[object Object]

An ops team gets 200 supplier invoices a week and three people spend mornings keying them into BC. AI Builder’s invoice processing model handles 85 percent of those invoices with no training and the remaining 15 percent with a clear exception queue. The path from idea to working pilot is days, not months. The catch is capacity planning; AI Builder credits run out faster than most teams expect.

Pre-Built Models

Out-of-box: text classification, sentiment, entity extraction, document processing (forms, invoices, receipts), object detection. Zero training needed for many use cases. The pre-built models are tuned on broad data and work well for standard scenarios; they struggle on industry-specific document layouts or jargon-heavy text.

Pre-built model fits:
- Standard invoices and receipts
- General sentiment classification
- Common entity extraction (names, dates, amounts)
- Object detection for everyday objects

Test the pre-built first. If accuracy is above 80 percent on a 50-sample test, ship it. If below, escalate to a custom model.

Custom Models

Train on your own data: custom document types, custom classification. Requires labeled training data; 50 to 200 samples per class to get decent accuracy. The training experience is web-based; the maker uploads samples, labels them in the UI, and triggers training that runs server-side.

Custom model lifecycle:
1. Collect 50 to 200 representative samples
2. Label in the AI Builder studio
3. Train (server-side, takes minutes to hours)
4. Quick test with held-back samples
5. Publish to flows or apps
6. Monitor accuracy in production
7. Retrain periodically as data drifts

The retrain step is the one teams skip. Models drift as the underlying data shifts; quarterly retraining is a reasonable default.

Capacity

AI Builder consumption is metered in credits. Monitor usage per maker. Forecast before broad rollout or you will hit limits mid-month. Credits are consumed per prediction, with prices varying by model type. Document processing costs more per call than text classification.

Capacity planning rule of thumb:
- Estimate calls per day per use case
- Multiply by per-call credit cost
- Add 30% buffer for peak loads
- Compare to monthly credit allocation

A pilot that uses 2000 invoices a month consumes about 2000 document processing credits, which fits in a starter capacity. Scale to 50000 invoices and you need a meaningfully larger capacity allocation.

Integration

AI Builder actions plug into Power Automate and Power Apps. No glue code needed. Action input to model to structured output to downstream logic. The action signature is a typed input (file, text) and a typed output (JSON), which means downstream actions can pick fields by name.

// Power Automate action result
{
  "vendorName": "Contoso",
  "invoiceNumber": "INV-12345",
  "totalAmount": 4567.89,
  "invoiceDate": "2026-04-15",
  "lineItems": [...]
}

Map these fields to your downstream system in the next action. The flow handles the type conversion.

When It Fits

Invoice processing, contract analysis, feedback sentiment, simple vision. When the job is “classify or extract from a document”, AI Builder is usually faster than rolling custom. The crossover point is roughly: if the job is well-defined and you have less than 10 days of bandwidth, AI Builder beats custom; if the job is novel and the volume justifies a multi-week build, custom wins.

AI Builder fits when:
- The job is classification or extraction
- The output schema is straightforward
- The volume justifies the credit cost
- The accuracy target is achievable with off-the-shelf models

Confidence Thresholds

Every AI Builder prediction returns a confidence score. Use it to gate the autonomous-versus-review decision. Above 0.90, process automatically. Below 0.90, queue for human review. The exact threshold depends on the cost of a wrong answer in your business.

threshold 0.90: 70% auto, 0.5% rework
threshold 0.85: 80% auto, 1.2% rework
threshold 0.80: 88% auto, 2.5% rework

Tune weekly for the first month, then settle on the value that balances throughput and rework cost.

Governance and Sharing

Custom models are owned by a maker and shared with users or teams. Without explicit governance, a model owned by a departed employee becomes orphaned. Assign every production model to a service account and add a backup owner.

What to do this week

Pick one document type, run the pre-built model against 50 samples, and measure accuracy. If it clears 80 percent, build the flow and pilot. Forecast credit consumption for the next quarter and confirm the capacity covers the rollout.

[object Object]
Share