Build an AI Invoice Triage Flow

Build a Power Automate flow that routes invoices using AI confidence scores.

Introduction

30 Second Summary

Invoice teams lose time when uncertain details look just as trustworthy as clear ones. A review rule gives people a visible signal before a weak prediction moves forward.

In this project, you will build an invoice triage automation in Power Automate. Your flow turns each uploaded invoice into a visible confidence-based decision.

What You'll Build

You'll watch AI Invoice Triage turn a synthetic invoice into visible details plus a REVIEW REQUIRED or AUTO-READY result.

By the end of this project, you'll have:

An upload-to-result flow that accepts an invoice file before finishing with green check marks.
Transparent invoice details showing invoice ID, vendor, total, plus confidence from AI Builder or a clearly labelled Prototype mode fixture.
Evidence of both decision routes in run history through one REVIEW REQUIRED result plus one AUTO-READY result.
Secret Mission: Add invoice ID confidence to the gate so either weak prediction triggers REVIEW REQUIRED.

Are there any prerequisites?

You'll need a work or school Microsoft account on a Windows computer with Microsoft Edge. No prior Power Automate experience is required.

Before We Start

This checkpoint locks in the business problem: you are building an AI-powered invoice triage flow that exposes invoice data and sends low-confidence results to human review. Access to Power Automate, a usable environment, Dataverse, and AI Builder or Copilot capacity still needs verification in the next step.

Set Up and Choose Your AI Path

Your invoice triage flow needs a place to run before it can accept files. Power Auto8mate stores each cloud flow inside a selected Power Platform environment.

Live invoice extraction through AI Builder depends on tenant access. It also depends on confirmed organization capacity.

This step establishes an honest Live AI or Prototype path before any automation is built. Your chosen path keeps every later result clear about where its invoice data came from.

In this step, get ready to:

Confirm Power Automate access in a usable environment.
Prepare a synthetic English invoice for safe testing.
Choose Live AI mode or Prototype mode from the available evidence.

Verify your Power Automate access

The environment determines where your flow is stored. It also determines which AI capabilities your organization makes available.

Press the Windows key to open the search bar.
Type Microsoft Edge and press Enter to open it.
Visit the Power Automate web portal in Microsoft Edge.
Sign in with your work or school account.
Complete any organization sign-in prompt that appears.
Select the environment switcher in the Power Automate header.
Choose an environment where your account can create flows.

Why Does the Environment Matter?

Flows created in one environment are not visible from another. Selecting the intended environment keeps your work easy to find.

AI Builder also requires Dataverse in the environment. Access to the AI model later in this step provides useful evidence about that environment.

License pages can feel dense. The only question here is whether the visible information confirms capacity for live AI execution.

Select Settings in the Power Automate header.
Select View My Licenses.
Review the visible user-license information.
Review the visible environment-capacity information.
Look for confirmed AI Builder credits or Copilot Credits.
Treat live capacity as unconfirmed when the page does not clearly show it.

What Confirms Live AI Access?

Live AI mode requires the invoice action to be available in your environment. Your organization must also confirm available AI Builder credits or Copilot Credits.

Visible access to Power Automate alone does not confirm live AI capacity. Prototype mode keeps the workflow lesson available when that evidence is missing.

You have cleared the first access hurdle. Power Automate is open in a known environment with its visible license information inspected.

Prepare a synthetic sample invoice

A synthetic invoice lets you test extraction without exposing anyone's personal or financial information. The model accepts one clear JPEG, PNG, or PDF invoice up to 20 MB.

Visit the official AI Builder sample-data page.
Select English version.
Choose one synthetic invoice from the downloaded English sample set.
Confirm the chosen invoice is a JPEG, PNG, or PDF file.
Use this synthetic invoice for the manual tests in this project.

Keep Invoice Data Synthetic

Real invoices can contain names, addresses, payment details, or account references. The official sample gives you realistic test data without exposing that information.

Your test file is ready. You now have an English synthetic invoice that can safely travel through the flow.

Test the model and choose your path

The AI Models page lets you inspect a prebuilt model before placing it inside a flow. This check reveals whether the invoice model is available in your selected environment.

Testing Versus Running

Testing a prebuilt model on the AI Models page does not consume credits. Running AI Builder features inside a flow consumes AI Builder credits or Copilot Credits.

This model-page check is safe to perform. Live AI mode still requires confirmed organization capacity.

Switch back to Power Automate in Microsoft Edge.
Select More in the left navigation pane.
Select AI hub.
Under Discover an AI capability, select AI models.
Look for Extract information from invoices in the most popular section.

Before you test the sample, which invoice fields do you expect the prebuilt model to identify?

✔️ Live capacity is confirmed

Use this path when the invoice model is available. Your organization must also have confirmed runtime capacity.

Select Extract information from invoices.
Select Use prebuilt model.
Upload the synthetic English sample invoice in the model test area.

You should see extracted values such as invoice ID, vendor name, and invoice total. This result confirms that the prebuilt model can read the synthetic invoice.

Choose Live AI mode for this project.

What Live AI Mode Means

Live AI mode uses AI Builder's Extract information from invoices action. Each executed AI action can consume organization capacity.

ⓧ Capacity is unconfirmed

The model-page test can work without proving that your flow has runtime capacity. Unconfirmed capacity makes Prototype mode the transparent choice.

Select Extract information from invoices.
Select Use prebuilt model.
Upload the synthetic English sample invoice in the model test area.

You should see extracted invoice values from the free model-page test. This result confirms model access without claiming that live flow execution is available.

Choose Prototype mode for this project.

What Prototype Mode Means

Prototype mode uses clearly labeled Compose fixtures with simulated values. Every later result remains explicit about the source of its data.

ⓧ The invoice model is unavailable

The invoice model may be hidden when the selected environment lacks the required access. Its absence leaves live execution unconfirmed.

Choose Prototype mode for this project.

Prototype mode keeps the flow runnable with clearly labeled simulated values. It does not present fixture data as an AI prediction.

Can't See AI Hub or the Invoice Model?

Switch back to the Power Automate header to confirm that the intended environment is still selected.
Use Prototype mode when the selected environment does not expose AI Builder or confirmed capacity.

AI Builder requires an environment with Dataverse. Your organization can also restrict model access or capacity.

Help me identify why AI Builder is unavailable in my Power Automate environment. If you still need help, share a redacted screenshot with the NextWork community.

Before the final check, which mode do you expect your access evidence to support?

Check the Power Automate header for the environment you selected.
Confirm that your synthetic English sample invoice is ready for manual testing.
Compare your license evidence with the outcome you followed above.

You should see your selected environment in the Power Automate header. Your evidence should support one defensible choice: Live AI mode with confirmed capacity or Prototype mode with clearly labeled simulated values.

Record your final AI path in the checkpoint below.

Your setup now has a safe sample invoice plus an honest AI path. Next up, you will build the first runnable version of AI Invoice Triage.

Build the Invoice Intake Flow

You chose an honest Live AI mode or Prototype mode path in the previous step.

AI can only help after an invoice file reaches the automation. A small intake flow proves that path before you add extraction.

In this step, you will build an instant cloud flow in Power Automate. A manual test will return a visible confirmation from your synthetic invoice.

In this step, get ready to:

Create a manually triggered instant cloud flow named AI Invoice Triage.
Add a file input named My invoice.
Return Invoice received for triage from a successful test.

Create the instant cloud flow

An instant cloud flow runs when its trigger fires. A manual trigger starts only when you invoke it.

Why Use an Instant Cloud Flow?

Desktop flows automate web or desktop interfaces.

An instant cloud flow keeps this project focused on the file input. Its manual trigger avoids dependencies on storage or messaging services.

Return to the Power Automate tab from earlier.
Select My flows in the left navigation.
Select New flow.
Select Instant cloud flow.
Enter AI Invoice Triage as the flow name.
Select Manually trigger a flow.
Select Create.

You now have the flow's starting signal in place. The designer shows Manually trigger a flow at the top.

Add the invoice file input

Inputs define the data a manual run asks you to provide. A file input gives the trigger a place to receive the invoice itself.

Select +Add an input on the Manually trigger a flow card.
Select File.

You will see a file control labeled File Content inside the trigger.

Rename File Content to My invoice.

The trigger now asks for My invoice whenever you start a manual run.

Return a visible intake confirmation

The Compose action displays a fixed receipt message. That message proves the file trigger reached the next stage.

Select the add-action control below the trigger.
Select Data Operation in the action picker.

The action picker now shows the data operations available to your flow.

Select Compose.
Enter Invoice received for triage in Inputs.
Select Save.

Your invoice intake structure is now saved. Double-check its configuration before the first run.

✔️ Awesome, I've got everything!

Your saved flow is ready for its first manual intake test.

ⓧ I'd like to double check the full flow

The flow name is AI Invoice Triage. The trigger is Manually trigger a flow.

The trigger has one File input named My invoice.

The flow has one Compose action. Its Inputs value is Invoice received for triage.

Before you test the flow, what do you think the Compose output will show after you choose the invoice?

Select Test in the upper-right corner.
Choose Manually.
Select Test.
Choose the synthetic sample invoice from earlier for My invoice.
Select Run flow.

The run completes with green check marks on the successful steps.

Expand Compose in the run results.

You will see Invoice received for triage in the Compose output. That is your first complete invoice intake run.

Run Not Completing?

Confirm that My invoice contains the synthetic sample invoice before you select Run flow.
Confirm that the sample invoice uses JPEG, PNG, or PDF format.
Keep the sample invoice at 20 MB or smaller.
Check that the Compose input contains Invoice received for triage exactly.

Help me troubleshoot my invoice intake flow.

Your intake flow now accepts a file. Next, you will turn that file into visible invoice fields.

Extract Invoice Data

Your Power Automate instant cloud flow now accepts the synthetic invoice. Its confirmation only proves that the file arrived.

This step turns the invoice into four visible values. Your chosen path supplies either AI Builder dynamic content or clearly labelled Compose fixtures.

In this step, get ready to:

Connect the invoice to the data source for your chosen AI path.
Expose invoice ID, vendor, total, and confidence as separate outputs.
Run the flow to inspect each extracted or simulated value.

Connect the invoice data source

The flow needs an honest source for its invoice values. Live AI mode reads the uploaded file through a prebuilt invoice model. Prototype mode uses fixed fixtures that remain clearly identified as simulated data.

Live AI mode

Return to the AI Invoice Triage designer from the previous step.
Select the plus sign between Manually trigger a flow and the existing Compose action.
Enter Extract information from invoices in the Add an action search field.
Select Extract information from invoices from the AI Builder results.
Complete any connection prompt with the same work or school account.
Select the Invoice file input.
Select My invoice from the dynamic content panel.
Keep Asynchronous Pattern set to On.

You should now see the invoice extraction action between the manual trigger and the intake confirmation.

Select Save on the command bar.

Why use the prebuilt invoice model?

The prebuilt model already recognizes common invoice fields. This keeps the project focused on workflow decisions instead of model training.

Prototype mode

Prototype fixtures preserve the structure of the workflow when live AI execution is unavailable. Their fixed values make the source of the demonstration transparent.

Return to the AI Invoice Triage designer from the previous step.
Select the plus sign beneath the existing intake confirmation action.
Enter Compose in the Add an action search field.
Select Compose under Data Operation.
Enter INV-PROTOTYPE-001 in the Inputs field.
Select the Compose text on the action card title bar.
Enter Invoice ID as the action name.

You should now see an Invoice ID fixture card containing the prototype identifier.

Add another Compose action beneath Invoice ID.
Enter Contoso Supplies in the Inputs field.
Select the Compose text on the new action card title bar.
Enter Vendor as the action name.
Select Save on the command bar.

The designer now shows two clearly labelled prototype fixtures. No AI action sits in the runnable path.

Surface the four invoice outputs

Separate Compose actions make each invoice value easy to inspect in the run details. The action names also reveal exactly what each output represents.

Live AI mode

Select the plus sign beneath the existing intake confirmation action.
Enter Compose in the Add an action search field.
Select Compose under Data Operation.
Select the Inputs field.
Select Invoice ID from the dynamic content panel.
Select the Compose text on the action card title bar.
Enter Invoice ID as the action name.

The first visible output now maps the model's invoice identifier.

Add another Compose action beneath Invoice ID.
Select the new action's Inputs field.
Select Vendor name from the dynamic content panel.
Select the Compose text on the new action card title bar.
Enter Vendor as the action name.

Why use separate outputs?

Each Compose action gives one field its own card in the run details. You can inspect a missing or unexpected value without searching through the full AI response.

Add another Compose action beneath Vendor.
Select the new action's Inputs field.
Select Invoice total (number) from the dynamic content panel.
Select the Compose text on the new action card title bar.
Enter Invoice total as the action name.

You should now see separate action cards for the invoice ID, vendor, and total.

Add another Compose action beneath Invoice total.
Select the new action's Inputs field.
Select Confidence of invoice total from the dynamic content panel.
Select the Compose text on the new action card title bar.
Enter Invoice total confidence as the action name.
Select Save on the command bar.

The canvas now includes all four visible invoice outputs beneath the extraction action.

Prototype mode

Add another Compose action beneath Vendor.
Enter 1250 in the Inputs field.
Select the Compose text on the new action card title bar.
Enter Invoice total as the action name.

The prototype path now exposes its invoice amount as a separate output.

Add another Compose action beneath Invoice total.
Enter 0.60 in the Inputs field.
Select the Compose text on the new action card title bar.
Enter Invoice total confidence as the action name.
Select Save on the command bar.

The canvas now includes four clearly labelled fixtures. Their values are INV-PROTOTYPE-001, Contoso Supplies, 1250, and 0.60.

What does confidence mean?

AI Builder confidence scores range from 0 for low confidence to 1 for high confidence. The score represents how confident the model is about a specific extracted field.

Prototype mode uses the fixed value 0.60 to make uncertainty visible. The fixture remains clearly identified as simulated data.

Test and inspect the results

The next run reveals how far the current flow can take an invoice. Before you test it, do you think these outputs can decide whether the invoice needs human review?

Select Save on the command bar.
Select Test.
Choose Manually.
Select Test.
Select the file picker for My invoice.
Choose the synthetic sample invoice from earlier.
Select Run flow.

You should see green check marks beside the completed trigger and actions.

Expand Invoice ID to inspect its Outputs.
Expand Vendor to inspect its Outputs.
Expand Invoice total to inspect its Outputs.
Expand Invoice total confidence to inspect its Outputs.

In Live AI mode, you will see values derived from the synthetic sample invoice. In Prototype mode, you will see INV-PROTOTYPE-001, Contoso Supplies, 1250, and 0.60.

Good work. Your successful run now exposes four invoice values with an honest source.

No review decision appears in the run. That shortfall is intentional because the flow still treats every confidence value equally.

Missing values or blocked live execution?

A Live AI run can be blocked when your environment lacks available organization capacity. The run details may identify EntitlementNotAvailable, NoCapacity, or QuotaExceeded.

Delete the four Compose output actions that depend on AI Builder dynamic content.
Delete the Extract information from invoices action.
Follow the Prototype mode instructions in the first two substeps.
Record that Live AI execution requires administrator action.
Repeat the manual test with the synthetic sample invoice.

If an output is blank, compare its Inputs value with the matching dynamic content or prototype fixture above.

Help me diagnose my invoice output.

Your flow can now expose invoice data and its confidence. Next, you will turn that uncertainty into a visible review decision.

Add the Confidence Gate

Your Power Automate flow now exposes the invoice ID, vendor, total, and invoice-total confidence. It still accepts every result without making a review decision.

This step adds conditional routing based on the invoice-total confidence. A confidence score below 0.65 sends the invoice to human review.

In this step, get ready to:

Compare Invoice total confidence with 0.65 by using a Condition.
Assign visible decisions to the True and False branches.
Prove the gate with a successful manual run.

Define the review threshold

Both AI paths already expose a Compose output named Invoice total confidence. The gate can therefore use the same rule in Live AI mode or Prototype mode.

Return to the AI Invoice Triage designer from earlier.
Select the plus button directly below Invoice total confidence.
Choose Condition from the action picker.
Select the first value field in the Condition card.
Choose the dynamic content from the Compose action named Invoice total confidence.
Set the comparison operator to less than.
Enter 0.65 in the final value field.

You should now see a Condition card with True and False branches beneath the comparison.

How the gate works

An AI Builder invoice confidence score ranges from 0 to 1. A score near 0 indicates low confidence.

This project uses 0.65 as its review threshold. The Condition follows the True branch when the score is lower.

Create the review route

A true result means the invoice total carries too much uncertainty for automatic handling. The branch needs a visible output that calls for human review.

Locate the True branch in the Condition card.
Select the add-action control inside the True branch.
Choose Data Operation from the action picker.
Choose Compose.
Enter REVIEW REQUIRED in the Inputs field.

You should see the Compose card inside the True branch with REVIEW REQUIRED in its Inputs field.

Complete and test the gate

A false result means the confidence meets or exceeds the review threshold. This branch marks the invoice as ready for automatic handling.

Locate the False branch in the Condition card.
Select the add-action control inside the False branch.
Choose Data Operation from the action picker.
Choose Compose.
Enter AUTO-READY in the Inputs field.

You should see the second Compose card inside the False branch with AUTO-READY in its Inputs field.

Select Save in the designer toolbar.

Before you test, picture the branch that your current Invoice total confidence value should take.

Select Test in the upper-right designer toolbar.
Choose Manually as the test type.
Select Test again to continue to the manual inputs.
Choose the synthetic sample invoice in the My invoice file input.
Select Run flow.

You should see green check marks on the successful trigger and actions.

Expand the completed Condition action.
Expand the Compose action in the executed branch.

You should see either REVIEW REQUIRED or AUTO-READY in the executed Compose output.

In Prototype mode, the 0.60 fixture takes the True branch. You will see REVIEW REQUIRED as the decision.

That is the decision gate working. Your flow now turns uncertainty into a visible business decision.

Decision output missing?

Confirm that the first Condition value comes from the Compose action named Invoice total confidence.
Check that the comparison operator is less than.
Verify that both Compose cards sit inside their matching branches.

Help me debug the confidence gate in my invoice flow.

Your flow now turns invoice uncertainty into a review decision. Next, you will exercise both branches with controlled confidence values.

Prove Both Routes and Inspect Run History

Your confidence gate already turns uncertainty into a visible decision. A single successful run only proves that the flow can finish.

In this step, you will use controlled values to prove that both branches work. You will inspect the evidence in Power Automate's run history.

In this step, get ready to:

Prove that a confidence value of 0.60 produces REVIEW REQUIRED.
Prove that a confidence value of 0.90 produces AUTO-READY.
Restore the normal confidence source before inspecting both successful runs.

Test the review route

A test fixture gives the Condition a controlled input. The fixed value makes the expected branch deterministic.

The first fixture sits below the 0.65 threshold. This creates evidence for the human-review route.

Return to the AI Invoice Triage designer.
Select the existing Condition.
Remove Invoice total confidence from the left value field.
Enter 0.60 in the empty left value field.
Select Save.

Before you run the flow, which branch do you expect a confidence value below 0.65 to execute?

Select Test.
Choose Manually.
Select Test.
Choose the synthetic sample invoice for My invoice.
Select Run flow.
Wait for the manual run to finish.

You'll see green check marks on the successful steps. The True branch contains the executed Compose action.

Expand Condition to inspect its inputs.
Expand the executed Compose action in the True branch.

The Condition inputs show 0.60. The executed Compose output reads REVIEW REQUIRED.

Review route not executing?

Confirm that the left value in the Condition is 0.60.
Confirm that the comparison checks whether the left value is less than 0.65.
Expand the run steps to locate the first action showing an unsuccessful result.

Help me diagnose why my confidence Condition did not execute the review branch.

Test the auto-ready route

The second controlled value sits above the threshold. This run proves that acceptable confidence reaches the automatic route.

Return to the AI Invoice Triage designer.
Select the existing Condition.
Remove 0.60 from the left value field.
Enter 0.90 in the empty left value field.
Select Save.

Before you test again, which branch do you expect a confidence value above 0.65 to execute?

Select Test.
Choose Manually.
Select Test.
Choose the synthetic sample invoice for My invoice.
Select Run flow.
Wait for the manual run to finish.

You'll see green check marks again. This time the False branch contains the executed Compose action.

Expand Condition to inspect its inputs.
Expand the executed Compose action in the False branch.

The Condition inputs show 0.90. The executed Compose output reads AUTO-READY.

Controlled testing is complete. The saved flow now needs its normal confidence source restored.

Live AI mode

Live AI mode normally receives its confidence from the invoice-processing action's dynamic content.

Return to the AI Invoice Triage designer.
Select the existing Condition.
Remove 0.90 from the left value field.
Select the empty left value field.
Select Confidence of invoice total from Dynamic content.
Select Save.

Prototype mode

Prototype mode normally uses the clearly labeled confidence fixture. Its restored value remains below the review threshold.

Return to the AI Invoice Triage designer.
Select the existing Condition.
Remove 0.90 from the left value field.
Enter 0.60 in the empty left value field.
Select Save.

The saved Condition now shows Confidence of invoice total in Live AI mode or 0.60 in Prototype mode.

Normal confidence source not restored?

Use Confidence of invoice total only when your flow follows Live AI mode.
Use 0.60 only when your flow follows Prototype mode.
Select Save after restoring the correct source.

Help me restore the normal confidence source in my AI Invoice Triage flow.

Inspect the run history

The 28-day run history preserves the inputs and outputs from completed runs. These details connect each test value to the branch that executed.

Before you inspect the two runs, what differences do you expect to find in their Condition inputs and Compose outputs?

Return to My flows.
Select AI Invoice Triage.
Select Details.

You'll see the 28-day run history on the flow's Details page. The two latest successful manual runs are your controlled tests.

Select the newer successful run from the 28-day run history.
Expand Manually trigger a flow to inspect its inputs and outputs.
Expand the action that supplies Invoice total confidence to inspect its inputs and outputs.
Expand Condition to inspect its inputs and outputs.
Expand the executed Compose action in the False branch.

The newer Condition inputs show the controlled value 0.90. The executed Compose output reads AUTO-READY.

Return to the AI Invoice Triage Details page.
Select the older controlled test from the 28-day run history.
Expand Manually trigger a flow to inspect its inputs and outputs.
Expand the action that supplies Invoice total confidence to inspect its inputs and outputs.
Expand Condition to inspect its inputs and outputs.
Expand the executed Compose action in the True branch.

The older Condition inputs show the controlled value 0.60. The executed Compose output reads REVIEW REQUIRED.

Can't identify the two controlled runs?

Use the run timestamps to distinguish the newer 0.90 test from the older 0.60 test.
Expand Condition in each run to verify the controlled input.
Expand the executed Compose action to verify the decision output.

Help me find the 0.60 and 0.90 tests in my Power Automate run history.

Strong finish. Your run history now proves that low confidence reaches human review while high confidence reaches the auto-ready route.

Secret mission

Add a Second Review Signal

One confidence score can hide a weak prediction elsewhere on an invoice. Add invoice ID confidence to the gate so either weak signal sends the invoice to human review.

Clean Up Your Resources

Choose whether to keep your Power Automate flow available, turn it off for later, or delete it. Each Live AI invocation can consume organization AI Builder credits or Copilot Credits.

Resources you used:

The AI Invoice Triage instant cloud flow, including its 28-day run-history records.
The local synthetic sample invoice used during the manual tests.

Keep everything running

No action needed. Choose this if you want to demonstrate the flow or continue testing its review logic.

Leave AI Invoice Triage turned on for future manual tests.
Keep the synthetic sample invoice in its current local folder for future runs.
Use Prototype mode for future practice when Live AI capacity is unavailable or unconfirmed.
Remember that the flow runs only when invoked. Each future Live AI mode test can consume organization capacity.

Pause - I'll come back to this later

Turn off the flow to prevent new manual runs while preserving its design. Your run history stays available for later inspection.

Return to My flows in Power Automate.
Open the vertical ellipsis for AI Invoice Triage.
Select Turn off.
Keep the synthetic sample invoice in its current local folder.
Select Turn on from the same menu when you are ready to resume testing.

Delete - I don't want to use this again

Deleting the flow is permanent, so take a moment to make sure you have finished using its run evidence. The synthetic invoice contains test data that you can also remove.

Return to the AI Invoice Triage Details page from earlier.
Select All runs to check for pending work.
Wait for every pending run to finish.
Select Delete on the flow Details page.
Return to My flows.

You should no longer see AI Invoice Triage in your flow list. Its project-specific run history is no longer available from that flow.

Use your Windows file browser to return to the local folder containing the synthetic sample invoice from earlier.
Select the synthetic sample invoice.
Select Delete.

The synthetic sample invoice should no longer appear in that folder.

Nice Work!

You did it. Your AI Invoice Triage flow in Power Automate now turns an uploaded invoice into a visible routing decision.

What you learned:

Built an instant cloud flow with the My invoice file input. Confirmed the visible Invoice received for triage output after a manual run.
Used AI Builder in Live AI mode or transparent Compose fixtures in Prototype mode. Displayed invoice ID, vendor, invoice total, and confidence values without misrepresenting their source.
Converted confidence scores into a review gate at 0.65. Controlled tests produced REVIEW REQUIRED and AUTO-READY. The inspected run history preserved the evidence behind both decisions.
Secret Mission: Extended the review gate with invoice ID confidence. A test with total confidence 0.90 and invoice ID confidence 0.60 still produced REVIEW REQUIRED.

Ready to quiz yourself?