The low-code AI stack
The Power Platform lets people build working automation and applications without writing much code. When you add AI Builder to the mix, those low-code makers can also classify documents, extract fields from invoices, and call a prompt — all inside flows and apps they already know how to build.
This matters because the people closest to a process are often the ones who understand it best. Power Automate, Power Apps, and AI Builder put practical AI within their reach. The job of the Deploy practitioner is to make that power safe, governed, and reliable — not to gatekeep it away.
By the end of this unit- Describe how Power Automate, Power Apps, and AI Builder fit together as a low-code AI stack.
- Choose the right AI Builder model type for a given task.
- Design a document-processing and approval flow that combines extraction, human review, and routing.
Three components, one platform
- The workflow engine — triggers, actions, conditions, and approvals
- Cloud flows for automation, desktop flows for legacy UI automation (RPA)
- Where most document and approval automation lives
- Calls AI Builder models as ordinary actions
- The interface layer — canvas and model-driven apps
- Gives people a screen to review extracted data, approve, or correct
- Reads and writes to Dataverse and connected systems
- Can call AI Builder models directly from a screen
- The AI layer — prebuilt and custom models
- Document processing, category classification, prediction, and prompts
- Consumed inside flows and apps without data-science skills
- Runs on Microsoft-hosted models; consumption is metered by credits
AI Builder models
AI Builder offers two families of models: prebuilt models that work immediately, and custom models you train on your own data. Choosing correctly saves a great deal of effort.
Prebuilt models — use first
Invoice processing, receipt processing, identity-document reading, text recognition, key-phrase extraction, sentiment analysis, and language detection all work without training. If a prebuilt model covers your case, you can be in production the same day.
Custom document processing — when layouts are yours
If you process a bespoke form — a particular order sheet or claim layout — train a custom document-processing model on a sample of your documents. It learns the fields and tables specific to your format. Expect to provide and tag several example documents.
Custom prediction and classification
For structured outcomes — will this lead convert, which category does this case belong to — a custom prediction or category-classification model trained on historical Dataverse data can score new records inside a flow.
Prompts — generative tasks
AI Builder prompts let you call a generative model with your own instruction to summarise, draft, or reformat text inside a flow. Useful for the generative steps in a process, with the usual need to review output before it is used.
AI Builder consumption is metered in credits, allocated per environment. A flow that calls a model on every record at high volume can exhaust an allocation quickly. Estimate volume against your credit capacity before going live, and monitor consumption afterwards.
Document processing patterns
Document processing is the most common, highest-value low-code AI pattern: turn a stream of incoming documents into structured data with a human in the loop where confidence is low. The reliable pattern follows a consistent shape.
1. Trigger on arrival
A new file in a SharePoint library, an email attachment, or a form submission starts the flow. Capture the source and a unique identifier immediately so the document can be traced.
2. Extract with AI Builder
Call the prebuilt invoice model — or your custom document model — to pull the fields you need. Each field comes back with a confidence value. Do not ignore confidence; it is the basis of the next step.
3. Branch on confidence
If all key fields exceed your confidence threshold, proceed automatically. If any fall below it, route the document to a person for review in a Power App or an approval. Straight-through processing for the easy cases; human attention only where it adds value.
4. Write back and audit
Write the confirmed data to the system of record, store the original document, and log what happened — extracted values, confidence, who reviewed it. The audit trail is part of the deliverable, not an extra.
For a document type your organisation handles, what confidence threshold would you set before allowing straight-through processing? What is the cost of a wrong field slipping through versus the cost of routing too much to people? That trade-off is the heart of the design.
Your invoice extraction flow returns a supplier name with low confidence but a total amount with high confidence. What is the appropriate design response?
Approvals and orchestration
Most business processes contain a decision point that a person must own. Power Automate's approvals turn that decision into a tracked, auditable step — and combined with AI extraction they let the machine do the gathering while the person does the deciding.
Give approvers the context, not just the question
Design for the rejection path
Keep a human accountable for AI-assisted decisions
Time-outs and escalation
Build the practical skills with these paths:
AI Builder: an Introduction ↗
Get started with Power Apps ↗
Automate a business process using Power Automate (Applied Skill) ↗
End of Unit 12
You should now be able to:
- Position Power Automate, Power Apps, and AI Builder as a coherent low-code AI stack.
- Select the right AI Builder model type and account for credit consumption.
- Design a confidence-gated document-and-approval flow with a clear rejection path.
Unit review
You need to extract fields from a standard supplier invoice with no special layout. Which AI Builder option is the fastest route to production?
Why is the confidence value returned by an extraction model important?
What is the most common mistake when building an approval flow?
Where should accountability sit in an AI-assisted approval process?
End of module
You have completed Course 12: Power Platform and AI. Next: Data Pipelines for AI — building reliable ingestion and transformation pipelines for AI workloads.