This presentation covers the landscape of AI-enabled enterprise search.
The presentation was given at Sinequa's INFORM2019 events in both NYC and Paris.
Learn more about AI-enabled enterprise search on Emerj: https://emerj.com/?s=enterprise+search
Injustice - Developers Among Us (SciFiDevCon 2024)
The Digital Workplace Powered by Intelligent Search
1. The Digital Workplace Powered by
Intelligent Search, Today and Tomorrow
Daniel Faggella
CEO at Emerj Artificial Intelligence Research
2. Presentation Outline
● Background in Brief
● Enterprise Search - Then and Now
● Intelligent Search Use-Case Overview
○ 1 - Tagging and Clustering
○ 2 - 360º View of the Customer or User
○ 3 - Concept and Advanced Entity Search
● Future Forecast
● /end
emerj.com @danfaggella
3. We help large organizations (the World Bank,
global pharma giants, etc) make critical strategic
decisions about AI strategy and AI impact.
● AI market sizing, growth-rate analysis
● Competitive intelligence and strategy
● Vendor selection and AI adoption
Presenting our AI Research at
United Nations HQ, NYC
5. Search structured documents.
Emphasis on predictable formats and
direct-match keyword searching.
Metadata is applied manually and
painstakingly.
Difficulty: Integration, defining metadata
ontologies, solving a defined use-case.
Digital text is searchable.
Enterprise Search, Then and Now
Search unstructured documents.
Emphasis on “understanding”, clustering, and
metadata.
Metadata is applied programmatically and at
scale.
Difficulty: Integration, defining metadata
ontologies, solving a defined use-case.
Digital documents, scanned paper
documents, images, microfiche -- all is
searchable.
6. ● There is still plenty of value in older enterprise search
approaches, by allowing information accessibility, and in
organizing previously unorganized data.
● AI and ML approaches take these benefits to the next
level, by:
○ Making more information available (OCR, advanced
metadata, etc…)
○ Allowing users to ask more questions of the data
itself (reporting on broader patterns, finding more
direct answers)
Enterprise Search, Then and Now
emerj.com @danfaggella
7. ● Newer AI vendors underestimate the significant
integration challenges to bring AI into the enterprise.
● AI-enabled search and discovery applications are not
unique in this respect.
Enterprise Search, Then and Now
emerj.com @danfaggella
11. ● Adding tags and meta data manually, and training
systems
○ (Note: The value here still relies on human
ability to determine the use-case and the
meta data ontology! That’s beyond AI)
● This data can be added retroactively to an entire
corpus, or added upon entry
● Potential metrics of success:
○ Improved speed and efficiency of any
business process involving search
1. Enrichment and Classification
Examples:
Proactively protect confidential
information by having AI
categorize the confidentiality of
documents - based on initial
human training (rather than
relying on all employees to
intuitively know the confidentiality
level).
Manufacturing: Search through
production orders for mentions of
specific cluases or terms.
12. ● Enabling sales and support people with a full
view of a user or customer’s situation / history
● Potential metrics of success:
○ Improving customer service satisfaction
○ Reduction of time-to-resolution for CS
○ Improved upsell close rates for salespeople
● 75% of the applications of enterprise search in
the financial services sector feature Customer
Information Retrieval as a main featured
capability, more than any other use-case.
2. 360º View of the Customer or User
Examples:
A call center rep might see (a) a
summary of recent support calls
and chats with a customer, (b)
what those calls were about,
combined with (c) the ability to
find contracts and docs related to
that customer.
In the future, this use-case may
also involve suggesting next
actions or approaches to the call
center rep (“coaching”).
13. ● Previous search systems could search for terms:
○ “Wells Fargo”
○ “Pharma”
○ etc
● ML enables broader inquiry ability, including:
○ “Contracts for X service over 18 months
long”
○ “Invoices that don’t reference the service
paid for”
○ etc
3. Concept Search
Examples:
Banking: Search for all
documents that reference
LIBOR, or LIBOR-related
terminology.
Life Sciences: Search toxicology
reports that mention specific
types of complex or broad
symptoms.
15. Present
A huge bulk of the value of enterprise AI search comes not from advanced AI features, but
from:
● Tagging and clustering
● Entity recognition
● An established process to integrate and connect data systems, and determine meta tag
ontologies and structures and help the client
Today, value doesn’t lie in the fanciest AI tricks. Value lies in accessing data and making it
reasonably accessible to the people who need it. Proper integration, knowledge of workflows,
and basic, working functionality seems to be most important today.
5-Year Future Forecast
16. Source:
Emerj Artificial Intelligence
Research
“Enterprise Search and
Discovery - AI Capability
Overview”
Level of
Advantage
Competitive Advantage
High Client relationships with data access (storage, analytics, etc)
Middle-High Client relationships without data access (trust)
Middle Knowledge of the subject-matter (types of data)
Middle Knowledge of systems and workflows (processes, IT systems)
Middle-Low Data science talent (experience with applied AI, ability to iterate
models)
17. Source:
Emerj Artificial Intelligence
Research
“Enterprise Search and
Discovery - AI Capability
Overview”
Level of
Difficulty
Feature
High Providing “Answers” (Receiving sentence answers, not reports or
lists, e.g. “There are 47 contracts with XYZ type of clause included in
them since 2012”)
Middle-High Insights (e.g. Predictive analytics, notifications of anomalous activitiy
– notifying users to activity before)
Middle Natural language search (e.g. Receiving a text answer to a natural
text question like “How many of our client accounts have spent over
$1MM in the last 6 months?”)
Middle Enrichment and classification (e.g. Automatically tag documents on
intake - and/or suggest relevant metadata for human users to approve)
Low Entity search (e.g. Finding places, people, things, companies, or
concepts - in text data)
Low Reporting (e.g. Finding the number of docs in various categories,
instances of entities over time, etc)
18. 5-Year Future Forecast
5-Year Future
More specific search functions will be built as part of workflows (in compliance, in customer
service, etc), allowing human users to instantly reach the info the need when they need it.
Purpose-built solutions suited to specific use-cases will become more the norm. Insight
applications and increasingly broader “concept search” will define future development.
So will accessibility. Interfaces will develop to allow non-technical experts to set up custom
searches, and derive insights as they see fit.
Developing meta tag ontologies and determining how to connect to disparate data silos will
still be a massive challenge, but there will be best-practices that make it less painful than it is
today.