How to Fix Power Platform AI Builder Problems
Why This Is Happening
I've been troubleshooting Power Platform AI Builder issues for years, and I can tell you one thing with certainty: the error messages Microsoft surfaces in this product are some of the least helpful in their entire ecosystem. You get something like "Model training failed" with a spinner that just disappears, or a Power Automate flow that silently returns null from an AI Builder action with zero explanation. I know how maddening that is , especially when you've already spent hours labeling training data or building the downstream automation that depends on a working model.
Power Platform AI Builder troubleshooting is complicated by the fact that it sits at the intersection of several independent systems: Microsoft Dataverse, Power Automate, Power Apps, and Azure Cognitive Services under the hood. When something breaks, the fault could live in any of those layers , and the AI Builder UI itself often doesn't tell you which one.
Here's the breakdown of what actually causes most AI Builder failures in practice:
- Capacity unit exhaustion: AI Builder runs on a credit system called AI Builder capacity. Every training run, every prediction call, every document processed consumes units. When you hit zero, or when your tenant admin hasn't allocated units to your environment, models stop working completely, often with no obvious error.
- Environment configuration issues: AI Builder requires a Dataverse-enabled environment. If your Power Platform environment was created without Dataverse, or if the AI Builder feature toggle was never enabled at the tenant level, you'll see models fail to train or publish.
- Stale browser sessions and cached tokens: The Power Platform maker portal at make.powerapps.com is a heavily JavaScript-driven single-page application. Corrupted session state causes all sorts of ghost failures, models that claim to be training forever, publish buttons that do nothing, dashboards that show wrong credit counts.
- License mismatches: AI Builder requires a Power Apps or Power Automate Premium license, or a dedicated AI Builder add-on. If a user's license expired, was reassigned, or if the tenant is on a trial that ended, predictions fail at runtime, sometimes hours after they worked fine.
- Model-specific data quality problems: For custom models (object detection, form processing, custom prediction), training failure is often a data issue, not enough labeled samples, inconsistent image resolution, PDF documents that contain scanned images rather than selectable text, or training/test splits that don't have enough coverage across all labels.
- Regional service outages and throttling: AI Builder predictions hit Azure backend endpoints. Transient HTTP 429 or 503 responses from Microsoft's infrastructure get surfaced in Power Automate flows as generic action failures, with no retry logic built in by default.
Who sees these problems most often? Developers building their first AI-powered automation, citizen developers who inherited a working solution that suddenly stopped, and IT admins who changed license assignments or reorganized environments without realizing the downstream impact on AI Builder capacity allocation. If any of those sound like you, you're in the right place. Browse all Microsoft fix guides →
The Quick Fix, Try This First
Before you go deep on diagnostics, run this sequence. I've seen it resolve roughly 60% of AI Builder issues without any further work.
Step one: Check your AI Builder capacity right now. Go to the Power Platform admin center at admin.powerplatform.microsoft.com, click Resources in the left nav, then select Capacity. Look at the AI Builder row. You need to see a positive number under "Available." If it says 0 or shows nothing, that's your problem, you've exhausted your capacity units or none have been allocated to this environment.
If capacity looks fine, move on. Open make.powerapps.com, select your environment from the top-right environment picker, then navigate to AI Hub in the left sidebar. Find the failing model and click it. Look at the Details pane on the right, specifically the Last trained date and Performance score. A model with a performance score below 70% will produce unreliable or empty predictions even when it technically runs.
Step two: Force a hard browser refresh. This one sounds too simple, but the Power Platform portal caches aggressively. Press Ctrl + Shift + Delete in Chrome or Edge, check "Cached images and files" and "Cookies and other site data" for the make.powerapps.com domain only, then hit Clear data. Sign back in and retry whatever was failing.
Step three: Re-test in an InPrivate/Incognito window. If the issue disappears in a private window, you have a browser profile corruption problem, not an AI Builder problem. That's a much easier fix.
Step four: Check the Power Platform service health dashboard. Go to admin.microsoft.com, click Health → Service health, and filter for "Power Platform" and "AI Builder." If Microsoft has an active incident, there's nothing to troubleshoot on your end, you just wait.
Power Platform AI Builder isn't free, and the license requirements are more nuanced than Microsoft's documentation makes clear. Here's exactly what you need and how to verify it.
Every user who creates, trains, or runs AI Builder models needs one of the following: a Power Apps Premium license, a Power Automate Premium license, or an explicit AI Builder add-on. Trial licenses include limited AI Builder capacity, but they expire after 30 days and often deplete mid-project without warning.
To check your own license: go to admin.microsoft.com → Users → Active users → find your account → click Licenses and apps. Confirm "Power Apps Premium" or "Power Automate Premium" shows a green checkmark. If you see "Power Apps (free)" or "Power Automate (free)," that's your problem right there.
For capacity units, the fuel AI Builder actually runs on, the check is slightly different. In the Power Platform admin center:
admin.powerplatform.microsoft.com
→ Resources
→ Capacity
→ AI Builder tabYou'll see total allocated capacity and how much is consumed. One AI Builder capacity unit equals 1 million service credits. Custom model training costs around 10 credits per page processed. Prebuilt model predictions (like sentiment analysis or business card reading) cost 1 credit per call. If you're running high-volume automations, it's easy to burn through capacity faster than expected.
To allocate capacity to a specific environment: from the Capacity page, click Allocate to environments, select your environment from the dropdown, and enter the number of units to assign. Changes take effect within a few minutes, no restart needed. After allocation, go back to your failing model and retry. You should see the training or prediction resume normally.
AI Builder requires Microsoft Dataverse. Full stop. If your environment was created as a "default" environment or was set up for Power Automate without explicitly enabling Dataverse, AI Builder will appear partially functional in the UI but fail when you actually try to use it, training jobs will hang, and published models won't respond.
To verify Dataverse is enabled, go to the Power Platform admin center, click Environments, select your environment, and look at the Details panel. Under "Type," you should see "Production," "Sandbox," or "Developer", all of these support Dataverse. If you see "Default" with no Dataverse listed under Resources, that environment is limited.
Separately, confirm that AI Builder is enabled at the tenant level. This is a global toggle that tenant admins sometimes disable to control costs:
admin.powerplatform.microsoft.com
→ Settings (gear icon, top right)
→ Tenant settings
→ AI Builder preview models → OnAlso check the environment-level setting. Click your target environment → Settings → Features → scroll to "AI Builder" and confirm the toggle is set to On.
If Dataverse isn't present, you'll need to create a new environment with Dataverse enabled. You can't retroactively add Dataverse to an existing environment that was created without it, that's a current Platform limitation. Once you have a Dataverse-enabled environment, you'll need to recreate your AI models there. Your flow and app connections can be updated to point to the new environment.
After confirming both settings, try opening AI Hub in the affected environment again. If the model list loads and shows your models with status indicators, you're past this hurdle.
When a model trains successfully but predictions return wrong results, empty outputs, or confidence scores below 50%, the model itself needs work, not the plumbing around it. This is the most common root cause of AI Builder form processing not working correctly and AI Builder object detection failing on real-world inputs.
First, check your training data quality. Navigate to AI Hub → click your model → Edit model. For document processing models, Microsoft recommends a minimum of 5 sample documents, but realistically you need 15–20 for decent accuracy, and 50+ for production-grade performance. Each document should be a real example of what the model will see in production, not ideal clean scans, but the messy, skewed, mixed-format documents your users actually submit.
For custom prediction models (the tabular data type), check the Training report after your last training run. This report shows:
- Feature importance, which columns actually influenced predictions
- Model accuracy score, anything below 80% is a red flag for production use
- Confusion matrix, where the model makes mistakes
To force a retrain with existing data:
AI Hub → Your model → Edit model
→ Next (skip data selection changes)
→ Train → Train nowAfter training completes (typically 5–30 minutes depending on data volume), click Publish. This is a step many people miss, a trained model is not automatically available to your flows and apps until you explicitly publish it. If your model was previously published and you've retrained, you need to publish again to push the new version live.
Once published, test with the Quick test function inside AI Hub before rerunning your full flow. This confirms the model returns results before you go debugging Power Automate.
Power Automate AI Builder connection errors and silent prediction failures are their own category of problem. The flow looks fine, runs without a red error, but the output from the AI Builder action is empty, null, or contains fields that never populated.
Start by opening your flow in make.powerautomate.com and examining the AI Builder action directly. Click on the action step (it's usually named something like "Predict" or "Process and save information from documents"). Look at the Connection field, it should reference a valid connection with a green dot. If it shows a warning triangle or says "Invalid," click the connection name, delete it, and re-add the connection by clicking Add new connection and signing in with your Microsoft 365 account.
For flows using the AI Builder: Predict action with a custom model, confirm the model ID is correct. Model IDs change when you delete and recreate a model. To find your current model ID:
AI Hub → Your model → Settings (three-dot menu)
→ Copy model IDPaste this back into your Power Automate action's Model field.
A very common cause of AI Builder prediction failing silently in flows is missing error handling. By default, Power Automate will not throw a visible error if an AI Builder action receives an HTTP 429 (throttled) or 503 response, it just returns empty. Add a Configure run after setting on the steps after your AI Builder action to handle "has failed" and "has timed out" branches, then log the full action outputs to a SharePoint list or send yourself an email with the raw response. This reveals exactly what Microsoft's backend is returning.
Also check that the user account under which the flow runs has an active AI Builder-eligible license. Flows running under a service account or shared mailbox often fail here because those accounts don't have the right license assigned.
If you're embedding an AI Builder component, like the Form Processor, Object Detector, or Text Recognizer, directly in a Power App canvas, you have a separate set of failure modes to investigate.
The most common Power Apps AI Builder component error is the component appearing blank or showing a spinning loader indefinitely. Nine times out of ten this is a data connection issue. In Power Apps Studio, click Data (cylinder icon in the left sidebar) and confirm that the AI Builder connection is listed and shows a green status. If it's red or missing, remove it and re-add it via Add data → search for "AI Builder" → select your model.
For the Form Processor component specifically, check that you're passing a compatible file type. Supported formats are PDF, JPEG, PNG, BMP, and TIFF. Files over 20 MB or PDFs with more than 200 pages will silently fail. Add an explicit file-size check in your app before passing the document to the component:
If(
DataURIToBlob(UploadedFile.Value).Size > 20971520,
Notify("File too large for AI Builder", NotificationType.Error),
AIFormProcessorComponent.Document = UploadedFile.Value
)For the Object Detector component, confirm your model is published (not just trained) and that the camera or image input your app is using matches the resolution range your training data covered. Models trained exclusively on high-resolution images often fail on low-quality webcam captures.
Finally, if you recently updated your Power App to a newer format version (select File → Settings → App settings → check "App format version"), AI Builder components may need to be removed and re-added to the screen, their property bindings sometimes break across major format version upgrades. Save a backup of your app before making this change.
Advanced Troubleshooting
If you've worked through all five steps above and AI Builder is still misbehaving, you're into the territory of environment-level or enterprise configuration problems. These take more digging but they do have solutions.
Check the Dataverse AI Builder Tables Directly
AI Builder stores model metadata, training jobs, and prediction results in specific Dataverse tables. If these tables are corrupted or missing permissions, everything breaks silently. Open make.powerapps.com, click Dataverse → Tables → search for "AIModel." You should see tables like msdyn_AIModel, msdyn_AIConfiguration, and msdyn_AIPredictionHistory. If these tables are missing or show permission errors, your Dataverse environment has a provisioning problem that typically requires a support ticket to resolve.
Audit the Power Platform Data Loss Prevention Policy
Enterprise IT teams frequently configure Data Loss Prevention (DLP) policies in the Power Platform admin center that block connectors. If your DLP policy places the AI Builder connector in the "Blocked" category, every AI Builder flow action will fail at runtime, often with an error that says "Connector is blocked by DLP policy" buried inside the flow run history. Check DLP policies here:
admin.powerplatform.microsoft.com
→ Policies → Data policies
→ Click each policy → Connectors tab
→ Search for "AI Builder" → confirm it's in "Business" not "Blocked"Review the Power Platform Activity Log
For organizations with Microsoft 365 E3/E5 licensing, the Microsoft Purview audit log captures AI Builder events. Go to compliance.microsoft.com → Audit → New search → set Activities to "Power Platform activities" and filter by "AIBuilder." This shows you exactly which model operations were attempted, by whom, and whether they succeeded or failed with error codes, far more useful than what the UI surfaces.
Network-Level Issues in Corporate Environments
AI Builder training jobs and prediction calls reach out to Azure endpoints at *.api.powerplatform.com and *.cognitiveservices.azure.com. If your corporate proxy, firewall, or SSL inspection appliance is intercepting these calls, training jobs will hang indefinitely and prediction calls will time out. Work with your network team to whitelist these endpoints and exclude them from SSL inspection. Microsoft's full list of required URLs for Power Platform is documented in their service URLs reference, your network team will need those ranges.
Domain-Joined Machine / Conditional Access Issues
If users hit AI Builder fine from personal devices but fail on corporate domain-joined machines, Conditional Access policies are the likely culprit. Check with your Azure AD / Entra ID admin whether a policy requires compliant devices or MFA re-authentication for the Power Platform app ID (475226c6-020e-4fb2-8a90-7a972cbfc1d4). Forcing a fresh token by signing out of all Microsoft 365 apps and signing back in often resolves temporary Conditional Access evaluation failures.
Prevention & Best Practices
Once you've got AI Builder working, the goal is to keep it that way. Most of the outages I've helped companies recover from were entirely avoidable with a few habits in place.
Monitor capacity consumption proactively. Don't wait until you hit zero. Set up a Power Automate flow that checks your AI Builder capacity on a weekly schedule using the Power Platform admin connector and sends an email alert when remaining units drop below 20%. This gives you time to purchase additional capacity before anything breaks in production. The Power Platform admin connector action you need is "Get organization capacity details."
Use separate environments for development and production. Labeling training data and running test predictions in your production environment burns real capacity units. Set up a dedicated sandbox or developer environment for AI Builder experimentation. Allocate a small fixed capacity to that environment so runaway tests can't deplete your production pool.
Version control your training data. Store your labeled training documents in SharePoint or Azure Blob Storage with versioning enabled. AI Builder doesn't keep a history of what data you trained on, if a model starts performing poorly after a data update, you need to be able to roll back to the previous training set. Without source control, you're flying blind.
Build retry logic into every flow that calls AI Builder. Transient failures from Microsoft's backend are real. Add a "Do until" loop around your AI Builder prediction action that retries up to 3 times with a 30-second delay between attempts. This alone eliminates most production incidents caused by momentary throttling or backend hiccups.
Retest published models after Microsoft updates. Microsoft updates AI Builder's prebuilt models regularly. These updates occasionally change output field names or confidence score ranges. Set a monthly calendar reminder to run a batch of known test inputs through your production models and verify the outputs match your expected values. Catch breaking changes before your users do.
- Set a Power Automate weekly alert when AI Builder capacity drops below 20% of your allocation
- Always train custom models on 50+ real-world examples, not idealized clean samples
- Wrap every AI Builder flow action with a "Configure run after" error branch that logs the full raw output
- Keep a dedicated sandbox environment for AI Builder experimentation so tests don't burn production capacity
Frequently Asked Questions
Why is my AI Builder model stuck on "Training" for hours?
A model that stays in "Training" status for more than 60–90 minutes is almost always stuck due to one of three causes: your environment's AI Builder capacity hit zero mid-training, the Dataverse service had a transient hiccup that orphaned the training job, or the training data itself has a format issue that's causing the backend to spin indefinitely. First, check your capacity in the admin center. If capacity is fine, cancel the training job by clicking the three-dot menu on the model and selecting "Cancel training," wait 5 minutes, then start a fresh training run. If the new job also stalls past 90 minutes, open a Microsoft support ticket, orphaned training jobs sometimes require backend intervention to clear.
AI Builder says "Not enough credits", but my admin says we have capacity?
This is almost always an allocation mismatch, not an actual shortage. Tenant-level capacity and environment-level allocated capacity are two different numbers. Your admin may have purchased 10 capacity units for the tenant but never explicitly allocated any to your specific environment. Have them go to Power Platform admin center → Environments → your environment → Resources → AI Builder capacity → Allocate units. Once units are assigned to your environment specifically, the "not enough credits" error resolves immediately, no rebuild of models required.
Can I use AI Builder without a Power Apps or Power Automate Premium license?
No, AI Builder requires at minimum a Power Apps Premium or Power Automate Premium license, or a standalone AI Builder add-on. The free and "seeded" versions of Power Apps and Power Automate that come bundled with Microsoft 365 E3/E5 do not include AI Builder access. There is a 30-day trial you can activate through the Power Platform admin center under Settings → AI Builder trial, which gives you access to a limited pool of capacity for evaluation. After that trial, you need a paid license or the models stop running.
Why does my form processing model fail on certain PDFs but work on others?
Form processing models trained in AI Builder work best on PDFs with selectable text (also called "digital" or "born-digital" PDFs). When someone scans a physical document and sends it as a PDF without running OCR, the file contains only an image, and your model needs to work harder to extract fields from it. If your failing PDFs are scanned images, try preprocessing them through the AI Builder "Text recognition" prebuilt model first to extract raw text, then pass that text through your custom model. Also confirm the failing PDFs aren't password-protected or encrypted, AI Builder cannot read locked documents at all.
My Power Automate flow AI Builder prediction action returns empty results, what's wrong?
Empty results from the AI Builder Predict action in Power Automate almost always mean one of four things: the model isn't published (only trained), the connection used by the flow has an expired token, the input data being sent to the model doesn't match the schema the model was trained on (mismatched column names or data types), or the backend returned a throttling error that Power Automate swallowed silently. Turn on run history, click a failed run, expand the AI Builder action, and examine both the "Inputs" and "Outputs" raw JSON. This will show you exactly what was sent and what came back, including any HTTP error codes that the UI hides from you.
How do I check how many AI Builder credits I have left in real time?
The clearest view is in the Power Platform admin center at admin.powerplatform.microsoft.com → Resources → Capacity → AI Builder tab. This shows both tenant-wide capacity and per-environment allocation. For a more granular breakdown of what's consuming your credits, go to make.powerapps.com → AI Hub → and look at the "Credits used" counter on each model's detail page. Note that usage reporting has up to a 24-hour lag, so the numbers you see reflect consumption from the previous day, not the current moment, factor that in when you're trying to diagnose a sudden capacity depletion.