7.4 Proposing Microsoft Power Platform AI Features & the AI Hub

Microsoft Power Platform AI Features & the AI Hub

Quick Answer: The AI hub is the consolidated AI surface inside the Power Apps and Power Automate maker portals, reached from the left navigation through ... More > AI hub, where makers create and manage AI models, prompts, and agents, and where the Monitor activity area exposes AI Builder activity for consumption and governance. Underneath it sits AI Builder, which supplies two classes of model: prebuilt models that are ready to use for scenarios common across businesses (business card reader, text recognition, receipt processing, invoice processing, ID reader, contract processing, key phrase extraction, language detection, sentiment analysis, text translation, text generation), and custom models that the maker builds, trains, and publishes on business-specific data (document processing, prediction, object detection, plus customisable category classification and entity extraction). Activity data is written to the AI Event Dataverse table (msdyn_aievent), which persists even after the model, flow, or app is deleted — making it both a governance asset and a data-retention obligation. Do not confuse this with Microsoft Purview AI Hub, which is a security and compliance surface for AI risk, not a Power Platform maker surface.

Architects arriving from Azure frequently under-use the Power Platform AI layer. They reach for a Foundry model deployment and a custom orchestration when the requirement was "read 4,000 supplier invoices a month and populate a Dataverse table" — a scenario the prebuilt invoice-processing model resolves with no model training, no endpoint, and no inference infrastructure to operate.

Proposing Microsoft Power Platform AI features, including the AI hub, is an explicit design-domain skill on the AB-100 blueprint. It is the low-code end of the build/buy/extend spectrum, and on the exam it is the correct answer far more often than its reputation suggests.

---

1. Navigating to the AI Hub

The AI hub is not a separate product or portal. It is a navigation node inside the maker experience:

1. Sign in to Power Apps or Power Automate.
2. On the left pane, select ... More > AI hub.
3. Select AI models to browse and create models. (Makers can pin AI models to keep it permanently on the menu.)

From the same hub, makers reach prompts and agents, and the Monitor activity tile leads to AI Builder consumption reporting. In Power Automate, the AI Builder activity page is surfaced under Automation Center > Monitor > AI Builder activity.

The architectural significance of a single hub is consolidation: models, prompts, agents, and their consumption telemetry share one governance surface, one Dataverse environment boundary, and one security model, instead of being scattered across per-feature configuration pages.

---

2. The AI Builder Model Catalogue

AI Builder classifies every model by data type (documents, text, structured data, images) and build type (prebuilt or custom). Prebuilt models go straight to productivity; custom models are built, trained, and published by the maker for data unique to the business.

Microsoft maps these to canonical business scenarios: receipt processing automates expense reports; contract processing extracts key clauses and data points; document processing automates customer application intake; prediction identifies fraudulent transactions; object detection automates inventory taking; category classification organises user feedback; sentiment analysis classifies customer feedback; key phrase extraction surfaces brand mentions.

The decision rule is data uniqueness. Use a prebuilt model when the document or text pattern is common across businesses — an invoice looks like an invoice everywhere. Use a custom model when the pattern is proprietary: your own claim form layout, your own product images, your own churn signal.

---

3. Prompts as a First-Class Power Platform Artefact

An AI prompt is a natural-language instruction backed by a large language model, created in the AI hub and consumed from Power Apps, Power Automate cloud flows, Copilot Studio agent flows, and Copilot Studio agents. Prompts accept typed inputs, return structured outputs, and are versioned as Dataverse artefacts inside a solution — which means they participate in normal application lifecycle management rather than living as untracked strings inside a flow action.

Two architectural consequences matter for AB-100:

• Reuse across surfaces. A single governed "summarise this service case in three bullets for a field technician" prompt can be invoked from a canvas app, a nightly flow, and an agent. Duplicating that instruction into three places is the anti-pattern the prompt artefact exists to prevent.
• Model choice inside a low-code surface. Foundry models for prompts are available inside Copilot Studio, Power Apps, and Power Automate, so a prompt is not permanently locked to a default model; model selection becomes a configuration decision subject to the organisation's permitted-model rule.

---

4. Monitoring, Consumption & Governance

The Monitor activity area of the AI hub provides tables for AI model and prompt activity, the data they process, and consumption tracking. The AI Builder activity page aggregates activity generated in Power Automate, Power Apps, and Microsoft Copilot Studio, defaulting to all models over the last seven days, with columns for:

Selecting an event reveals Processed by (typically the flow owner or the app user) and Model (the language model used for a prompt run).

Visibility is role-scoped. System Administrator and System Customizer see all activity for all AI models; Basic User and Environment Maker see only their own activity for models they can access. An architect designing a governance dashboard must account for this: a maker cannot self-serve tenant-wide consumption reporting.

4.1 The AI Event Table Is a Retention Obligation

Monitoring data is stored in the AI Event Dataverse table (msdyn_aievent), and Microsoft is explicit that it persists even if you delete the model, cloud flow, and app. Recording activity requires Create and Write access to that table. For text scenarios the table contains the input of the AI model predict actions, which means prompt inputs may constitute personal or sensitive data under the organisation's classification scheme.

The correct architectural response is a deliberate retention policy implemented as a recurring bulk-delete job in Dataverse:

1. Sign in to the Power Platform admin center.
2. Select Environments, choose the environment, then Settings.
3. Select Data management > Bulk deletion > New.
4. In Look for, select the AI Events table.
5. Add a filter, typically using OlderThanXDays, to match records beyond the retention window.
6. Name the job, choose an off-peak start time, and select Run this job after every n days to make it recurrent.

Bulk deletion requires the Bulk Delete privilege and is irreversible, so the filter must be tested before submission. The table can also be exported to CSV for audit evidence before purging.

4.2 Credits

Building and testing models and prompts does not consume AI Builder credits or Copilot credits. Running them in agents, agent flows, apps, or flows does. This asymmetry is the reason pilot programmes look free and production rollouts do not; the cost model must be sized on projected runs, not on the number of artefacts built.

---

5. Where Power Platform AI Fits in the Architecture

AB-100 Exam Tip: When a scenario describes bounded, high-volume, non-conversational intelligence inside a business process — read the document, extract the fields, classify the text, predict the outcome — the intended answer is an AI Builder model surfaced through the AI hub, not a Foundry deployment and not a Copilot Studio agent. Copilot Studio earns its place when the interaction is conversational or agentic; Foundry earns its place when the model itself must be custom or frontier-class. Also watch for the naming trap: Microsoft Purview AI Hub is a compliance and AI-risk surface and is never the answer to a Power Platform maker-tooling question.