Slide deck of my DyanmicsPower! Brussels session.
Session description:
AI Builder is here!
The Citizen Developer style tool to easily add Artificial Intelligence to your PowerApps and Flows.
In this session we will take a deep dive into this great new tool in the Power Platform stack. How to build your own models and how to embed them in your apps and flows.
We will look at the what is possible, but also what are the limits and boundaries you need to take in consideration.
We will cover best practices and tips and tricks.
In a nutshell: everything you need to know to get started yourself.
8. But first!
Let’s generate some demo data!
• Tweet something using #AIBuilder
• Use a language you like (Latin script)
9.
10. Why do I like the
Power Platform
tools
– including AI
Builder?
Evolved to become a Consultant at a
Microsoft Partner
Landed by accident in the
IT-department of an end customer
Started out as business user
12. GA on October 1st
incl. Dynamics 365 Sales Enterprise integration
Key changes • AI Builder add-on and trial offers
• AI capacity management in PPAC
• GA scenarios are solution aware
• Scalability and reliability of GA scenarios
GA scenarios now require trial
or paid capacity
GA Preview
New!
New!
New!
New!
15. Step by Step Guidance in the UI
ADD YOUR DATA TRAIN THE
MODEL
CHECK THE
PERFORMANCE
PUBLISH & USE
16. Prediction
Use your database and this AI model to determine the
likelihood of specific business outcomes.
Form Processing (preview)
Read, extract, and process data from scans, emails, PDFs, and
images.
Object Detection (preview)
Build an AI model that recognizes and counts visual objects.
Text Classification (preview)
Build an application that reviews, tags, and classifies large
volumes of text to track sentiment, improve customer
experiences, and mine valuable insights.
Custom models
17. Prediction - GA
Determine the likelihood of specific business outcomes – Boolean (Yes/No – True/False)
Input: Data in 1 CDS entity
Output: prediction written back into prediction field in CDS
Use automatic schedule in CDS and in Power Automate
Specifics: Uses CDS Entity
Only 2 outcome possibilities (Yet!)
Data requirements: Minimum of 50 records in total and 10 records historical outcome for each class.
Best result with at least 1000 records.
18. Coming
Q4 CY19
• Related entities
• Filtering of
training data set
• Scheduling
• Predict two or
more outcomes
• Predict a number
Q1 CY20
• Power Apps
component to
explain
prediction
output
Ideas
• Forecasting time
series
• Data lake
support
19. Be aware of
• Data quality – no empty fields / faulty data
• Data quantity – enough data with options that
reflects the expected outcome
• Unwanted bias – unselect irrelevant or misleading
fields
• Accuracy score between 50 – 100%
BUT! You are the one to determine if the
accuracy makes sense!
20. Form processing - Preview
Read, extract, and process data from scans, emails, PDFs, and images.
Output: Extracted key-value pairs and table data.
Use in Power Apps Canvas & Power Automate
Specifics: JPG, png or PDF – max 4 MB – Latin alphabet
High quality images – between 530 x 100 and 4200 x 4200 pixels
Data requirements: 5 – 10 sample documents with the same layout.
Must contain keys above or to the left of the value.
22. Coming
Q4 CY19
• Support for larger
file sizes
• Support for TIFF
files
Q1 CY20
• Manually tag fields
not automatically
detected
Ideas
• Combo box
detection
• Specialized model
(e.g. invoices)
• Multiple document
types in one model
23. Object detection - Preview
Build an AI model that recognizes and counts visual objects.
Output: Name and count of detected objects
Use in Power Apps Canvas & Power Automate
Specifics: JPG, png or BMP – max 6 MB
Data requirements: minimum of 15 images per object – preferably 50 or more.
25. Text classification- Preview
Build an application that reviews, tags, and classifies large volumes of text to track sentiment,
improve customer experiences, and mine valuable insights.
Output: List of tags with certainty score in separate entity
Auto generated on schedule in CDS & used in Power Automate
Specifics: Uses CDS Entity
Languages: English, French, German, Dutch, Italian, Spanish, Portuguese
Data requirements: Text and tags in text fields in same entity – All tags in one field using a delimeter
Max 5000 characters per text item - Minimum of 50 text items per tag.
26. Coming
Q4 CY19
• Prebuilt models
(e.g. Key Phrase,
Sentiment,
Language)
Q1 CY20
• Improved data labelling
and correction
experience
• Prebuilt models (e.g.
Entity, Intent)
• Power Apps and Power
Automate integration
• Solution import and
export
• GA
Ideas
• Multi-entity and
additional sources
• Filtering
• Continuous learning and
feedback
• New scenarios
27. Tips for your
training data
• No high rate of missing values
• Check for typos
• Check for inconsistent capitalizations
• Check for inconsistent or incorrect labels
Data quality
• Remove duplicate or irrelevant fields
• Remove fields with high correlation to the outcome
Unwanted bias
At least 100 records per tag
Balanced use of tags in your data
Training data similar to actual text
28. Business card reader
Pull contact information from business cards, store this information in
your database, and import it into your CRM system.
Sentiment analysis (preview)
Detect positive, negative, neutral and mixed sentiment in social
media, customer reviews or any text data you want to analyze.
Key phrase extraction (preview)
Extract the main points and key phrases in text documents.
Language detection (preview)
Use this model to identify the predominant language of a text
document.
Text recognition (preview)
Extract embedded printed and handwritten text from documents and
images into machine-readable character streams.
Prebuilt models
29. Coming
Q4 CY19
• Prebuilt models
on Build page
• Power Apps OCR
component
Q1 CY20
• Power Automate
templates
• Text recognition
app templates
• Text recognition v2
• GA
Ideas
• Receipts
• Speech
34. Update custom model
Create new version
• From published version
• Or
• From last trained version
Change settings or training
data
Publish
or
• Only for:
• Prediction model
• Text classification model
Retrain
model
37. Roles & security → CDS security roles
ENVIRONMENT MAKERS
CREATE AI MODELS
CDS USERS
USE MODELS
SYTEM ADMINS & CUSTOMIZERS
ACCESS ALL MODELS
38. AI Builder License
AI Builder Capacity Add-on
$ 500 / subscription
1 million service credits
Tenant level
Credit consumption
Training models
Using models
Credit calculation ???? – Calculator not available yet
Credit allocation
Allocate capacity to environments in
Power Platform admin center