Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
Process Automation for Unstructured Content
2 | Copyright © 2019 Indico
• Boston-based Enterprise AI
solution for process automation
founded 2013
• Unique deep-learning based
technology solving key barriers to
enterprise AI adoption
• Focus on “understanding” text,
image, documents and other
unstructured content
• Named to 2018 Gartner Cool
Vendors in September
About Indico
3 | Copyright © 2019 Indico
Agenda
What’s the Problem?
Traditional Approaches & Challenges
Introduction to AI for unstructured content
Re-Imagining document workflows
Conclusions
4 | Copyright © 2019 Indico
The Unstructured Content Challenge
Unstructured content in the
enterprise represents
unrealized opportunity &
unquantified risk…
5 | Copyright © 2019 Indico
The Trouble with Text
Context, not keywords, drives meaning
6 | Copyright © 2019 Indico
And then there’s this…
7 | Copyright © 2019 Indico
How it usually gets “solved”…
ManualHighly manual processes
involving many people
InconsistentIndividual understanding of
process rather than institutional
understanding
ExpensiveExcessive resources and
financial risk embedded
in process
8 | Copyright © 2019 Indico
The Goal
Decreased Cycle TimesEnhance internal and external
stakeholder experience
Redeployed ResourcesRedeploy critical resources to
higher-value opportunities
Increased EfficiencyDrive throughput and capture profitable
growth opportunities
8 | Copyright © 2019 Indico - Confidential
Decreased RiskConsistent & auditable workflows
with comprehensive “explainability”
9 | Copyright © 2019 Indico
Automation is Entering its Third Wave
Business Process
Management
Robotic Process
Automation
Intelligent Process
Automation
INPUT Structured Data Structured Data Unstructured Data
AUTOMATION
APPROACHProcess Models Task Models
Algorithmic/
Deep Learning Models
OUTPUT Deterministic Deterministic Probabilistic
VALUE
PROPOSITIONProcess Efficiency Digital Labor Arbitrage
Transparent and
scalable decisioning
10 | Copyright © 2019 Indico
AI & The “Automation Imperative”
10
“Across regions and industries, the survey results suggest that automating businesses is a global phenomenon.”
-
“The Automation Imperative”Sep 2018
AI Doubles the Success Rate of Automation Initiatives
11 | Copyright © 2019 Indico
Traditional Approaches
• Intense, time-consuming and
expensive manual effort
• Requires capturing all edge cases
• Brittle, requires endless
maintenance
Rule Engines
12 | Copyright © 2019 Indico
Traditional Approaches
• Assumes forms are
homogeneous
• Variances in layout
breaks OCR rules
• Intensive manual
setup effort
• Inability to handle
scanned PDFsOCR
13 | Copyright © 2019 Indico
Traditional Approaches
• Rigid representation of
language-parts of speech,
entities etc
• Requires extensive
dictionary-almost always a
pre-trained approach
• Must be endlessly “tuned”
to account for anomalies is
the data NLP
14 | Copyright © 2019 Indico
New Approach: AI?
Huge
Training Sets
Expensive
Hardware
Data Science
Expertise
Data
Leakage
15 | Copyright © 2019 Indico
Some Definitions
Machine Learning –
A field of computer science
that focuses on “teaching”
machines to make
decisions and
determinations based on
data rather than relying on
explicit programming
16 | Copyright © 2019 Indico
Some Definitions
Deep Learning –
A set of machine learning
algorithms that have become
increasing popular in recent
years due to their near-human
levels of performance for tasks
involved unstructured data –
primarily text, image, and audio
data.
17 | Copyright © 2019 Indico
Some Definitions
Transfer Learning –
A deep learning method
where a model
developed for a task is
reused as the starting
point for a model on a
second task.
18 | Copyright © 2019 Indico
We have had to understand the computer…
19 | Copyright © 2019 Indico
The computer learns to understand us…
20 | Copyright © 2019 Indico
Keys to Success: Labeled Data
1. Data
• Data prep is single most
important aspect of ML
• Labeled data that
sufficiently captures the
desired outcome is critical
21 | Copyright © 2019 Indico
Keys to Success: The SME
2. Expertise
• Subject matter experts
who understand the
business problem
• Some level of data
science understanding to
interpret and refine
approach
22 | Copyright © 2019 Indico
Keys to Success: Defining the Outcome
3. Definition of Success & ROI
Hypothesis
• A defined outcome (with consensus
across stakeholders) connected to a
tangible business benefit
• Ability to define the costs of the
current process- both hard costs and
opportunity costs
• Patience- several iterations will be
required. Scientific method.
23 | Copyright © 2019 Indico
Example AI-Based Approach to Unstructured Content
DATA LABELING
Inbound
Forms
Custom Extraction
Model
Custom Classification/
Transformation Model
Value 1
Value 2
Value 3Elements
B CA
DATA LABELING MANUAL REVIEW(AS NEEDED)
Structured
Output
</>
24 | Copyright © 2019 Indico
Conclusions
• Successful construct for AI-driven process automation is
“scaled decisioning”
• Pick a use case where there are heterogeneous inputs and
homogeneous outputs
• Capturing the understanding of the content and the process
directly from the SME increases probability of success
• AI can sometimes feel like a “black box”-pick an
approach/solution that includes explainabilty