Fix Power Platform AI Builder Issues, Full Troubleshooting Guide
Why This Is Happening
Power Platform AI Builder troubleshooting is one of those things that looks simple on the surface, until you're staring at a broken flow at 9 AM and your manager is asking why the invoice-processing automation stopped working overnight. I've seen this exact situation play out on dozens of enterprise tenants, and the frustrating part is always the same: the error messages Microsoft gives you are either cryptic codes like EntitlementNotAvailable or vague banners that say "credits consumed" without telling you what to actually do next.
AI Builder sits inside the Power Platform ecosystem, which means it inherits a layered permission and licensing model that most people never fully understand until something breaks. There are three distinct failure categories I see most often:
- Process/action deactivation: The underlying Dataverse processes that power AI Builder get turned off, either manually by a well-meaning admin or automatically during certain environment migrations and updates. When this happens, every AI Builder action in your flows and apps silently stops working.
- Credit exhaustion: AI Builder runs on a credit-based consumption model. When those credits run dry, or when the system thinks they've run dry due to a calculation delay of up to five days, flows start throwing errors like
QuotaExceededandCode="EntitlementNotAvailable"before you've even had a chance to react. - Licensing confusion: A premium Power Apps license comes with a small seeded number of AI Builder credits, but the app may still display a trial banner until the actual trial period ends. This trips up a lot of users who just purchased and assume something went wrong with their license activation.
There are also environment-specific issues, things like a 408 Dependency Timeout in Power Automate, a 429 TooManyRequests throttling error when you hammer AI Builder endpoints too fast, or document processing flows that refuse to save. Each of these has a specific root cause and a specific fix.
The good news: every single one of these is fixable without opening a Microsoft Support ticket, as long as you're an environment administrator or can get five minutes with one. This guide walks you through all of it, in order of what's most likely to solve your problem fastest. Browse all Microsoft fix guides →
The Quick Fix, Try This First
If your AI Builder actions are suddenly not working across flows and apps, and nothing changed on your end, the single most likely culprit is that the AI Builder processes in your Dataverse environment got deactivated. This is the number one cause of AI Builder actions being disabled or deactivated, and it affects the whole environment at once, which is why it feels so catastrophic when it happens.
Here's what you do. Sign in to the Power Platform admin center at admin.powerplatform.microsoft.com. Select the environment where things are broken. Go to Settings > Resources > All legacy settings. From the Settings menu that appears in the classic interface, click Processes.
Change the view to All Processes, this is important, because by default it only shows active ones. Look for any of the following AI Builder processes that show as deactivated:
- IsPaiEnabled
- Predict
- Predict by Reference
- PredictionSchema
- Train
- QuickTest
- BatchPrediction
- ScheduleTraining
- SchedulePrediction
- ScheduleRetrain
- UnschedulePrediction
- UnscheduleTraining
- CancelTraining
- PublishAIConfiguration
- UnpublishAIConfiguration
Select all deactivated ones from this list and activate them. Depending on the environment, this can take a minute or two to propagate. Once done, go back and test your flow or app, in the vast majority of cases, this is all it takes to get things running again.
If that doesn't fix it, don't stop here. Keep reading, the next most common cause is credit exhaustion, and that has its own fix path.
Let's do this properly from the beginning so nothing gets missed. You need to be the administrator of the environment, a regular user account won't have access to these settings.
Navigate to the Power Platform admin center:
https://admin.powerplatform.microsoft.comIn the left navigation, you'll see a list of environments. Click on the specific environment where AI Builder has stopped working. Once inside the environment detail view, click the Settings button in the top toolbar, not the sidebar item, the one that appears in the environment header area.
From there, go to Resources > All legacy settings. A new browser tab or window will open, taking you into the classic Dynamics 365 settings interface. This feels outdated, but this is the right place, Microsoft hasn't migrated process management to the new admin center yet.
Once inside, click the Settings dropdown at the top of the page, then select Processes. You'll land on a Processes list view. By default, this shows only "My Processes" or "Active Processes" depending on your version. Use the view dropdown and switch it to All Processes.
Now search or scroll to find the AI Builder processes listed above. Select each deactivated one (they'll have a "Draft" or "Deactivated" status), and use the Activate button in the toolbar. Do this for every process in the AI Builder list.
What you should see after this works: go back to your Power Automate flow and run it manually. The AI Builder action should execute without errors. If the action was previously returning a blank output or skipping silently, it should now produce actual results.
If your flow is throwing errors like All AI Builder credits in this environment have been consumed, or you're seeing error detail codes of EntitlementNotAvailable or QuotaExceeded, or the message Credit usage exceeds allocation, this is a credit problem, not a process problem.
Here's the thing that surprises most people: credits aren't blocked the moment you hit 100%. Microsoft blocks AI Builder actions at 125% of your allocated credits. This means by the time you see the error, you've already gone 25% over your limit. And because the system recalculates consumption with a delay of up to five days, you might not even see the error in real time.
To check your credit situation, go to the admin center and navigate to Resources > Capacity in the left pane. On the right side, click the Add-ons tab. You'll see a list of environments with their allocated AI Builder credit counts.
If your environment isn't listed, or shows zero credits, that means it's drawing from tenant-level unassigned credits rather than environment-level ones. To see the total tenant credits, switch to the Summary tab in that same Capacity pane and scroll down to the Add-ons section.
To add more credits to your specific environment, click Assign to an environment in the Add-ons view, select your environment from the panel that slides in, and modify the credit count. If you don't have spare tenant credits to reallocate, you'll need to either pull from another environment or purchase additional AI Builder capacity.
One more thing: check that the Allow users to consume unassigned credits setting is enabled. Go to Power Platform admin center > Settings > AI Builder credit page to toggle this on if it's off.
Two of the most common HTTP-level errors in AI Builder flows are the 408 Dependency Timeout and the 429 TooManyRequests. They're different problems with different causes, and mixing up the fixes will waste your time.
408, Dependency Timeout: This means the AI Builder service took too long to respond. It's not necessarily your fault. It happens when the model you're calling is processing a large document, when the Power Platform region is under load, or when your flow doesn't have retry logic. The fix is to add a Retry Policy to the AI Builder action in your flow. In the flow designer, click the three dots (…) on the AI Builder action, go to Settings, and set the Retry Policy to Fixed Interval with at least 3 retries and a 60-second interval. This alone resolves most 408 situations.
429, TooManyRequests: This is a throttling error. You're calling AI Builder faster than the service allows. This typically shows up in high-volume flows that process many items in a loop without any pacing. The fix here is to introduce concurrency controls and delays. In Power Automate, find the loop or parallel branch that's triggering the AI Builder action repeatedly, open its settings, and reduce concurrency to 1. You can also add a Delay action between iterations, even a 2-3 second delay is often enough to stop the throttling.
For both errors, also check whether your AI Builder model itself is trained and published. A model that's in a broken training state can cause dependency failures that look like timeouts. In AI Builder, navigate to your model and confirm its status shows Published, not Training, Draft, or Failed.
After making these changes, re-run your flow. A successful run with the retry policy in place will show you in the flow run history exactly how many retries were needed, useful data for right-sizing your retry configuration going forward.
This one is genuinely annoying because it looks like your purchase didn't go through, but it did. Here's what's happening: when you purchase a premium Power Apps license, you automatically get a small number of seeded AI Builder credits included with that license. But if you previously started a trial, the trial banner persists in the top of the AI Builder interface until the trial period actually expires on its own.
The important thing to understand is this: the trial banner is cosmetic at this stage. Your licensed credits are there. Your flows and apps are consuming your purchased credits, not trial credits. The banner will disappear on its own once the original trial period ends, you don't need to cancel anything or contact support.
To verify that your purchased credits are in place and active, go to the Power Platform admin center and check the Capacity > Add-ons view as described in Step 2. If you see your environment listed with an allocated credit count that matches what you purchased, your license is working correctly.
If you want to speed up the resolution, or if the trial banner is blocking functionality (not just appearing cosmetically), contact your Microsoft account team or tenant admin to check whether the trial and purchased licenses are conflicting at the Azure Active Directory level. In enterprise tenants, sometimes a license assignment conflict needs to be resolved manually in the Microsoft 365 admin center under Billing > Licenses.
One thing that does not work: unassigning and reassigning your license through the M365 admin center to "reset" it. This rarely changes the trial banner behavior and can sometimes temporarily disrupt access. Just wait it out, the trial banner goes away cleanly at expiry.
Two more scenarios that trip people up: the AI Builder action runs without errors but the output is empty, and separately, a model that gets stuck with a "Training failed and the model can't be used" status.
Empty output from AI Builder actions: When an AI Builder action completes but the downstream action returns empty data or an error saying the output isn't available, the most common cause is that the model isn't properly published, or the output schema isn't being referenced correctly in the flow. First, confirm the model is in Published status inside the AI Builder portal. Then, in your flow, delete and re-add the AI Builder action, sometimes the dynamic content picker in Power Automate caches an old schema and doesn't pick up new output fields from a retrained model. Re-adding the action forces it to pull a fresh schema. Also check that the action's response is not inside a conditional branch that's silently evaluating to false.
Training failed error: If your model's status shows as failed after training, and AI Builder won't let you use it, you have a few options. First, try running a quick test from inside the AI Builder model page, sometimes a training failure is transient and a QuickTest call will succeed anyway. If the model is genuinely broken, you'll need to retrain it. Before retraining, review your training data: common causes of training failures include corrupt files in your document set, too few training samples (AI Builder requires a minimum number of labeled examples depending on model type), or labels that are inconsistently applied.
For document processing models specifically, check the Pages setting when saving the flow. There's a known issue where saving a document processing flow with certain page configurations fails, if you're getting an error specifically when saving (not running), try simplifying the page range setting or removing and re-adding the document processing action.
After a successful retrain, always click Publish explicitly, training alone doesn't make the model live. The Published state is what your flows and apps consume.
Advanced Troubleshooting
If the standard fixes above haven't resolved your AI Builder troubleshooting situation, it's time to dig into the environment and tenant configuration at a deeper level. These scenarios come up less often, but when they do, they're typically harder to diagnose because they involve interactions between systems rather than a single misconfigured setting.
Plug-in Steps Being Turned Off
In Power Apps specifically, there's a known issue where plug-in steps associated with AI Builder can get turned off, particularly in environments that have undergone solution imports, managed layer changes, or recent updates from the publisher. When this happens, AI Builder components in canvas and model-driven apps stop functioning, but the error message you see often points to the component rather than the underlying plug-in step.
To check this, go to the Power Platform admin center, navigate to your environment, then go to Settings > Resources > All legacy settings > Settings > Plugins. This opens the Plug-in Trace Log and Step Registration views. Look for any AI Builder-related plug-in steps that show as disabled. Re-enable them, then test your app again.
Understanding Credit Consumption in Different Contexts
One subtlety that causes real confusion in enterprise tenants: AI Builder credits don't work the same way everywhere in Power Platform. In Power Apps and Power Automate, AI Builder actions consume AI Builder credits first, and only fall back to Copilot Credits if those are exhausted or unavailable. But in Microsoft Copilot Studio, including agents and agent flows, AI Builder actions consume only Copilot Credits, never AI Builder credits. This matters a lot if you're looking at credit dashboards and wondering why your AI Builder credit balance looks fine while your Copilot Studio-based automations are failing.
If you're running AI Builder inside Copilot Studio and hitting credit errors, you need to look at your Copilot Credit allocation, not your AI Builder credit allocation. These are managed separately in the admin center.
Environment-Level vs. Tenant-Level Credit Sources
When you're diagnosing a QuotaExceeded error, knowing whether your environment uses allocated or unassigned credits changes the fix path. If credits come from the environment level, you can reallocate without purchasing more. If they come from the tenant level (because the environment isn't listed in the Add-ons tab, or shows zero allocated), you need to either enable the Allow users to consume unassigned credits setting or explicitly allocate credits to that environment from the pool. Admins sometimes disable the unassigned credit setting as a cost control measure, which then silently blocks entire environments from using AI Builder.
Image Rotation Issues with AI Builder Components
If you're using an AI Builder camera component in a Power App and the scanned images are coming out rotated incorrectly, this is a known display bug related to EXIF metadata handling on certain mobile devices. The workaround is to pre-process images through a flow that corrects rotation before passing them to the AI Builder model. There's no admin-level setting that fixes this, it has to be handled at the app or flow level.
Business Card Reader Not Working for Some Users
If the AI Builder business card reader component works for some users in your org but not others, the cause is almost always a per-user permission or license issue rather than an environment problem. Check that the affected users have the appropriate Power Apps license assigned and that their security roles in the environment include the AI Builder user permissions. A common mistake is assigning environment access but forgetting to grant the AI Builder-specific role.
Prevention & Best Practices
Once you've fixed an AI Builder issue, the last thing you want is to go through this again next month. The good news is that most of the problems covered in this guide are entirely preventable with a few proactive habits.
Monitor your credit consumption before it becomes a crisis. The Power Platform admin center sends notification emails to environment admins when consumption approaches the allocated limit, but those emails only go to the admin account, not to the flow owners or app makers who are actually affected. Set up a shared distribution list or alias as the admin contact so the right people get the warning. Better yet, build a scheduled Power Automate flow that queries the capacity APIs and posts a summary to your team's Teams channel weekly.
Don't let credit calculation delays catch you off guard. The fact that AI Builder credit consumption can lag by up to five days means you might blow past your limit and not know it until flows start failing days later. Factor this into your capacity planning, treat your allocated credits as if 80% of the limit is your real ceiling, and set alerts at that threshold.
Document every AI Builder process activation. When you activate those fifteen processes in the admin center, write it down in your runbook with the date and who did it. If your organization does regular environment refreshes or solution deployments, these processes can get deactivated again. Having a documented baseline means your next admin doesn't spend two hours figuring out what broke, they check the runbook first.
Use retry policies in every flow that calls AI Builder. Make this a standard part of your flow development template. A fixed-interval retry with 3 attempts and 60-second intervals handles the majority of transient 408 timeout errors without any human intervention.
Keep your AI Builder models trained on fresh data. A model trained six months ago on outdated documents will gradually drift in accuracy, and a poorly performing model can cause unexpected outputs that look like bugs. Schedule quarterly model reviews and retrains as part of your maintenance cycle.
- Set up credit consumption alerts at 80% of allocation in the Power Platform admin center, don't wait for the 125% hard block
- Add retry policies (3 retries, 60-second fixed interval) to every AI Builder action in Power Automate as a development standard
- After any environment refresh or solution deployment, immediately verify the fifteen AI Builder processes are still in Active status
- Assign AI Builder credit monitoring responsibility to a named person on your team, "everyone's job" means no one's job
Frequently Asked Questions
Why does my AI Builder action work for some flows but not others in the same environment?
This usually points to a model-level issue rather than an environment-level one. Each AI Builder action in a flow is tied to a specific published model. If one flow is calling a model that's in Draft or Training state while another calls a properly Published model, you'll see exactly this split behavior. Go into AI Builder, check the status of every model your flows reference, and make sure each one shows Published. Also check whether the affected flows are owned by different users, a model shared with limited permissions might be accessible to some flow owners but not others.
How long does it take for AI Builder credits to reset after they're consumed?
AI Builder credit consumption resets at the beginning of each calendar month, it's not a rolling 30-day window. So if you exhaust your credits on the 28th, you're looking at only a few days before the counter resets. The catch is the calculation delay: the system can take up to five days to register consumption, which means your credit dashboard might show available credits even when you're actually over your limit. Microsoft does not bill you retroactively for this overage, the excess consumption is absorbed without charge, but your actions will be blocked during that period.
What's the difference between AI Builder credits and Copilot Credits, and which one do I need?
AI Builder credits are used when you call AI Builder actions from Power Apps or Power Automate. Copilot Credits are used when AI Builder actions run inside Microsoft Copilot Studio, agents and agent flows consume only Copilot Credits, never AI Builder credits, even if the action itself is technically an AI Builder operation. If you're building primarily in Power Apps and Power Automate, AI Builder credits are what you need to purchase. If you're building Copilot Studio agents, focus on Copilot Credit capacity. Both pools are managed separately in the Power Platform admin center under Resources > Capacity.
I reactivated all the AI Builder processes but they keep getting deactivated again, why?
Processes reverting to deactivated after you manually activate them is a sign of a deeper conflict in your environment. The most common causes are: a managed solution in your environment that has those processes set to deactivated as part of its configuration (solution layers can override manual changes), or a custom plug-in or workflow that's programmatically deactivating them as a side effect. Check your Solutions list in the maker portal and look for any solution that includes AI Builder processes. You may need to unmanage or update that solution, or work with the solution publisher to fix the deactivation behavior. This is also worth flagging to Microsoft Support with your environment ID.
My AI Builder model training keeps failing, what are the most common reasons?
Training failures usually come down to one of four things: not enough training samples (every AI Builder model type has a minimum, and getting close to that minimum produces fragile models that often fail), inconsistent or missing labels in your training data, corrupt or unsupported file formats in your document set, or a transient service issue on Microsoft's end. Start by reviewing your training data quality, remove any files that are password-protected, heavily scanned with poor image quality, or in edge-case formats. If you're above the minimum sample count and your data looks clean, try retraining once more; transient failures do happen. If it fails a second time, check the AI Builder model page for any specific error detail and bring that to Microsoft Support.
Can I use AI Builder if my organization hasn't purchased a standalone AI Builder license?
Yes, with limitations. A premium Power Apps per-user license includes a small number of seeded AI Builder credits, enough for light testing and low-volume use cases. These seeded credits are not the same as purchasing AI Builder capacity outright, and they won't support production-grade automation workloads that process hundreds or thousands of documents per month. If you're hitting credit errors regularly on seeded credits, that's a sign you need to purchase dedicated AI Builder capacity for your tenant or environment. Your Microsoft account team or a Power Platform licensing specialist can help you estimate the right credit volume based on your actual consumption patterns.