AI-Augmented Architecture: The Definitive Guide for EA Teams in 2026

AI-Augmented Architecture is the way architects should be doing the job on Sparx EA: the architect decides what should be done, and AI does the mechanics in the tool and the model. A team that works this way finishes the model, keeps it current, and can then say quickly what changed and what a proposal would affect. A CIO can ask, in a Teams chat, “Which applications are approaching end of life, and what capabilities do they support?” and get a structured answer drawn from the repository, read-only, through Microsoft 365 Copilot. That answer is the result. The work behind it is a model that someone finished and keeps current, and that is the work AI now does.

This guide covers the work, the skills it asks of architects, the condition the model needs to be in, and which tools help with which job. Most teams are not working this way yet. The technology is no longer what holds them back. What holds them back is whether the architects can state what they want and whether the model is in shape for AI to build on.

Why this is happening now

Three developments converged through 2025 and into 2026 to make AI-Augmented Architecture possible and, for many teams, overdue.

AI that can act in the tool and the model. The Model Context Protocol (MCP), introduced by Anthropic in late 2024 and adopted across the major AI vendors during 2025, gives AI assistants a common way to reach external data and tools. Any source that publishes an MCP server can be used by any MCP-capable client. Sparx EA core does not ship an AI assistant or an MCP server, so the capability comes from products that arrived in 2026. Kernaro AI Hub, from Sparx Systems and released in January 2026, includes a server-deployed, read-only MCP server. AI Power Tools for Sparx EA, which followed in April 2026, is free, published by NovoCircle, and gives Claude and GitHub Copilot full read/write access to the repository, with visual diagram validation through the EA interface.

Models that can follow an instruction through structured data. The current generation of large language models can take a clear instruction and carry out a long sequence of modeling steps against architecture data, then explain what they did. Earlier generations could not do that reliably.

A backlog that hand work cannot clear. Architecture teams are asked for more with the same people. Most produce a good prototype, a pilot for one department, or coverage of the top initiatives, and then run out of time to finish the model for the rest of the organization. Leaders ask what changed and what a proposal would affect, and the team cannot answer quickly because the model is incomplete or out of date.

The technology is no longer the constraint. The constraint is whether architects can state what they want, and whether the model is in shape to build on.

The work AI takes over

A large share of an architect’s week goes to mechanics that require no architectural judgment:

• Drawing and redrawing the diagrams for each application
• Completing the mappings between applications, capabilities, processes and technology
• Updating the model for what the project teams delivered last month
• Checking elements against the modeling standard
• Producing the views, reports and slide decks that stakeholders ask for

These outputs are valuable, and producing them by hand uses the hours that should go to weighing options, managing trade-offs, engaging stakeholders on hard decisions, and building the mental models good architecture depends on. AI can do the mechanics well enough that the architect’s time moves to the decisions. The teams that build the skill now will have the strongest practices a few years out.

Four kinds of work

The work falls into four kinds. In each one the architect states the intent and judges the result, and AI does the work in between.

Architecture Modeling

Today the team builds diagrams and fills in the model by hand, so coverage stops where the time runs out. AI-augmented, the architect states what the model should contain and the standard it follows, and AI creates the elements, draws the diagrams for every application and completes the mappings across the portfolio. A spreadsheet, an inventory or a document becomes properly typed, connected elements in the repository. The skill is deciding the structure and the level of detail: what to model, at what granularity, and what to leave off a diagram.

Architecture Analysis

Once the model is complete and current, it can say what depends on what, what changed since the last review, and what a proposal would affect. AI traces dependencies across the whole repository instead of the few facets a person can hold at once, and compares versions to show what changed, when and by whom. The skill is choosing which questions matter and deciding what the answer means for the business, not only for the graph.

“What breaks if we decommission Application X?” traverses the dependency graph from the application through the capabilities it supports, the processes those capabilities enable, the user communities those processes serve, and the technology it depends on. In a model that is complete and current, that comes back in seconds. In a model that is not, no tool can invent the missing links.

Architecture Governance

AI checks every element against the modeling standard and lists what is missing or inconsistent: elements without required properties, relationships that look absent against the pattern of the rest of the model, classifications that differ from similar elements. Review time then goes to rationale instead of mechanics. The architect decides what the standard is, which rules reflect the real standard and which are habit, and what to do about each finding. Automation checks completeness. People own correctness.

Stakeholder Engagement

Program managers, domain leads and technology owners have recurring questions. Which applications support a given domain? Which projects touch the customer data architecture? What is the technology-risk profile of the retail banking platform? AI produces the dashboards and views that answer them from the model as it stands today, in the stakeholder’s vocabulary. The architect decides what each audience needs to see and in which words, and says how complete the underlying data is. Stakeholders get answers without a queue, and the architects get their time back for the questions that need an architect.

The skill is stating intent and judging the result

AI is only as good as the person directing it. Take the diagrams people use to find their way around a model. Today an architect draws navigation diagrams by hand, and they go stale as the model changes. The AI-augmented way starts with the people who will use them. The architect decides what experience a stakeholder should have when they open the model, where they should land and how they should drill down, and states that intent. AI builds the navigation diagrams and keeps them current. The skill is designing the experience and stating the intent clearly. The drawing is no longer the architect’s job.

Precision matters in the instruction. “Applications with lifecycle status ‘Strategic’ that support ‘Customer Data Management’ through an ArchiMate Serving relationship” gets a better result than “what supports onboarding?” every time. Language models are probabilistic, and a vague instruction against a clean repository can still return a confident answer that is partly wrong.

Judging the result is the other half. The architect reviews what AI built or found, checks it against the intent and the facts, and decides whether it stands. A change that reaches the model has been looked at by a person who can be held responsible for it.

What your architects need to be able to do

Readiness is mostly about people. Four abilities decide how much of the work a team can hand to AI.

State the intent. Describe the outcome in terms specific enough to act on: the experience a stakeholder should have, the standard a model must meet, the question that needs answering.

Judge the result. Review what AI produced, recognize when it is wrong, and trace the problem back to the instruction or to the data.

Know the architecture. AI amplifies expertise and does not supply it. Deciding that an unclassified system should be Business-Critical takes someone who knows the system.

Keep the model current as a habit. When a project delivers, the update goes into the model through the same process every time, with AI doing the mechanics and an architect reviewing.

These skills are taught on your own repository and built through mentoring until they stick. The architects track covers how, and the readiness assessment scores where a team stands today.

The condition of the model

The second half of readiness is the model. AI builds on what the repository contains, and it cannot invent structure that was never modeled. A question like “Which applications support our customer onboarding capability?” depends on three things being true:

1. Applications modeled as elements with a governed «ApplicationComponent» stereotype
2. Capabilities modeled with a governed «BusinessCapability» stereotype
3. An explicit relationship between the applications and the capability, of the type your MDG maps as application-supports-capability

Miss any one and the answer comes back incomplete. Miss all three and it comes back empty. The model has to be fixed, and the AI is the means of fixing it quickly.

AI can build on a model that is in reasonable shape and finish it. It cannot make up structure that was never modeled.

What a model in good shape looks like

Consistent stereotypes. Every significant element carries a stereotype from a governed MDG Technology profile. No significant element sits on a bare UML type.

Populated key properties. Owner, lifecycle status, maturity level and investment direction are filled in on strategic elements consistently, not sporadically.

Explicit cross-layer relationships. The links that power cross-layer analysis, application-supports-capability and technology-hosts-application, are modeled with the relationship types your MDG recognizes.

Named and described elements. Names mean something to domain stakeholders as well as to their author, and short descriptions give AI the context to say what each thing is.

A model that does not meet this bar is the starting point for the work and does not block it. We can build and update the model for you, including the MDG and rulesets that keep it consistent, and we can help you decide where to start with a plan your management will fund.

Which tools help with which job

Several tools run on the same platform. Which ones you use depends on the job and on what you already license. AI tools for Sparx EA sets them side by side, and the AI Power Tools documentation (opens in a new tab) has the feature detail.

Building and updating the model

Claude and GitHub Copilot with AI Power Tools work on the live repository from the architect’s machine, with read/write access. For Claude, the Claude Desktop extension is the MCP server and the plugin supplies the skills; both are needed. GitHub Copilot gets the server through a VS Code extension, and VS Code loads the same skills. This is the route for building diagrams, completing mappings and applying project updates.

Kernaro Assist is an add-in inside Enterprise Architect for natural-language modeling, analysis and validation, for architects who want the assistance in the tool they already have open.

Prolaborate 6 with AI generates descriptions, summarizes discussions and diagrams, rephrases, and answers questions about the model. It is inactive by default and needs a Kernaro AI Hub license and the customer’s LLM.

Answering questions across the organization

Kernaro AI Hub is server-deployed and its MCP server is read-only. It relies on an MDG Technology defined for your repository that maps the physical Sparx schema onto a clean GraphQL schema. It serves Model Chat, impact analysis, cross-tool answers from Jira, Confluence, Azure DevOps and SharePoint, and a report builder, in plain language for analysts and stakeholders. Because it is read-only, it answers questions and does not change the model. Shared, server-deployed access builds on Pro Cloud Server. Other MCP-capable assistants, including Salesforce Agentforce, can query the same read-only server, and Tableau can be fed from the same repository.

Dashboards, Copilot and your data platform

Once the model is finished and current, the people who never open EA can use it. We can help you get your EA data into Power BI, Microsoft 365 Copilot, Microsoft Fabric or your data warehouse. When you have Kernaro AI Hub, the Kernaro Connector for Power BI is our preferred route. Without Hub, AI Power Tools generates a Power BI model from your EA data. Microsoft 365 Copilot answers read-only from a published Power BI semantic model today and, with Kernaro AI Hub, through Hub’s read-only API. A reporting database can be mirrored into Fabric, or used as the feed to your warehouse. Reporting and Power BI covers the routes.

Microsoft 365 Copilot and Fabric work inside the organization’s Microsoft 365 tenant, which suits regulated industries with data-residency requirements; confirm the specifics with your security team.

Choosing without locking in

Because Kernaro is bring-your-own LLM, Kernaro AI Hub and Kernaro Assist run on whichever provider you choose: Azure OpenAI, OpenAI, Claude or an enterprise endpoint. As models improve, you change the provider and keep the platform. Organizations that want the strongest tool in each category regardless of platform, or that run several clients for different teams, can combine the routes above, since each one reads and writes the same Sparx EA repository.

What it does not do

Most disappointment with these tools traces back to expecting one of the following.

It does not replace architectural judgment

AI can tell you which applications are at end of life. It cannot tell you whether the case for replacing them outweighs the risk of the replacement program. That is judgment drawn from domain experience and organizational context. AI-Augmented Architecture amplifies architects, and treating it as a substitute produces worse architecture.

It does not compensate for a poorly structured model

No tool extracts reliable results from a model with untyped elements, sparse properties and missing relationships. Start with the structure, then let AI do the volume. Buying the tool before doing the structural work is the most common way to be let down by it.

It does not keep the model current without a habit

AI makes updates cheap, and someone still has to ask for them. If models are built during projects and then left to drift, the answers will describe an organization that no longer exists, and stakeholders who get a few stale answers stop trusting the model and then the practice behind it. Keeping the model current takes a commitment to send every project delivery through the same update process.

It does not turn vague instructions into good results

Treat what AI produces as a draft that an architect checks. A precise instruction gets a precise result, and a vague one gets a confident guess.

How to get started

The path runs in sequence: decide where to start, get the model and the team ready, set up the tools and do a first job, then extend. Skipping ahead is where most teams stumble.

Stage 1: Decide where to start

A Paralysis to a Plan engagement is the right opening move. It produces an assessment of the repository’s governance quality and the team’s skills, a capability map and application landscape as a baseline, a gap analysis, and a prioritized roadmap with a plan your management will fund. It answers one question: where do we start, and what does it take?

Stage 2: Get the model and the team ready

Where the assessment finds gaps in the model, we can close them, building the diagrams, the mappings and the governed profiles in your repository. At the same time we teach your architects to state intent and judge results on that repository. The work needs only the elements and relationships the first job will use.

Stage 3: Set up the tools and do a first job

Choose one job with a visible result, for example the navigable diagrams for one domain, or the mappings for one part of the portfolio. Configure the Solution sets up the AI tools on your Sparx EA and gets the data ready to work with. AI Power Tools for Sparx EA is free to try on an architect’s own machine. The other tools are licensed by their vendors.

Stage 4: Extend and keep it current

Once the first job is done, extend to the next domain, send each project delivery through the update process, and add the dashboards stakeholders ask for. A well-run program always has a next increment worth doing.

Frequently asked questions

What is AI-Augmented Architecture?

AI-Augmented Architecture is the way architects should be doing the job on Sparx EA: the architect decides what should be done, and AI does the mechanics in the tool and the model. The repository stays in Sparx EA. With the model finished and kept current, dashboards and the AI assistants your teams already use can answer questions from it as a result.

What do architects need to be able to do?

State the intent clearly, judge what AI produces, and know the architecture well enough to make the call. These are skills, and they are taught on your own repository. Setting up a tool takes hours. Building the skill to direct it takes practice.

Which tool should we use?

It depends on the job. Claude or GitHub Copilot with AI Power Tools for Sparx EA can build and update the model. Kernaro Assist works inside Enterprise Architect. Kernaro AI Hub is a web layer for Model Chat, impact analysis and report building, with a read-only MCP server. Power BI and Microsoft 365 Copilot serve stakeholders from the finished model. AI tools for Sparx EA compares them.

Does Sparx EA have a built-in MCP server?

No. Sparx EA core ships no MCP server. Two exist alongside it: the read-only MCP server included in Kernaro AI Hub, server-deployed for enterprise-wide access; and AI Power Tools for Sparx EA, a free local MCP server from NovoCircle with full read/write and diagram validation through the EA user interface, which works with Claude (with a skills library) and with GitHub Copilot (the MCP server only, through a VS Code extension).

What is MCP, and why does it matter for EA?

MCP, the Model Context Protocol, introduced by Anthropic and widely adopted across AI vendors, is a standard way for assistants to reach external data sources and tools. Before MCP, each AI-to-data integration was bespoke. With it, any source that publishes an MCP server (including the one in Kernaro AI Hub) is reachable by any MCP-capable client, so the Sparx EA repository can be reached from assistants such as Claude and GitHub Copilot through one mechanism.

What is Kernaro AI Hub’s MCP server, and who makes it?

Kernaro AI Hub is a Sparx Systems product released in January 2026. It includes a server-deployed, read-only MCP server that AI assistants use to query the repository, with a GraphQL layer behind it. That layer relies on an MDG Technology that maps the physical Sparx schema onto a clean GraphQL schema. For Power BI and, read-only, Microsoft 365 Copilot, the Kernaro Connector for Power BI, an extension to Kernaro AI Hub, is our preferred route.

Does the model have to be in good shape first?

It has to be in reasonable shape. AI builds on what the repository contains. If elements lack stereotypes, properties are sparse or relationships are missing, the results come back incomplete or misleading. The model does not have to be finished, because finishing it is the first job. Governing the model through MDG Technology is the quality gate, and building the model includes establishing or validating it.

What is Kernaro, and how does it relate to Sparx EA?

Kernaro is the AI family from Sparx Systems for Sparx EA. Kernaro Assist is an add-in inside Enterprise Architect for natural-language modeling, analysis and validation. Kernaro AI Hub is a web layer over the repository for Model Chat, impact analysis, integrations and report building, and it includes the read-only MCP server. Kernaro is bring-your-own LLM: your organization supplies the keys and pays the provider for usage. These are specialist tools that sit beside general-purpose assistants like Claude or Copilot rather than replacing them.

Can we use Azure OpenAI instead of Microsoft Copilot?

Yes. Kernaro is bring-your-own LLM, so your organization can supply keys for Azure OpenAI, OpenAI, Anthropic Claude or an enterprise endpoint and pay that provider for usage. That helps organizations that want Azure-native data residency or a specific model. We help you choose and set up the combination in Configure the Solution.

What does it cost to implement?

The Sparx Services engagement is scoped to your environment: the condition of the repository, how much of the model needs building, and how many dashboards you want all factor in, so we work it out together rather than quoting a list price. That scope is separate from third-party product licensing, which you arrange directly: Power BI, Microsoft 365 Copilot and Fabric from Microsoft, Kernaro AI Hub and Prolaborate from Sparx Systems, and Claude or GitHub Copilot from their vendors. AI Power Tools for Sparx EA is free. Talk to us about scope.

Is it suitable for regulated industries?

It can be, and the Microsoft route is the usual starting point. Microsoft 365 Copilot and Fabric work inside the organization’s Microsoft 365 tenant, which is how many financial services, healthcare and government organizations meet data-residency requirements; confirm the specifics with your security team. Because Kernaro is bring-your-own LLM, you also choose the model provider. For UK government and US defense, Government Cloud (GCC and GCC High) tenants provide the appropriate classifications. Sparx Services has delivered into financial services, government, and defense contexts.

Where to start

AI-Augmented Architecture is not a research project, and you cannot buy it off a shelf and switch it on. It is a way of working that architects learn, applied to a model that gets finished and kept current. We can teach your architects to work this way on your own repository, or build the model for you and hand it over. If you also need the finished model in front of stakeholders, we can get your EA data into Power BI, Microsoft 365 Copilot, Fabric or your data warehouse. For the full definition, see What Is AI-Augmented Architecture?, and for the wider picture of how the work divides, see AI-Augmented Architecture.