Putting AI to work on your model: what to watch for in the first months

The job here is getting an architecture team to build and maintain its Sparx EA model with AI. Several tools can do it. This piece uses Kernaro Assist as the worked example, because it has the most to settle before the first prompt: a license tier, a model supplier and a guardrail. Kernaro Assist shipped in May 2026. Nobody, including us, has a long track record with it, and this is not a report of results. It is a planning piece, drawn from Sparx’s published material and from how we teach AI-Augmented Architecture. Sparx Services is independent of Sparx Systems Pty Ltd. We teach the tools and do not sell them.

If your architects chose Claude or GitHub Copilot with the free AI Power Tools for Sparx EA, checks 3 to 5 and the measures below apply unchanged. Checks 1 and 2 become different questions, because the tooling is free and the model comes with the Claude or GitHub Copilot subscription your organization already holds.

What Kernaro Assist is, in one paragraph

Kernaro Assist is an add-in inside Enterprise Architect. It brings natural-language modeling, Model Chat, model analysis, validation and event-driven agents to the architect’s desktop, and it works with EA structures and your MDGs. It is one of several ways to add AI to Sparx EA, and our comparison of the AI options places it next to the others. This piece covers what to check when you adopt it.

Check 1: which license you are on

Read Sparx’s license table before you plan anything else, because it decides what your pilot can show. As we read it, Standard uses machine-based activation with no central administration. Under Standard, shared and cloud repositories are read-only, agents are unavailable and there is no central audit trail. Agents come with the Advanced license, and SSO, roles, named seats, shared agents and the audit trail come when Advanced is connected to Kernaro AI Hub.

Plan the pilot on a repository the license lets you write to, and say in advance which parts of the evaluation depend on the Enterprise tier. A team that pilots agents under Standard will conclude, wrongly, that the capability is missing.

Check 2: your LLM and who pays for it

Kernaro Assist does not bundle a model. The customer supplies keys for Azure OpenAI, OpenAI, Anthropic Claude or an enterprise endpoint, and pays the provider for usage. That moves some questions to your side of the table:

• Which provider does your organization already approve, and under what data terms?
• Who holds the keys, and who sees the usage bill?
• What repository content is allowed to leave your network when a prompt runs?

Different models may draft differently. Record which one you used for each trial so a change in output is traceable.

Check 3: how people instruct it, and what review means

Natural-language modeling and Model Chat let an architect describe what they want and get proposed content in EA. The proposal is a draft, and its quality starts with the instruction. Today an architect clicks through the UI and fills in fields. AI-augmented, the architect states the intent, and the skill is stating it precisely: which type, which package, which stereotype, which naming pattern and what to flag instead of guess. Teach this on your own repository, with your own conventions, before you let people loose.

Then decide what a reviewer checks on each kind of content.

• Names and descriptions: do they follow your naming pattern and your description standard, or do near-duplicate labels creep in for the same concept?
• Relationships: are the type, direction, cardinality and stereotype valid under your metamodel? A plausible sentence such as “System A integrates with System B” can still be modeled wrongly.
• Placement: did the element land in the right package, with the right properties?

MDG conventions are what make an AI draft checkable. Without them, the reviewer is comparing the draft to nothing.

Relationship metadata deserves the most attention, because a wrong relationship corrupts every query and diagram built on top of it. Tighten the MDG, meaning which relationship types are valid and which stereotypes are mandatory, before you let anyone approve AI-proposed relationships at speed.

Check 4: validation and event-driven agents

Validation is where Kernaro Assist can take over tedious audit work: checking that required elements and metadata are present. Event-driven agents run on Broadcast Events, so a rule can fire when something changes in the repository instead of at an annual review. Agents need the Advanced license.

The discipline is in how you write the rules:

• State the rule precisely first. “Every application has an owner” leaves open vendor-owned, retired and still-planned applications. Decide the exceptions before you encode them.
• Start narrow. One or two rules you fully understand beat a dozen you cannot explain.
• Watch the noise. A rule that fires on things nobody cares about will be switched off. Keep a short log of each rule, who owns it and when it was last reviewed.

Automation checks completeness. Whether the owner named in the field is the right owner is a human judgment, and it stays one.

Check 5: which jobs belong to other helpers

Kernaro Assist works on the model inside EA. Work across other tools sits with Kernaro AI Hub, the web layer that connects to Jira, Confluence, Azure DevOps and SharePoint and serves analysts and stakeholders. If your first use case is cross-tool analysis or stakeholder self-service, evaluate Hub, or a combination, instead of judging Assist on a job it is not built for. If the first use case is a large piece of modeling work stated in a chat, such as diagrams for every application, Claude or GitHub Copilot with AI Power Tools for Sparx EA is the helper to test alongside it.

Measure before you start

These are the four things we would settle with a team before the first prompt. None needs a number from us. The numbers should be yours, taken from your own repository.

Questions we would keep open

We would rather name these than pretend they are answered.

How much review is enough as volume grows?

When an assistant can draft content quickly, a policy of reviewing every entry may not hold. Moving toward exception review and periodic spot checks is a governance decision. Make it explicitly, with someone accountable, instead of letting it happen by default.

How do you train people to review, not just author?

Approving a draft is a different skill from writing one. Plan to teach reviewers what a plausible-but-wrong relationship or description looks like in your metamodel.

Is the model consistent, or just internally tidy?

Content can satisfy every rule and still emphasize the wrong aspects of the architecture. Completeness checks will not catch that. Periodic review by an architect who knows the domain will.

What this means for your practice

Treat the first months as a structured trial: a known license tier, a named LLM, a hardened MDG, people taught to instruct and review, a few precise rules and measures agreed in advance. Then you will have your own evidence in place of anyone’s anecdote, ours included. For the wider picture of how this reshapes the role, see AI-Augmented Architecture.

Configure the Solution is the engagement where we select the combination of tools, connect it to your Sparx EA and prove it on your data. We can also teach your architects to work this way on your own repository, or build and update the model for you and hand it over.