Build a Power Automate Flow That Uses Azure AI Vision or Custom Vision

Image classification is no longer a niche AI project. With Microsoft Power Automate, Azure AI services, SharePoint, Teams, and optional AI Builder models, organizations can route images for analysis, approvals, tagging, and review without building a full custom application.

From Microsoft Flow to Power Automate

Microsoft Flow is now Microsoft Power Automate, part of Microsoft Power Platform. The core idea from the original article is still valuable: trigger a workflow, submit an image for analysis, evaluate the confidence score, then take action. What has changed is the product landscape. Azure Cognitive Services is now positioned as Azure AI services, image analysis is commonly handled through Azure AI Vision, custom classification can still be built with Azure AI Custom Vision where appropriate, and low-code makers may also use AI Builder inside Power Platform.

Choose the right image AI option

For a modern implementation, start by choosing the right service. Use Azure AI Vision Image Analysis when you need general image understanding such as tags, captions, object detection, image metadata, or visual features through an Azure-hosted API. Use Azure AI Custom Vision when you need to train a model to recognize your own categories, products, defects, equipment types, labels, or field-inspection conditions. Use AI Builder image classification when your team wants a more Power Platform-native, low-code experience and the scenario fits AI Builder capabilities, licensing, and data requirements. Azure AI Foundry can also be used as a current Microsoft entry point for discovering, managing, and building with Azure AI capabilities.

A modern workflow pattern

A typical 2026 workflow starts with a durable business trigger: a file added to a SharePoint document library, an image uploaded through Power Apps, an attachment received in a shared mailbox, a Microsoft Forms submission, a Teams-based request, or a mobile Power Automate button. The flow retrieves the image content, calls Azure AI Vision, Azure AI Custom Vision, or an AI Builder model, parses the prediction result and confidence score, applies business rules, stores the image and metadata in SharePoint or Dataverse, and notifies the right people in Microsoft Teams, Power Automate mobile notifications, or the Approvals app. Social media automation is still possible in some scenarios, but X, LinkedIn, and Instagram connector/API access, licensing, throttling, and terms have changed significantly, so business-owned repositories are usually more reliable triggers.

Example: classify images uploaded to SharePoint

One practical design is to monitor a SharePoint Online library for new images. When a file is created, Power Automate gets the file content and sends it to the selected AI service. The response may include tags, classes, objects, and confidence values. If the confidence is above a defined threshold, the flow updates SharePoint columns such as Category, Confidence, ReviewStatus, ModelVersion, and ProcessedDate. If the confidence is below the threshold, the flow creates an approval, posts an adaptive card in Microsoft Teams, or assigns a human reviewer. After review, the final decision is written back to SharePoint so the library becomes searchable, auditable, and easier to manage.

Build steps at a high level

In the current Power Automate designer, create an automated cloud flow and select a trigger such as 'When a file is created' for SharePoint. Add an action to get the file content. If you are using Azure AI Vision or Custom Vision, configure the endpoint and authentication method according to the Azure resource you created, often through a connector, custom connector, or HTTP action. If you are using AI Builder, add the appropriate AI Builder prediction action. Then parse the response, compare the confidence score with your threshold, and add a condition branch. The high-confidence branch can update metadata, move the file to an approved asset library, or notify a channel. The low-confidence branch should request review, capture the reviewer decision, and log the outcome for future model tuning.

Licensing and cost considerations

Licensing depends on the design. Standard Microsoft 365 connectors such as SharePoint, Teams, Outlook, and OneDrive may be included with many Microsoft 365 plans, but premium connectors, HTTP actions, custom connectors, Dataverse, unattended automation, and some advanced Power Platform features may require additional Power Automate or Power Platform licensing. AI Builder scenarios consume AI Builder capacity/credits. Azure AI Vision and Azure AI Custom Vision are Azure consumption services and should be budgeted, monitored, and governed like other Azure resources. For organizations buying through CSP, Microsoft licensing is commonly transacted under the New Commerce Experience (NCE), so term length, cancellation windows, seat changes, and add-on licensing should be reviewed before rollout.

Security, governance, and compliance

Image workflows can contain sensitive information, so governance should be designed from the start. Use Microsoft Entra ID groups for access control, least-privilege permissions for SharePoint libraries and Azure resources, environment separation for development/test/production, and Power Platform data loss prevention policies to control which connectors can be used together. Store secrets in approved locations such as Azure Key Vault where appropriate, avoid hard-coding keys in flows, and review whether managed identities or secured custom connectors fit the architecture. Define retention rules for uploaded images and prediction results, enable audit logging, document model ownership, and ensure that low-confidence or high-risk predictions include human-in-the-loop review.

Where this helps in 2026

The same pattern can support field inspections, product image tagging, retail shelf checks, equipment-condition review, marketing asset classification, compliance screening, warranty intake, facilities management, and document library enrichment. The workflow can start small with one SharePoint library and one approval path, then expand to Power Apps, Teams, Dataverse, and Azure AI as requirements mature. Accuracy is not automatic: custom models require representative training data, testing, monitoring, and periodic retraining. Business teams should treat confidence scores as decision support, not as a guarantee, especially when the output affects compliance, safety, finance, or customer-facing decisions.

Key takeaways

• Microsoft Flow is now Power Automate, and Azure Cognitive Services is now generally positioned as Azure AI services.
• Use Azure AI Vision for general image analysis, Azure AI Custom Vision for trained custom categories, and AI Builder for low-code Power Platform scenarios.
• A modern image workflow should include threshold-based branching, SharePoint or Dataverse metadata, Teams or Approvals notifications, and human review for low-confidence results.
• Licensing may involve Microsoft 365, Power Automate premium, AI Builder capacity, Azure consumption, and CSP/NCE subscription terms.
• Security and governance require Microsoft Entra ID access control, DLP policies, environment strategy, auditability, and careful handling of image data.

IT Partner can help you design and implement a governed Power Automate and Azure AI image-processing workflow, including licensing review, SharePoint architecture, Teams approvals, and Power Platform security.