Assist (which is a separate in-EA extension with a richer feature set). It is not an MCP connection (which exposes the repository to external AI tools). It is a native feature of the EA client itself.
Best understood as: configuration that connects EA to AI models for use within EA. The other integrations covered in this section work in the reverse direction: they expose EA data to external AI tools. The AI Modeling Workbench keeps the AI interaction inside EA.
For architects who want to test multiple AI providers against real architecture tasks before committing to a production integration, the Workbench is a practical starting point. For organizations already committed to a specific integration path, the Workbench adds flexibility that may or may not be needed.
Multiple AI providers, configurable within the EA client, available to architects during modeling work.
The Workbench allows architects to configure AI model connections within EA: specifying endpoints, credentials, and model parameters for one or more LLM providers. Once configured, those models are available to assist with tasks within the EA client: documentation generation, element description writing, analysis, and other tasks that an LLM can help with when given EA model context.
The multi-provider capability is the distinctive feature. Rather than committing to a single AI provider at the client configuration level, the Workbench allows multiple providers to be configured simultaneously. Architects can test different models against the same tasks, select different providers for different task types, or maintain fallback options.
For organizations in evaluation phases, this is valuable: it allows side-by-side comparison of AI providers using actual architecture work, in the actual EA client environment, before any significant integration investment is made.
AI assistance is available within the EA modeling environment without requiring separate tools or logins. Documentation, descriptions, and analysis assistance are accessible during active modeling sessions. The multi-provider configuration means flexibility to use the model that works best for a given task type.
The evaluation capability is the primary management-level value. Before investing in a Connect engagement to deploy a production AI integration, the Workbench allows realistic comparison of AI providers using your actual EA data. This is a more grounded basis for integration decisions than vendor demos.
The Workbench is a lower-commitment entry point into AI-augmented architecture. No EA GraphLink deployment, no external integrations, no additional licensing beyond EA and API access. This allows organizations to begin AI-augmented architecture work quickly, assess the benefit, and plan larger investments based on showed value.
Evaluation Phase: Testing Multiple AI Providers Against Real EA Tasks
An organization is deciding between Claude, GPT-4o, and a locally deployed model for their AI-augmented architecture work. Before committing to a production integration, they configure all three in the AI Modeling Workbench. Architects run real documentation and analysis tasks against the same EA models using each provider. The comparison is based on actual results with actual data, not synthetic benchmarks. The evaluation phase produces a clear recommendation grounded in operational experience.
Hybrid Deployment: Different Providers for Different Task Types
An organization has found through evaluation that GPT-4o performs better for documentation generation tasks, while Claude produces more useful analysis and reasoning output for complex architectural decisions. They configure both in the Workbench and use each for its strongest task type. Architects select the appropriate provider based on what they’re doing. This granularity isn’t available if you’re committed to a single model for all tasks.
Compliance-Driven Model Selection: Configuring Only Approved Providers
An organization’s AI governance policy specifies which AI providers are approved for use with organizational data. The Workbench allows configuration of only approved providers: making it impossible for architects to inadvertently use unapproved models during EA work. The configuration becomes a governance control, not just a preference setting.
On-Premises AI: Connecting to Locally Deployed LLMs
An organization with strict data controls needs AI assistance but cannot use cloud AI APIs. They deploy a local language model (Llama 3, Mistral, or similar) within their private infrastructure. The AI Modeling Workbench can be configured to connect to this local endpoint, providing AI assistance within EA without any data leaving the organization’s environment. This is particularly relevant for defense, classified-adjacent, or highly sensitive commercial environments.
| Role | Primary Benefit | Secondary Benefit |
|---|---|---|
| EA Architect | AI assistance during active modeling without leaving the EA client | Flexibility to use different providers for different tasks |
| Architecture Manager | Realistic evaluation of AI providers before production investment | Governance control over which models architects use |
| IT Executive | Low-commitment entry point for AI-augmented architecture | Evidence base for Connect engagement investment decisions |
| Systems Architect | AI assistance in EA during technical design work | On-premises option for air-gapped environments |
The evaluation use case is the strongest argument for the Workbench.
Before committing to a production AI integration: which involves EA GraphLink deployment, external connectivity configuration, licensing decisions, and potentially a Connect engagement of six figures: it is reasonable to want evidence that the AI integration will work for your organization’s specific architecture tasks, with your specific repository content.
The AI Modeling Workbench provides that evidence at low cost. You configure API access to one or two providers, set up the Workbench, run your real architecture tasks, and observe results with your actual data. This is a more reliable basis for integration decisions than any vendor demonstration.
On-premises AI is otherwise difficult to access.
For organizations that need AI assistance but cannot use cloud APIs, the Workbench’s ability to connect to locally deployed models is the only viable path for in-EA AI assistance. EA GraphLink and external MCP integrations all connect to cloud AI services. The Workbench is the mechanism for connecting to on-premises AI infrastructure.
Lower complexity for EA architect AI use cases.
If the primary use case is helping architects with documentation, description writing, and analysis tasks within EA: not exposing the repository to external stakeholders: the Workbench provides that capability without the infrastructure overhead of a full EA GraphLink and MCP deployment.
Honest recommendation: most organizations get more value from committing to one integration deeply than from spreading across multiple providers.
The AI Modeling Workbench’s multi-provider flexibility is genuinely useful in two scenarios: during evaluation, and when there’s a clear and sustained operational need for different providers for different tasks. Outside those scenarios, multi-model configuration adds complexity without adding proportional value.
In production deployments, architects consistently using two or three different AI providers: switching between them based on task type: often find the cognitive overhead outweighs the quality benefit. The differences between current frontier models for typical architecture tasks are narrower than they were in earlier model generations. The practical case for production multi-model deployment has weakened as model quality has converged.
The stronger production investment is typically: choose the right integration for your ecosystem and use case (Copilot for M365 stakeholders, Agentforce for Salesforce, Kernaro Assist for in-EA architects), deploy it well, and govern it consistently. The Workbench’s flexibility is for evaluation and edge cases: not the backbone of a production AI architecture practice.
Kernaro Assist offers a richer in-EA experience.
If the goal is in-EA AI assistance for architecture architects, Kernaro Assist is the purpose-built option. It provides a richer feature set than the generic LLM configuration the AI Modeling Workbench offers. The Workbench is a configuration mechanism; Kernaro Assist is a developed product with architecture-specific features. If in-EA AI assistance is the primary need and Kernaro Assist is available, start there rather than with the Workbench.
Compatible Sparx EA version. The AI Modeling Workbench is a native EA feature: check your current EA version against Sparx Systems’ documentation for compatibility. If your EA version doesn’t include the Workbench, an EA upgrade is required. Confirm with your Sparx Systems account manager.
AI provider API keys or credentials. For cloud AI providers (OpenAI, Anthropic, Google, Azure OpenAI), you need API credentials from each provider you intend to configure. These are commercial relationships with the respective AI vendors. Sparx Services is not a reseller of AI provider access: you procure directly.
Network access from EA client machines to AI providers. If the EA client machines are in a restricted network environment, outbound access to AI provider API endpoints must be permitted. For cloud providers, this typically means HTTPS outbound access to provider domains. For on-premises LLMs, network routing to the model server must be configured.
No EA GraphLink required: this is explicitly not a prerequisite for the AI Modeling Workbench. This is one of the key differences from every other integration covered in this section.
For the evaluation use case:
The value is measured in avoided cost: a well-structured evaluation phase prevents committing to a production integration that doesn’t deliver expected results. If a Connect engagement costs $80K–$185K, and the evaluation phase catches a misfit before that commitment, the Workbench’s evaluation cost is small relative to the avoided waste.
For production architect use:
(Documentation and annotation tasks per week × time saved per task × architect hourly rate × 52 weeks)
Example:
5 documentation tasks/week × 30 min saved × $130/hour × 50 weeks = $16,250/year per architect
For a team of five architects, this is approximately $80K in recovered capacity annually. Net against API costs and setup time.
For the on-premises AI use case:
The value is risk mitigation and policy compliance: similar to the Azure OpenAI calculation. AI assistance that would otherwise be blocked by policy becomes permissible. Quantify as the value of AI-assisted architecture productivity that would otherwise not exist.
Kernaro Assist
Purpose-built in-EA AI extension from Sparx Systems. Richer feature set than the generic LLM configuration the AI Modeling Workbench provides. Architecture-domain-specific features rather than general-purpose LLM access. If in-EA AI assistance for architects is the primary need, Kernaro Assist is the stronger product. The Workbench is a configuration panel; Kernaro Assist is a developed capability. Note: Kernaro Assist is currently in Beta: confirm availability and feature scope with Sparx Systems.
EA GraphLink + MCP + Single Provider
For organizations where the goal is AI integration for a production use case: not evaluation or on-premises access: the EA GraphLink path with a single provider (Claude, Copilot, Agentforce, etc.) is simpler and deeper. The Workbench’s multi-provider flexibility adds overhead that typically isn’t needed in production. The MCP path exposes the repository outbound to external AI tools, which is a different and complementary capability.
EA GraphLink Foundation
If the real goal is connecting the EA repository to external AI tools and platforms: rather than configuring AI within the EA client itself: the right starting point is EA GraphLink Foundation, not the AI Modeling Workbench. These are different paths for different use cases. EA GraphLink enables the outbound integrations (Copilot, Agentforce, Claude via MCP). The Workbench enables AI within the EA client.
What is the AI Modeling Workbench and how is it different from Kernaro Assist?
The AI Modeling Workbench is a native Sparx EA feature: built into the EA client: that allows configuration of multiple LLM providers for use during modeling work. Kernaro Assist is a separate extension developed by Sparx Systems with architecture-specific AI features, currently in Beta. The Workbench is a general-purpose LLM configuration panel; Kernaro Assist is a purpose-built architecture AI tool. Both operate within the EA client. If Kernaro Assist’s features match your needs, it provides a richer experience than the Workbench’s generic LLM configuration.
Do I need EA GraphLink to use the AI Modeling Workbench?
No. This is one of the key characteristics of the Workbench: it’s a native EA feature that connects the EA client directly to AI providers. EA GraphLink is not required. The Workbench is an option for organizations that want in-EA AI assistance without deploying EA GraphLink infrastructure.
Can I use the AI Modeling Workbench with on-premises LLMs?
Yes. The Workbench can be configured to point to any AI endpoint: including locally deployed language models: as long as the endpoint is accessible from the EA client machine. This makes the Workbench the primary option for organizations with strict data controls that need AI assistance but cannot use cloud AI APIs. Confirm endpoint compatibility with Sparx Systems documentation for your specific EA version.
How does this relate to the MCP server in EA GraphLink?
They operate in opposite directions. EA GraphLink’s MCP server exposes the EA repository outward to external AI tools: Copilot queries EA, Claude queries EA. The AI Modeling Workbench connects AI providers inward to the EA client: the EA client calls an AI provider during modeling work. They are complementary, not competing. An organization can run both simultaneously.
Is the AI Modeling Workbench included with my EA license?
The AI Modeling Workbench is a native EA feature, available in compatible EA versions without additional licensing. API costs for the AI providers you configure are charged by those providers: separate from your EA license. Confirm your EA version compatibility with Sparx Systems.
Should I start with the AI Modeling Workbench or go straight to EA GraphLink + Connect?
If you’re in evaluation mode and want to test AI providers against real architecture tasks before making a production commitment, start with the Workbench. It’s lower cost and lower commitment. If you have a clear integration requirement: specific AI tools, specific stakeholder access needs, specific ecosystem: go directly to EA GraphLink and Connect. The Workbench is for evaluation and in-EA architect use; EA GraphLink + Connect is for production integrations that reach stakeholders beyond the EA client.
Can I configure Azure OpenAI as a provider in the AI Modeling Workbench?
Yes. Azure OpenAI exposes an API endpoint compatible with the same interface as the public OpenAI API. The Workbench can be configured with an Azure OpenAI endpoint, API key, and deployment name. This is the path for organizations that need the compliance characteristics of Azure OpenAI (data within Azure tenant) but want to use it within the EA client rather than through an external MCP connection. Your Azure team provisions the Azure OpenAI resource; you configure its endpoint in the Workbench.
What AI tasks can I use the AI Modeling Workbench for?
The Workbench enables AI assistance for tasks that the EA client can initiate with AI provider context: primarily documentation generation, element description writing, analysis assistance, and modeling support tasks. The specific feature set depends on your EA version. Check current Sparx Systems product documentation for the task types supported in your version, as the feature set is actively developed.
The AI Modeling Workbench is a practical starting point for organizations that want to experience AI-augmented architecture work before committing to a production integration, or that need in-EA AI assistance for architect use cases.
If you’re in evaluation mode and want help structuring a comparison of AI providers for your architecture practice, that’s part of what Discover covers.
If evaluation has clarified the right integration path and you’re ready to move to production deployment:
[Start Your Connect Engagement]
Questions about where the AI Modeling Workbench fits in your broader AI architecture strategy?
[Let’s Discuss Your Options]
Talk to a Sparx Services architect about where your organization is on the journey and what the next stage looks like.