2. 2
Get your documents
processed intelligently
Teach your robots to understand documents
using AI-enhanced skills for data extraction
and interpretation.
Drag and drop these capabilities directly into
your automation workflows to embed AI
3. 3
• Like forms, passports, licenses, time
sheets
• Fixed in format and can contain
handwriting, signatures, checkboxes
• Like invoices, receipts, purchase
orders, medical bills, utility bills
• Containing fixed and variable parts
like tables
• Like contracts, agreements, emails,
scripts, drug prescriptions, news
• No fixed format, free-form
sentences/paragraphs
Which documents can be handled by
Document Understanding?
Structured documents Semi-structured documents Unstructured documents
4. 4
Load taxonomy for
your documents
How Document Understanding works
Digitize images using
multiple OCRs
Classify documents Extract named entities
in a taxonomy
Validate and train
supervised models
Export extracted data
2 3 4 5 6
Digitize Classify
Train & Validate
Extract
2 3
5
4
Structured
Semi-structured
Unstructured
Export
RPA
BPM
API
Other systems
6
1
Define taxonomy –
document types and fields
1
5. 5
Taxonomy manager is
used once at the start to
define the collection of
documents that you
would want to process
as well as business rules.
Additionally, you can
describe what data you
would like to extract.
Load taxonomy
6. 6
Move-For-You Co.
We move so you don’t need to move
PO: NP74006735
1 February 2020
PAYABLE WITHIN 15 DAYS OF RECEIPT
20800 ALMADEN AVE, SUITE 404
SAN JOSE, CA 95120-0520
T: +1 425 555 9876
F: +1 425 555 3456
E: billing@moveforyou-co.com
www.moveforyou-co.com
Bill To:
Tony Tzeng
12345 Mango Lane
Seattle, WA 98108
INVOICE DETAILS
Packing services
Storage fees (1 month)
House move (white-glove service)
Vehicle storage and transport
Sales tax 10%
Total Fee including Tax
FEE
$1,282.00
$1,884.00
$5,320.00
$5,186.00
$1,367.20
$15,039.20
Methods of payment
Personal Check: Move-For-You LLC
Wire Transfer: BigBank Co., Account 123456789-0987ABC
Invoice No: 456200-TZE1
Digitize text in the documents
using OCR
7. 7
Classify and split the documents
Documents scanned into one file
isn’t a problem – owing to
classifiers, the robot can identify
the document types and split the file
to process the documents
accordingly.
Document Understanding offers
different classification capabilities
ranging from keyword-based to
ML-based classification.
8. 8
Validate classification of the
documents
Classification Station is
used to check, correct, and
confirm the results of
document classification and
splitting.
9. 9
You can easily configure
data extraction to choose
most suitable extractor
for each field.
Use a combination of rule-
based and model-based
approaches to ensure
smooth and accurate
processing of different
documents.
Extract data from the documents
10. 10
Data extraction – from rules to AI
to a hybrid approach
RegEx-Based Extractor
Structured fields
Machine Learning Extractor
Mostly semi-structured documents
Form Extractor
Structured documents (forms)
A combination of both –
rule-based and AI-based approaches
Mostly documents combining both structured and less structured formats
Rule-based
Forms AI
Structured documents (forms)
AI-based
Hybrid approach
11. 11
Make use of pre-trained ML
models to process invoices,
receipts, utility bills, ID cards, and
many more document types.
Retrain the models to optimize
them for your custom documents
and improve the model accuracy
over time!
Bring your own model or third
party models and incorporate
them in your automations.
Pre-trained ML models
12. 12
Validate the extracted
Information and handle
exceptions using
Validation Station.
Now, ML models can also
be retrained using the data
confirmed or corrected in
the Validation Station.
Validate extraction of the
documents
13. 13
Let the classifiers and
extractors learn from the
data corrected and
validated in the
Classification Station and
Validation Station
respectively.
Learn about sharing data for model retraining here.
Train classifiers and extractors
14. 14
Export the extracted data
End-to-end intelligent
document processing
Start & continue the document
processing workflows with other
automation components. Export the data
for further usage/automation – for
example, to an Excel spreadsheet, send
as email, SAP, and so on.
16. 16
Logic in the process to automatically determine if field(s) extracted correctly
Art of the possible with business rules
Character
validation
Basic math
Date validation
External source
verification
Invoice # = 7 digits
Total = Line Items + Tax
Valid date format
Verify against PO Record
PO
Number
Vendor ID Vendor
Name
5928452 12345 PROTECH
3758292 98734 LORANA
17. How to get started
Define which documents you want to process
Explore how UiPath Document Understanding can automate this
Give it a try – start Enterprise Trial
Get trained on Document Understanding in UiPath Academy
or via instructor-led training from UiPath
Enterprise Trial, Academy Course, and more at
uipath.com/document-understanding
19. 19
Document Understanding: AI & ML improvements
More use cases and faster
deployment
• Automate more processes out of the box
with pre-trained & retrainable ML models
• Pay slips & personal earnings
statements
• Certificates of origin
• EU declarations of conformity
• Children’s product certificates
• Certificates of incorporation
• Shipping invoices
• CMS1500
• Train ML models quickly with one-click
ML for document classification and data
extraction (in public preview)
20. 20
Document Understanding: business rules
More accurate extraction
with custom business rules
• Set up your custom business rules,
including mathematical rules, to
reduce the need for human validation
and ensure higher accuracy & straight-
through processing rates
21. 21
Document Understanding as a service
Seamless integration with
third-party tools
• Easily integrate core Document
Understanding capabilities with any
third-party tools via API (in public
preview):
• Discovery
• Digitization
• Classification
• Extraction
• Validation
22. 22
Roadmap: 2023.10 and beyond
2
Accelerate time
to value
• AI-assisted active learning-
based training (tag the
minimum, real-time
retraining)
• Pre-labeling documents with
GPT
• GPT-enabled document
classification and data
extraction
1
Expand unstructured
document intelligence
• GPT for querying documents
with natural language
• Enhanced Communications
Mining + Document
Understanding integration
4
Deployment insights
& operations
• Dashboards & document
audit, containing metrics for
STP rate, time saved, and
more
• Validation connected with
Integration Services to enable
data lookups in external
applications
3
• Improved UX to help users
build better models faster
• Enhanced capability
discovery, guided labeling
experience and improved
model insights
User
experience
23. 23
Example AI use cases
Professional Svc
Data Extraction
from Charts
RFP Opportunities
Classification
Deal Guidance
Financial Svc
Fraud
Detection
Personal Loan
Approval
KYC – Entity
Identification
Retail
Packaging Quality
Evaluation
Inventory
Management
Merchandising
Planning
Healthcare
Real Time Pregnancy
Risk Evaluation
Patient Receivables
Management
Propensity of Claim
Denial Prediction
General
Help Desk
Answers
Customer Churn
Prediction
Resume Matching
Auditing – Anomaly
Detection
AML Alert
Classification
Product
Recommendation
Fraudulent Medical
Claim Prediction
Customer Complaints
- Email Classification
Legal – Win/loss rate
prediction
ID Information
Extraction
Pricing Optimization
Readmission
Prediction
Quality - Visual
Inspection
*Review appendix for details
25. 25
Use cases
Real world use cases,
and customer quotes
Product deep dive
AI Center overview,
deployment types, Email
AI, release highlights,
and roadmap
Next steps
How to get started, and
what to expect
Intro
What’s AI Center, ML
models, demo, use cases,
and key differentiators
Agenda
26. 26
We’ve built AI into every part of the
UiPath platform
Classify emails
Forms in VDI
Semi-structured
data
Unstructured
data
Mine processes
Mine tasks
Speech to text
Chatbots
Understand
documents
AI
Copy/paste
Email
event
Fill forms
Extract
structured
data
Log in
to apps
Capture
tasks
Scrape
data
Read/write to
databases
Activities
Moving
file/folders
RPA
Process
Mining
Task
Mining
Document
Understanding
AI Computer
Vision
Custom/
Ecosystem
Skills
Action
Center
Chatbots
AI Center Insights
BUILD
Teach your robots new
AI skills
DISCOVER
Scientifically identify
automation opportunities
MANAGE
Deploy, scale, and
manage AI
ENGAGE
Bring humans in the
AI loop
27. 27
AI Center enables you to insert AI into your workflow with drag-and-drop ease
1 2 3
ML Skill
(Humans) Validate exceptions
Choose a model
from these options:
Choose from over 25
pre-built models
From UiPath
Retrain the model
Select from dozens of
UiPath pre-build models
Pick a model
From UiPath marketplace
Build an ML Skill with
UiPath AI Center
Drag and drop ML Skill into
RPA workflows in UiPath Studio
Bring your own
29. 29
AI Center key differentiators
Bring your own
models
Models from UiPath
and partners
Multiple deployment
options
Continuously
improve ML models
with retraining
Deep integration with
other UiPath
products
Human validation
31. 31
Deployment Infrastructure Why? Why Not?
Main
Prerequisites
Installation
Experience
Out-of-the-box Model
Delivery
UiPath Automation
Cloud
SaaS (Infrastructure managed
by UiPath)
• Fully managed by UiPath. No
additional infrastructure needed
• Scaling all handled by UiPath
• Always get latest features first
• Can work with Robots connected
to on-prem Orchestrator
• Data Sensitivity
Simply an internet
connection
Sign-up and get started in
seconds
Automatically available and always up-to-date
UiPath Automation
Suite
Service Fabric deployed onto
Customer Infrastructure
• AI Center comes bundled with
entire UiPath suite
• Products come pre-integrated
• Want to maintain
existing Orchestrator
deployment on
Windows
• Don’t need most
UiPath products
• Very locked down
Linux servers (e.g.
STIG-hardened)
• One or more (depending
on scale) Linux VM
• Data disk(s)
• SQL Server
EITHER:
• Input values into
inputs.json file and
run setup.sh from
Linux Shell OR
• Input values via CLI
for interactive install
in Linux Shell
CONNECTED MODE:
• Daily CRON jobs ensure latest models are
always available
AIRGAPPED MODE:
• One time image bundle must be run via
Linux Shell for base images.
• Then, each model package is uploaded
via UI as Models packaged as OOTB
model.
AI Center Standalone
Service Fabric deployed onto
Customer Infrastructure
PLEASE NOTE: This is
different from the previous
version of AI Center on
Replicated, as we have
standardized on the Rancher-
based infrastructure provided
by Automation Suite)
• Want to maintain existing
Orchestrator deployment on
Windows
• Can deploy to on-prem data
center or your own cloud
subscription
• Don’t have existing Kubernetes
footprint
• Very locked down
Linux servers (e.g.
STIG-hardened)
• One or more (depending
on scale) Linux VM
• Data disk(s)
• SQL Server
• Orchestrator 20.10 or
higher
Input values into
inputs.json file and run
setup.sh from Linux Shell
CONNECTED MODE:
• Daily CRON jobs ensure latest models are
always available
AIRGAPPED MODE:
• One time image bundle must be run via
Linux Shell for base images.
• Then, each model package is uploaded
via UI as Models packaged as OOTB
model.
AI Center Deployment Types
• Please visit AI Center documentation pages for more details on this topic.
33. 33
Use any ML models you choose—we are open
Deploy Consume Manage Improve
Bring your own ML model. Or
choose a pre-built model from
UiPath or UiPath partners
AI Center offers dozens of ML
models, enabling hundreds of use
cases
Jump start your first AI use case
with a model without having to
build them
Easily deploy the model as an ML
skill
34. 34
Drag and drop the ML model into your automation
workflow
Deploy Consume Manage Improve
Drag and drop the ML activity to an
RPA workflow
Select an ML skill from the drop-
down list
Use multiple ML skills in one
workflow if needed
Easily test the ML skills before you
run
Scale the ML skill to as many
robots as you want
35. 35
Monitor and manage your model
Deploy Consume Manage Improve
View all datasets, ML packages,
and pipelines on the Dashboard
page
Get end-to-end visibility on ML
model use
Update model versions with a
few clicks
Keep track of your data, model
performance, user actions, and
ML pipelines
36. 36
Continuously improve your model
Deploy Consume Manage Improve
Label data, and configure custom fields
Train a pre-built model with your own data
Set confidence threshold
Handle exceptions, validate predictions, route
validated data back for retraining
Choose model version, package version and
dataset
Choose when to run the training pipeline—
now, time-based, recurring
Get accuracy score, and training run report
(data statistics, evaluation statistics, confusion matrix, per class
statistics)
37. 37
Deploy ML Skills as REST API
Use ML skill from UiPath Studio
Use ML skill from a different tenant
Use ML skill from other UiPath
products or from third-party apps
Use Orchestrator on-prem and AI
Center in cloud
Four different ways of using ML skills
38. 38
How to use UiPath ML models
Model
training &
evaluation
Model
deployment
Model
maintain &
retraining
Data
preparation
& cleaning
Prepare high-quality data
Your training data should be
as representative as of your
production data
Label your data correctly
Mislabeled data result in poor
model performance
Distribute evenly
Make sure your dataset is
evenly balanced across
your classes
Separate your data
Split your dataset into
training and test data
Train on base version
When retraining, use a bigger
dataset and retrain on the base
version
Work in iterations
Train and evaluate the model
on a small dataset. Reiterete
until you are happy with the
performance.
Use GPU for speed
GPU can be used for serving
and training. GPU delivers 5-
10x improvement in speed
Maintain model quality
Monitor model performance,
retrain the model periodically
ML Package version
After each training, the
knowledge is stored in the
next minor version
Bring human in the loop
Bring humans to validate
model predictions, and save
validated data for retraining
40. 40
AI Enables Automation of Processes That Include
Uncertainty
You cannot determine an outcome
with 100% certainty.
High Variability
There is too much variability for
rules based.
Unstructured Data
Information like articles, documents,
images, videos and emails.
Property
Valuation
Loan
Defaults
Inventory
Forecasts
Resume
Matching
Purchase
Decisions
Language
Translation
Invoice
Extraction
Email
Routing
Speech to
Text
41. 41
ML models and use case examples
Image Classification Semantic Similarity
Semantic
Similarity
Answer
questions
vis email
based on
FAQ docs
Fetch
articles
based on
emails or
search
queries
Plagiarism
detection
on forms
Detect
intent
Image
Classification
Vehicles
Clothes
Damaged
products
Logos
This model shows which sentences
or words correlate with each other.
This model classified images into
different classes.
This model identifies key elements
from text, then classifies them into
pre-defined categories.
Named
Entity
Recognition
Emails
Letters
Webpages
Call
transcripts
Named Entity Recognition
42. 42
Real-world Email AI use cases
Top US Bank
• Millions of emails, with different SLA
requirements
• With Email AI, many hours are removed.
People can now focus on what they need to do
instead of sorting, and pre-filling info from one
system to another
Financial Services company
• Emails by business customers
• Email AI extracts income data from those emails
Printing, envelope and label provider
• Orders to print labels
• Email AI extracts info like label color, height,
weight, quantity, materials, and other info in
emails
• Orders for different financial products, placed
by customers via email
• Email AI extracts information from emails,
prioritizes based on business value, and auto-
fill their OMS
Large investment management org
Metal solutions provider
• Spot RFQs and offers to buy excess materials
• Email AI helps determine what’s worth
spending time on, separating real opportunities
from the junk
Financial services company in Japan
• Notes details sent by issuers
• With Email AI, they extract info from the
emails, auto-fill Bloomberg terminal, create
quotes for issuers
Webinar link
44. 44
Roadmap to automating more processes
Take AI Center
Academy course
Sign up for trial
Get UiPath
support
Prove ROI and
scale up
Enterprise Cloud Trial
Enterprise On-premises Trial
• Upskill your RPA developers
• One of the most popular
UiPath Academy courses in
the past two years
• All ML models included
• Cloud and on-premises
• Identify a use case – choose
‘low-hanging fruit’ projects
rather than ‘moon shot’
projects
• Watch AI use case demo
videos
• POC
• AI Starter Pack
• Immersion Lab
• AI Center training course
• AI playbook
• AI 101 webinar
• Instructor-led training
• Professional Services
• Example metrics: cost
reduction, quicker customer
response time, faster cash
collection
• Apply AI in more regions,
more business units
• Write a playbook and
standardize the process
45. 45
AI Center Roadmap 2023
Enhance automations with ML
Feature Enhancements
• Data Labeling – Upload raw text data and annotate in the labelling tool for free text, and consume
labeled data in training models
• Tenant-level AI Units – Track and set limits of AI Unit consumption at the tenant-level
ML Packages and Skills
• Faster model training and deployment times
• Test ML skills on AI Center before consuming in an automation workflow
• AI Center API Availability – Enables programmatic management of ML Skills and Pipelines
• Usability improvements including outage alerts, search / filter options, tooltips and accessibility
Infrastructure, Security & Recovery
• At-rest Data Encryption – Encrypt data with your own key
• Tenant migration – enable AI Center projects to be migrated within and across geos
• Resource queues for on-prem hardware
• One-click access to logs and audit information
Get more done with a
digital workforce that
seamlessly collaborates
with your people and
automates work via UI
and API, powered with
native integrated AI
Automate
46. 46
Improved experience for pipeline
parameter settings
Auto-generated parameters for
ML pipelines provides an easier
user experience without coding.
With this feature, users no
longer have to remember or
search for the parameters for a
model anymore.
Notes de l'éditeur
Taxonomy Manager
OCR engines
Keyword-based Classifier
Intelligent Keyword Classifier
ML Classifier
Classification Station
RegEx-based Extractor
Form Extractor
Intelligent Form Extractor
ML Extractor
Invoices, Receipts – in GA
Purchase Orders, Utility Bills, Invoices for Australia, India, Japan – in preview
ML Extractor
AI Center
Validation Station
Let’s build momentum here highlighting new capabilities for classification & splitting
Two classification methods available already: a Keyword Based Classifier, focused on learning TITLES and searching for them, with no capability to split files into multiple docs; and the newly launched Intelligent Keyword Classifier, focused on document similarity from a word content perspective, also with learning capabilities, and also capable of identifying and suggesting splits between different documents within the same file.
2min 13 sec
Vendor Confirmation
Name
Address
Two and Three-Way Matching
Invoice Item Quantity
Invoice Item Price
Cost Center Validation
Date Format
Sanity Check ID #’s
Screenshot – new model for Shipping Invoices
2min 30 sec
LLM – large language models
We are seeing use cases across all industries. Refer to the Appendix to get deep dives on several of these.
Start with traditional RPA – The core of RPA is to mimic human actions in interacting with digital system. Traditional RPA is a process-driven technology to automate repetitive, rule-based and non-value added task in business processes. Typical tasks would include auto-keying, screen/form integration, application or data integration, automated decisions acting on structured data. Focus is on ‘execution’ of tasks.
When you add AI – By adding RPA to AI, you can automate more tasks that are knowledge-intensive. You can address scenarios that are predictive scenarios like whether a customer will default on their loans and ones that involve high-variability like matching resume job skills to open roles in the company.
AI can:
Understand unstructured and semi-structured data
Detect fraud
Classify and summarize text
Read sentiment
Detect objects in images and do many more
Discover – one of the reasons automation programs don't scale to where they need to is due to the challenge of finding the next automation. This is where AI can be a natural way of discovering the next automation opportunity
Gartner in ‘Emerging Technologies: RPA software advancements’
RPA scripts with integrated ML models allow buyers to automate more complex use cases and improve the scalability of their automation programs.
80%: https://www.i-search.com.cn/CompetitiveLandscape-RPA-Software.pdf
74%: https://www2.deloitte.com/us/en/insights/focus/cognitive-technologies/state-of-ai-and-intelligent-automation-in-business-survey.html
65%: https://www2.deloitte.com/us/en/insights/focus/cognitive-technologies/ai-adoption-in-the-workforce.html
AI Center is UiPath's machine learning platform, which enables customers to deploy, consume, manage and improve machine learning models developed inhouse, by UiPath, or by UiPath partners.