Using Prompts and AI Models in Apps and Flows

Two things to consume, two places to consume them

The last module built a prompt. AI hub also gives you AI models — trained capabilities such as reading a receipt or classifying text. Both can be called from a canvas app and from a cloud flow.

That gives four combinations, and the exam tests all four. Learn them as a grid rather than as a list.

The four ways to consume AI hub capabilities

	In a canvas app	In a cloud flow
Prompt	Add it under Data, then call it as a Power Fx function	Add the Run a prompt action and pick the prompt
AI model	Call the model from the formula bar with Power Fx, or insert an AI Builder component where one exists for that model	Add the model's action, or use the predict action

Prompts in a cloud flow

The action is Run a prompt.

Wiring it up

1. Add the Run a prompt action to the flow.
2. In the Prompt field, choose an existing prompt from the dropdown — or select New custom prompt to author one inline without leaving the designer.
3. Once a prompt is selected, its inputs appear below it. Fill each one with dynamic content from earlier steps in the flow.
4. The action returns output as flow variables, which downstream actions consume like any other dynamic content.

💡 A naming change worth knowing

This action used to be called Create text with GPT using a prompt. It was renamed to Run a prompt in May 2025. Older tutorials, blog posts, and screenshots still use the old name — if a question or a colleague’s documentation mentions it, they mean the same action.

Prompts in a canvas app

In an app, a prompt behaves like a function you can call.

Wiring it up

1. In Power Apps Studio, go to Data → + Add data and add your prompt. It joins the app the same way a table would.
2. Call it from a formula using .Predict() on the prompt name, passing the inputs as arguments.
3. The call returns a result you read a property from.

A button’s OnSelect might read:

Set(result, 'Task identifier'.Predict(TextInput1.Text))

and a label’s Text property then shows result.text.

Question

What are the two steps to use a prompt inside a canvas app?

Click or press Enter to reveal answer

Answer

First add the prompt as a data source under Data → + Add data. Then call it in a formula using .Predict() on the prompt name, passing its inputs as arguments, and read the returned result.

Click to flip back

The important idea is that it is the same prompt. Improve the wording in prompt builder and both the app and the flow pick up the improvement — you do not maintain two copies.

AI models: prebuilt versus custom

Now the other half of the cupboard. AI models live under AI hub → AI models, and they come in two build types.

Prebuilt versus custom AI models

Build type	What you do	Use when
Prebuilt	Nothing — the model is trained and ready to use immediately	The scenario is common across businesses: reading a receipt, detecting a language, scoring sentiment
Custom	Build it, train it on your data, then publish it	The data is unique to your business: your document layouts, your product images, your historical outcomes

The decision rule is short enough to memorise: common problem → prebuilt; your data → custom.

⚠️ Which models are custom

Most models are prebuilt. The ones that require you to build and train are the ones that depend on data only you have:

• Document processing — your document layouts
• Object detection — your products in images
• Prediction — your historical outcomes
• Category classification — available prebuilt and custom
• Entity extraction — available prebuilt and custom

That last pair is the trap. Category classification and entity extraction exist in both forms, so a question can legitimately go either way depending on whether the categories or entities are standard or specific to the business.

Picking the right model for a scenario

Scenario-to-model matching is exam bread and butter. The wording of the scenario gives it away.

Scenario to AI model

The scenario says	Model	Why
Automate expense reports from photos of receipts	Receipt processing	A receipt is a standard document shape across businesses
Pull key clauses and data points out of contracts	Contract processing	Purpose-built for contract documents
Extract fields from our own application form layout	Document processing	The layout is specific to the business, so the model must be trained on it
Work out whether feedback is positive or negative	Sentiment analysis	Sentiment is the literal capability
Sort feedback into topics	Category classification	Classification into categories
Pull product names and dates out of reviews	Entity extraction	Named things inside text
Work out what language a message is in	Language detection	Language identification
Turn support requests into our language	Text translation	Translation
Flag likely fraudulent transactions	Prediction	Learns an outcome from your historical structured data
Count our products on a warehouse shelf photo	Object detection	Trained on images of your specific products
Capture contact details from a business card	Business card reader	Purpose-built for business cards
Read text out of a photo and store it	Text recognition	Optical text extraction

Question

Kite Freight wants to extract fields from its own custom-designed damage report form. Prebuilt or custom model, and which one?

Click or press Enter to reveal answer

Answer

Custom document processing. The form layout is specific to Kite Freight, so the model must be built and trained on samples of that layout. A prebuilt model such as invoice or receipt processing only understands standard document shapes.

Click to flip back

AI models in a cloud flow

Each model exposes actions in the flow designer — for example an action to analyse sentiment, extract entities, or read a receipt. There is also a general predict action that works with many model types.

Then there is the setting that catches people out.

⚠️ The asynchronous pattern trap

For any AI Builder action in a cloud flow, leave the Asynchronous Pattern setting at its default of On. It lives in the action’s settings, not on the action card itself.

Turning it off stops results from the AI model being returned properly. The flow may appear to succeed while producing nothing useful downstream — which is far harder to diagnose than an outright failure. If a scenario describes an AI Builder action that runs without error but returns no usable output, this setting is the first thing to check.

AI models in a canvas app

Two routes, and knowing which model supports which route is the testable part.

Two ways to use an AI model in a canvas app

Route	How	Covers
AI Builder component	Insert a ready-made control from the Insert tab in Power Apps Studio	Business card reader, receipt processor, text recogniser, form processor, object detector
Formula bar	Call the model from Power Fx, like any other function	Sentiment analysis, entity extraction, key phrase extraction, language detection, category classification

Be careful about why the split exists, because this is a favourite trap. The route is model-specific — Microsoft’s wording is “two ways, depending on the model you will be using.” It is not determined by whether your scenario needs a user interface. A purpose-built component exists for only a short, fixed list of models; every other supported model is called through Power Fx, and you build any capture or display experience you need with ordinary controls. Power Fx supports all prebuilt and custom AI models, so “it needs a camera, therefore it must have a component” is exactly backwards.

Components and their build types

• Components backed by prebuilt models: business card reader (canvas and model-driven), receipt processor, text recogniser.
• Components backed by custom models: form processor, object detector — you must build and train the model before the component has anything to call.

Putting it together

Ravi wants drivers to photograph a damaged pallet, have the app describe what it sees, and have a flow decide whether to raise a claim.

💡 How the pieces line up

• The app captures the photo with an ordinary image or camera control and calls the image model from the formula bar with .Predict(). There is no purpose-built AI Builder component for image description — components exist only for a short, fixed list of models, and that is not one of them.
• The flow uses a Run a prompt action to turn the extracted detail plus the shipment record into a plain-English claim summary.
• The prompt takes the damage detail as an input and uses the claims-policy table as knowledge.
• The summary goes to an approval, not straight to the customer — the human review step from the previous module.
• The AI Builder action in that flow keeps Asynchronous Pattern on.

Quick check

Knowledge Check

A maker adds an AI Builder action to a cloud flow. The flow runs successfully but downstream actions receive no output from the model. What should be checked first?

AThe action's Asynchronous Pattern setting, which should be left OnBWhether the flow trigger is set to the correct table scopeCWhether the model has been shared with the flow ownerDWhether the environment has more than one connection reference

Check Answer

Knowledge Check

Which combination correctly describes calling a prompt from a canvas app?

AAdd the prompt under Data, then call it with .Predict() and read the returned resultBInsert the prompt from the Insert tab as an AI Builder componentCAdd a Run a prompt action to the app's OnStart propertyDReference the prompt directly by name without adding it as a data source

Check Answer

Knowledge Check

Kite Freight wants to automatically identify which of its own branded containers appear in warehouse photos. Which model type fits?

AA custom object detection modelBThe prebuilt text recognition modelCThe prebuilt image description modelDA custom prediction model

Check Answer

What to carry forward

• Learn the grid: prompt or AI model × app or flow.
• Flow + prompt = the Run a prompt action, formerly Create text with GPT using a prompt.
• App + prompt = add under Data, then .Predict().
• Flow + model = the model’s action or the predict action — and leave Asynchronous Pattern on.
• App + model = the formula bar with Power Fx is the general-purpose route; an AI Builder component only where one exists for that specific model. The choice follows the model, not whether the scenario “needs an interface.”
• The component list is short and worth knowing: business card reader, receipt processor and text recognizer use prebuilt models; form processor and object detector use custom ones. Anything else, reach for the formula bar.
• Prebuilt for problems every business has; custom for your documents, your images, your history.
• Category classification and entity extraction exist in both build types — read the scenario carefully.

Next you will step away from AI and into the declarative logic that runs whether or not anyone is watching: business rules and business process flows.