Introducing Ardoq Foundation Insights Agents

Ardoq Insights Agents are AI Agents that explore your architecture repository and look for a variety of architectural insights. You can think of them as architectural assistants, expanding your ability to track all the details in your repository. They will look for potential data quality issues, such as gaps or inconsistencies in your model, and can also find anomalies or insights that indicate architectural stories that could benefit from deeper investigation.

A key benefit of these Agents is their ability to look through large volumes of data and pick out the most valuable insights, giving you and your team more time to focus on taking action and delivering value to your stakeholders. We anticipate future versions being able to undertake follow-up action subject to user approval.

In this beta release, we are introducing initial versions of two complementary types of Insights Agent:

For this beta release, the two types of Insights Agents work independently, and both are limited to the metamodel that underpins Ardoq Foundation and models that conform to it. See Future Evolution of Insights Agents at the end of this article for details of how we intend these AI capabilities to grow.

To use the beta release of Foundation Insights Agents your Ardoq organization must have the FND - Foundation Solution deployed. For new Ardoq customers, this is the Solution that is already deployed when you start working with Ardoq. For older Ardoq customers who do not have the Foundation Solution deployed, you need to follow these instructions to add the Foundation Solution to your Ardoq organization.

This beta release requires that the Foundation Metamodel remains unaltered. This means that the following must not have been changed:

Future releases will provide more robust support for adjusted metamodels.

To use the Model Insights Agents you must have access to the Report AI Assistant. To check this, open any report and check that you see the following button in the top right corner:

If you don’t see this button, you should first ensure that you have AI Features enabled. If you don’t, see How do I enable AI features in my Ardoq instance? for details. Once AI Features are enabled, you can turn on the Report AI Assistant in Organization Settings → Features (you will need to be an Admin user). After this, the AI Assistant button will be available in all your reports.

This beta release is restricted to Insights Agents that work with the metamodel and selected reports that belong to the Foundation solution.

To have access to the reports and dashboard created by the Foundation Metamodel Insights Agent you need to request for these to be deployed into your Ardoq organization. These will soon be available by contacting Technical Support or your Customer Success Manager to arrange for their deployment. Successful deployment will result in new assets being added to your existing asset folders:

As a reminder, in this beta release the Metamodel Insights Agent itself is cannot be configured or run by end users. What is seen is the output of the agent that is, for now, managed and run by Ardoq internally: a collection of AIA-labeled Reports and an AIA-labeled Dashboard. Note that the AIA Reports contain gremlin code generated by the Metamodel Insights Agent. You should not attempt to modify this code, as it cannot be restored without redeploying the solution.

If you have changed the workspace names where Foundation components are held, you may update the list of a Report’s source workspaces by editing the report setup to ensure that the output contains your live data.

The Foundation Insights - AIA Dashboard contains a widget for each pattern identified by the agent following its analysis of the metamodel. This is a large set of possible insights. It is likely that some of these are not relevant for your organization; for example, there may be some component fields in the metamodel which you have decided not to populate for the time being. You can edit the Dashboard to rearrange the widgets, deleting those that are not relevant, changing the coloring or type of widget to suit your needs. In this way, you are tailoring your use of the agent’s insights. You should not amend the gremlin code of the underlying reports.

You do not need to take any action to gain access to the Model Insights Agents provided you have the Foundation solution deployed and the Report AI Assistant enabled (as described in Prerequisites above). In this initial release, the Model Insights Agents are automatically embedded within the Report AI Assistant associated with selected Foundation reports. In the list of Reports, you can identify those that contain Model Insights Agents by the agent icon that appears after the report name:

Note: in this first beta release, renaming the Report will cause the loss of this identifying icon.

For this beta release, the Metamodel Insights Agent is only represented by its output, which takes the form of a collection of generated Reports and an associated Dashboard. We recommend accessing these via the Dashboard Foundation Insights - AIA. This is divided into the following sections, reflecting the different underlying AI-generated Reports:

Each section contains a collection of Pie Chart Widgets, each of which focuses on a pattern that might be worthy of deeper investigation. These patterns are ones identified by the Foundation Metamodel Insights Agent as being of potential architectural significance. Each one corresponds to a column in a report that was generated by the agent, containing the value True if a given component has data that matches the pattern.

The True slice is coloured red, and represents the proportion of components that match the pattern (i.e. those that require investigation). As with all dashboard widgets, clicking on a pie segment opens the widget’s underlying report filtered to show just those components that are represented in the selected segment. Clicking on the True segment of the pie titled Unowned Critical Applications in the above example opens this filtered view of the AI generated report Application ownership gaps - AIA:

The user can now see the full list of all unowned applications that have a high score (4 or 5) for business criticality. And note that the AI Assistant button is available in the top right-hand corner should the user wish to explore the content of the report in more detail conversationally.

The Foundation Metamodel Insights Agent has identified a large number of patterns: 30 in all, across 7 different AI-generated Reports.

Not all of these will be interesting to all customers. For example, if you have not made use of the Reports To reference that allows you to show who reports to whom, four of the Dashboard widgets, under the heading People Structure, will be of no interest to you. We’ve chosen to leave all of the insights identified by the Agent in the Dashboard rather than impose our own preferences and limit your choice. The consequence of that decision is a large Dashboard with widgets that will vary in interest depending on how much of the Foundation metamodel you are using, and which of the patterns might be a useful “red flag” for you, given your organizational context.

In the future, the Agent will become part of the Ardoq platform, at which point we anticipate giving you the opportunity to configure its output. For this release, we recommend two ways that you can safely change the output of the Agent (the generated Reports and their accompanying Dashboard):

You should not, however, amend the Report’s agent-generated gremlin code.

Model Insights Agents are accessed via the AI Assistant attached to individual Reports. For this beta release, Model Insights Agents are only available for the following reports:

Agents linked to further Foundation reports are under development and will be released in the coming weeks.

When triggered, each Model Insights Agent examines the data that lies behind the report for insights that can be gained by looking at the report for evidence of identified patterns. These patterns may variously correspond to gaps or errors in the data, or anomalies that are worthy of further investigation. These might also turn out to be data errors, but they might simply be anomalies that can be explained with your knowledge about the distinctive nature of your organization. Or they may be indicative of an issue in your organization, such as conflicting data from different sources, that represents an opportunity for the architect to add value - bringing together stakeholders to build better understanding leading to better business decisions. For example, an application may be considered to be of the highest level of business criticality by its business owner, yet it receives only a low level of support from the IT Service team. Some of these patterns reveal opportunities for new architectural initiatives - potential stories worth developing into action that can lead to tangible business value.

Taken together, these aspects that are investigated by the Model Insights Agent are often referred to by architects as “architectural smells” (Garcia et al. 2009). The value of the Insights Agents is that they have the potential to spot valuable insights among a sea of report data that would likely be missed by a human expert architect.

Each of the Agents is accessed and triggered in the same way. Open the report that contains a Model Insights Agent and open the AI Assistant by clicking on its button in the top right corner of the Report:

The AI Assistant panel will open on the right-hand side of the Report. Towards the bottom of this panel, just above the prompt input field, you will see a button inviting you to run the Insights Agent.

Click on this button to invoke the Insights Agent. It will take a minute or two to fetch the data associated with the Report, perform some calculations then analyze it from a number of different perspectives before populating the upper portion of the AI Assistant window with the output. At the end of the output you can ask to see the rationale behind the agent’s recommendations.

The insights remain in the agent context, so you can continue to discuss the results and recommendations with the AI Assistant.

Like all chats conducted with the AI Assistant, you can recall past conversations, including invocations of the Insights Agent by clicking on the history icon at the top of the AI Assistant window:

This AI agent examines a metamodel and identifies a set of metamodel policies for healthy EA data that can be monitored with rules. For this beta release, the agent visible to users of the Ardoq platform. Instead, it has been run independently with the Foundation metamodel, and its output is being made available via a set of AIA Reports that encode the identified rules, and an AIA Dashboard from which their application to your data can be visualized.

To analyse the metamodel, the AI Agent needs to be given a definition of the metamodel with all its richness of expression. We know that using a simple description, as one does in a text prompt to a Large Language Model, results in knowledge getting “lost in translation”. Parts of the metamodel get misunderstood, while other parts are invented by it in hallucination, resulting in inaccurate and inconsistent recommendations. To overcome this, we use an ontological representation providing far more information density about the metamodel. Combined with the EA reasoning rubrics, the AI Agent is able to produce much higher quality policies and rules than we’ve seen using manual development.

The result is a set of recommended patterns that are genuine insights, that are then translated automatically into a set of reports ready for consumption and visualization by Ardoq users.

Each Model Insights Agent analyses a Report’s underlying data from the perspectives of multiple policies. Below is a summary, by Report, of the policies considered, and the corresponding questions addressed by the Report’s agent.

As with the Metamodel Insights Agent, we’ve developed unique ways to give the Model Insights Agents detailed context to be considered alongside the data that lies behind the report. This gives the AI agent detailed, relevant and accurate information, pertinent to each policy, to work on.

Something similar to these Agents can indeed be accomplished with suitable prompts in an MCP client (such as Microsoft Copilot) attached to Ardoq’s MCP Server. Depending on the prompts used, they are likely to give less accurate and consistent responses, and will likely degrade or fail with larger reports. This is where the design of Ardoq’s Insights Agents has a number distinct advantages:

This beta release of Foundation Insights Agents is just the first fruits of our research and development of insights agents. We anticipate further evolution of Insights Agents along the following lines:

This initial release has focused on the metamodel that lies behind the Foundation solution, and its resulting populated model. Future versions of both Metamodel and Model Insights Agents will be developed to support analysis of the metamodels and populated models from other Ardoq Solutions (such as Application Rationalization) and potentially metamodels that have been extended beyond them.

With this beta release of Ardoq Insights Agents, we are keen to learn from the experiences and ideas of early adopters. If you have suggestions for future iterations of Ardoq Insights Agents, please visit Ardoq AI Labs and submit your ideas.