Data Analytics

EDITOR: Seth Earley, Earley Information Science, seth@earley.com

There Is No AI
Without IA
Seth Earley, Earley Information Science

A
rtificial intelligence (AI) ence in 2015,4 uses deep learning customers. These initiatives in-
is increasingly hyped by and neural nets. DigitalGenius first clude improving personalization
vendors of all shapes and classifies the incoming questions of the user experience by present-
sizes—from well-funded into one of several categories for ing more relevant content; tuning
startups to the well-known software further processing: product infor- search results to return exactly
brands. Financial organizations are mation, account information, ac- what the user is interested in; and
building AI-driven investment advi- tion request, comparison question, improving the effectiveness of of-
sors.1 Chat bots provide everything recommendation question, and so fers and promotions. Organiza-
from customer service2 to sales as- on. These classifications qualify tions might also strive to increase
sistance.3 Although AI is receiv- as a foundational element of infor- response rates to email communi-
ing a lot of visibility, the fact that mation architecture. The starting cations; provide better customer
these technologies all require some point is contextualizing the query, self-service; increase participation
element of knowledge engineer- which is then passed to other mod- in user communities and other
ing, information architecture, and ules, including product informa- social media venues; and gener-
high-quality data sources is not tion systems and other databases ally enhance the product experi-
well known. Many vendors sidestep and APIs. Each of these systems ence through various other online
this question or claim that their al- and sources in turn needs to be mechanisms. In each of these
gorithms operate on unstructured well architected to return the cor- cases, the means of engagement is
information sources, “understand” rect information. If the informa- providing a relevant piece of data
those sources, interpret the user tion is not structured and curated or content (in the form of promo-
query, and present the result with- in some way, there is nothing for tions, offers, sales, next-best ac-
out predefined architectures or hu- the system to return. The ability of tion, products for cross sell and
man intervention. That very well DigitalGenius to use AI to engage upsell, answers to questions, and
might be true in certain circum- customers is predicated on having so on) at exactly the right time and
stances, but most applications re- high-quality, structured data. in the context that is most mean-
quire a significant amount of hard ingful and valuable to the user.
work on the part of humans before Digital Engagement: This is done by interpreting the
neural nets, machine learning, and The Right Information various signals that users provide
natural language processors can in Context through their current and past
work their magic. Organizations are in a never- interactions with the organiza-
DigitalGenius, a product that ending cycle of improving their tion. These signals include prior
received press coverage at a confer- digital means of engaging with purchases, real-time click-stream

58 IT Pro May/June 2016 Published by the IEEE Computer Society 1520-9202/16/$33.00 © 2016 IEEE
data, support center interactions, program is. A corpus of informa- ent the right content for the user’s
consumed content, preferences, tion contains the answers that the context, and requires not only
buying characteristics, demo- program is attempting to process that our customer data is clean,
graphics, firmographics, social and interpret. Structuring that in- properly structured, and inte-
media information, and any other formation for retrieval is referred grated across multiple systems
“electronic body language” that is to as knowledge engineering and the and processes but also that the
captured by marketing automa- structures are called knowledge system understand the relation-
tion and integration technolo- representation. ship between the user, his or her
gies. A search query could return specific task, the product, and the
a different result for a technical Ontologies as Knowledge content needed—all assembled
user than a nontechnical user, for Representations dynamically in real time. Building
example. At its core, search is a Knowledge representation con- these structures and relationships
recommendation engine. The sig- sists of taxonomies, controlled and harmonizing the architec-
nal is the search phrase, and the vocabularies, thesaurus struc- ture across the various back-end
recommendation is the result set. tures, and all of the relationships platforms and front-end systems
The more that is known about the bet ween terms and concepts. results in an enterprise ontology
user, the more the recommenda- These elements collectively make that enables a personalized, om-
tion can be tailored. If the recom- up the ontology. An ontology rep- nichannel experience. Some might
mendation is related to a product, resents a domain of knowledge call this an enterprise informa-
then clean, well-structured prod- and the information architecture tion architecture; however, there
uct data is a precondition. structures and mechanisms for is more to it than the data struc-
accessing and retrieving answers tures. Recall that the definition of
Personalization, in specific contexts. Ontologies an ontology includes real-world
User Signals, and
Recommendations
Getting these recommendations Every AI program interfaces with information, and
right and truly personalizing the
user experience requires that
the better that information is structured, the more
product data be correctly struc- effective the program is.
tured and organized, that con-
tent processes be integrated into
product onboarding, and that can also capture “common sense” logic and relationships. The ontol-
associations can be made among knowledge of objects, processes, ogy can contain knowledge about
products, content, and user intent materials, actions, events, and processes, customer needs, and
signals. The relationship of prod- myriad more classes of real-world content relationships.5
ucts to content is based on knowl- logical relationships. In this way,
edge of the user’s task and what an ontology forms a foundation Mining Content for
content would help him or her for computer reasoning even if the Product Relationships
accomplish it. The task might be answer to a question is not explic- Consumer and industrial prod-
a review, a how-to, product speci- itly contained in the corpus. The ucts need to be associated with
fications, reference materials, in- answer can be inferred from the content and user context, but
structions, diagrams and images, facts, terms, and relationships in content can also be mined to sug-
or other content that helps the the ontology. In a practical sense, gest products for a user context.
user decide on the purchase. this can make the system more For an industrial application, a
AI encompasses a class of ap- user friendly and forgiving when user might need parts and tools
plication that allows for easier the user makes requests using to complete maintenance on a
interaction with computers and phrase variations, and more capa- hydraulic system. Using adap-
also allows computers to take on ble when it encounters use cases tive pattern-recognition software
more of the types of problems that were not completely defined to mine reference manuals about
that were typically in the realm of when it was developed. In effect, hydraulic systems and repair, the
human cognition. Every AI pro- the system can “reason” and make system can extract a list of needed
gram interfaces with information, logical deductions. tools and related parts. A search
and the better that information is Correctly interpreting user sig- on hydraulic repair will present
structured, the more effective the nals enables the system to pres- a dynamically generated product

computer.org/ITPro  59
 Data Analytics

page based on the product rela- first car phone in the ’80s, in ad- In many cases, the data being
tionships and correlated with the dition to costing several thousand processed or corpus being ana-
company’s offerings. To some dollars, left little room for any- lyzed by AI systems is typically
information professionals, this thing else in the spacious trunk). less structured than better-orga-
might sound complex and cum- Most AI is quietly taken for grant- nized sources such as financial
bersome to implement; however, ed today. The word processor I am and transactional data. Learn-
there are emerging approaches using was once considered an ad- ing algorithms can be used to
that bring these aspirations closer vanced AI application! both extract meaning from am-
to reality. biguous queries and attempt to
Simplicity Is Hidden make sense of unstructured data
If It Works, It’s Not AI Complexity inputs. Humans might phrase
The concept of what constitutes Under the covers, AI is complex; questions using different termi-
AI has evolved as technology has however, this complexity is hid- nology, or they might ask overly
evolved. A colleague of mine has den from the user and is, in fact, broad questions. They are not
said, “It’s artificial intelligence the enabler of an easy, intuitive always clear about their objec-
until you know how it works.” An experience. It is not magic, and it tives—they don’t necessarily
interesting perspective indeed. I requires foundational structures know what they are looking for.
found support for this in mate- that can be reused across many This is why human sales people
rial from an MIT AI course (see dif ferent processes, depa r t- typically engage prospects in
http://bit.ly/1WTCUXK): ments, and applications. These conversations about their overall
structures are generally first de- needs, rather than asking them
There’s another part of AI […] veloped in silos and standalone what they are looking for. (At
that’s fundamentally about appli- tools; however, the true power least, the good ones do.)
cations. Some of these applications will be realized when consid- Inserting AI into the process is
you might not want to call “intel- ered in a holistic framework of more effective when users know
ligent” […] For instance, compilers machine-intelligence-enabled what they want and can clearly
used to be considered AI, because infrastructure. AI will change articulate it, and where there is
[…] statements [were in a] high- the business landscape but will a relatively straightforward an-
level language; and how could a require investments in product swer. The algorithm does what
computer possibly understand? and content architecture, cus- it does best in terms of handling
[The] work to make a computer un- tomer data, and analytics, and variations in how those ques-
derstand […] was taken to be AI. the harmonization of tools in tions are asked, interpreting the
Now […] we understand compilers, the customer engagement eco- meaning of the question, and
and there’s a theory of how to build system. The organizations that processing other unstructured
compilers […] well, it’s not AI any- adopt these approaches will gain signals that help further con-
more. […] When they finally get significant advantages over the textualize the user’s intent.
something working, it gets co-opted competition. There are many flavors of AI
by some other part of the field. So, and many classes of algorithm
by definition, no AI ever works; if it Clean Data Is the Price that comprise AI systems. How­
works, it’s not AI. of Admission ever, even when AI systems are
AI approaches are proposed as the used to find structure f rom
Seemingly intractable problems answer to enterprise challenges completely unstructured in-
have been solved by advances in around improving customer en- formation, they still require
processing power and capabilities. gagement by dealing with infor- structure at the data layer.
Not long ago, autonomous vehicles mation overload. Before those Given that the data being
were considered technologically approaches can be leveraged, searched by the A I system is
infeasible due to the volume of however, organizations need to unstructured, why do we need
data that needed to be processed in possess the data needed as inputs information architecture? Un-
real time. Speech recognition was to machine learning algorithms structured information is typi-
unreliable and required extensive that can in turn process these di- cally in the form of text from pages,
speaker-dependent training ses- verse signals from unstructured documents, comments, surveys,
sions. Mobile phones were once and structured sources. Clean, social media, or some other source.
“auto-mobile” phones, requiring well-structured, managed data is Though it is unstructured, there
a car trunk full of equipment (my assumed. are still parameters associated with

60 IT Pro May/June 2016

 Table 1. Example applications for AI technology.
Application/problem Typical approach Why it will not work How AI helps
Omnichannel selling Integrations across systems Content is typically out Harvesting signals of user intent
that are inflexible of synch, and systems across channels and devices;
have brittle integrations. intelligent integration across
diverse classes of customer-
engagement technologies.
Personalization Classification of users Assumptions are Incorporating multiple data
and recommendation according to personas rudimentary and simplistic, sources, ontologies, and reasoning
engines and presentation of limit user choice, and rarely algorithms.
content based on use cases anticipate needs.
Dynamic content Limited assembly based on The number of permutations Assembling content that meets
assembly curated content cannot be tracked, managed, a user’s preferences while
or understood. incorporating real-time data, such
as news and market performance
supported by knowledge
embedded in ontologies.
Field service View of maintenance The volume of information Processing performance data to
equipment maintenance schedules and analysis from sensors and hidden predict or prevent maintenance
of field reports. signals from degrading events.
performance is not perceivable
by humans.
Customer self-service Knowledge bases There are insufficient Enabling intelligent agent
and FAQs considerations of user context, knowledge base access using
language variations, fast- multiple variables to provide
changing content, and a high specific answers with metrics-
volume of possible use cases. driven feedback loops.
Call center support Reps trained through For technical and specialized Enabling intelligent agent access
on-the-job experience, knowledge, the level of integrating answers and content
knowledge bases experience needed is too from multiple knowledge and data
costly. sources using natural language
interfaces.
Product and tool Hand-built landing pages Maintenance of hand-curated Harvesting product relationships
presentation and assembled landing through knowledge acquisition
pages becomes costly and and mining of reference materials
unmanageable. based on maintenance tasks and
repair use cases.

the source and context. Social identified, but the input needs to Some vendors might debate the
media information requires vari- have structure. value of this approach, insisting
ous parameters to describe users, A common fallacy in consider- that their algorithms can handle
their posts, relationships, time and ing big data sources that form in- whatever you throw at them—
location of posts, links, hashtags, puts for machine learning is that however, I would argue that this
and so on. The information ar- because the data is “schema-less” is only the case when the ontolo-
chitecture question in this case (does not have a predefined struc- gies are self-contained in the tool.
characterizes the structures of the ture), no structure is required. Even so, there will always be gaps
input data, so that the system can Data still requires attribute defini- between what a tool developed for
be programmed to find patterns tions, normalization, and cleans- broad adoption can contain and
of interest. Even in the case of un- ing to apply machine learning and the specialized needs of an enter-
super vised machine learning pattern-identification algorithms.6 prise. Even if the tool is developed
(a class of application that derives As enterprises embark on the path for a specific industry, the differ-
signals from data that are not to machine learning and AI, they ences in processes found from or-
predefined by a human), the pro- should, first and foremost, be de- ganization to organization require
grammer still needs to describe veloping an enterprise ontology specialized vocabularies and con-
the data in the first place with at- that represents all of the knowl- textual knowledge relationships.
tributes and values. There might edge that any AI system they This is a significant undertaking;
not be predefined categories for deploy would process, analyze, le- however, not doing so misses an
outliers and patterns that are verage, or require. essential step in the process.

computer.org/ITPro  61
 Data Analytics

Table 2. Examples of AI tools with representative applications, limitations, considerations, and data sources.

Class of tool Best suited for Limitations

Inference engine Deriving product and data relationships from Significant effort required in certain cases for
unstructured content development of adaptive pattern-recognition
algorithms
Intelligent agent Highly selective search and information retrieval Breadth of possible use cases, variability of
for definable processes queries and questions, degree of manual
term-relationship mapping
Auto-classifier Large volumes of higher-quality content where Rules bases can become very complex,
training sets or clear rules can be applied high volumes of training content required,
variability of content
Entity extractors Predictable formats for data (Social Security Ambiguity of entity values (for “Washington,”
numbers, addresses, names, phone numbers, can be Washington State, George Washington,
account numbers) Washington, DC, and so on), variations in
format, quality of content
Unsupervised Pattern detection, recognition, and prediction; Depends on creation of foundational hypothesis
machine learning anomaly detection, outliers, hidden attributes, to define type of outlier or pattern (though
and relationships; discovering new patterns or not necessarily details of pattern); selection
segmenting audiences, content, or data into of class of algorithm requires technical
clusters and groupings sophistication and iterative testing across
many classes of unsupervised learning from
data signals to content, images, or transactions
Supervised Pattern detection based on training data and Requires training sets and example data that
machine learning used for well-understood patterns; discovering teach the algorithm what type of information
data, content, and relationships based on is being sought; needs large enough samples
examples; finding similar documents, purchase to test various hypotheses and an understanding
patterns, audiences with like characteristics; of the specific outcome sought; risk of finding
predicting outcomes patterns where none exist (over fitting)
Hybrid unsupervised Finding hidden patterns (unsupervised); using Similar in concept to creating a survey with
and supervised those patterns to train an algorithm to locate open-ended questions, then taking the answers
learning more instances of the data or content to the survey and using those results to form
closed-ended surveys

Much of what is described as AI that will allow for continuous AI approach will require a greater
is an extension of well-known ap- improvement of AI and cognitive level of investment, sponsorship at
proaches to addressing informa- computing systems. the executive level, program-level
tion management problems, all of governance, and an enterprise
which require clean foundational Identifying Use Cases span of influence. It will also re-
data and information structures as Differentiating AI use cases from quire a longer-term commitment
a starting point. The difference be- standard information manage- than a typical information man-
tween standard information man- ment use cases requires consid- agement project. Although there
agement and practical AI lies in ering the sources of data that are opportunities to deploy limit-
understanding the limits of these comprise the “signals” being pro- ed scope AI, fully leveraging it as a
technologies and where they can cessed, the type of task the user transformative class of technology
best be applied to address enter- faces, and the systems that will be should be part of an overall digital
prise challenges. part of the solution. The differ- transformation strategy in some
The remainder of this article ence in approaches to these prob- cases on the scale of enterprise re-
describes how your organization lems lies in how the sources of source planning (ERP) programs,
can identify use cases that will data will be curated and ingested, with commensurate support,
benefit from AI, identifies data how organizing principles need funding, and commitment. (Some
sources that offer reliable and to be derived and applied, the ERP programs can cost US$50 to
meaningful insights to train and sophistication of the functional- $100 million or more). Although
guide AI, and defines governance, ity desired, and the limitations of no organization will approach an
curation, and scalable processes the current solution in place. An unproven set of technologies with

62 IT Pro May/June 2016

 Decision factors Example data
Volume of content, quality of corpus, structure of data within Structured documents such as catalogs, engineering
content (use of tables, headers, other identifiers, and delimiters) specifications, knowledge bases, ERP data

Narrow scenarios, defined repeatable Policy and procedure manuals, chat logs, customer
tasks, sophisticated audience, standard terminology for domain service databases, support tickets, knowledge harvested
from human subject matter experts
High-value content that has a predictable structure, Semi-structured content such as regulations, journal
curated with editorial standards articles, news articles, policies and procedures, support
documentation, engineering documents
Security, compliance, and reference-based applications (looking Same semi-structured content sources mentioned
for personally identifiable information, sensitive account above, along with controlled vocabularies and
information, intellectual property protection, fact identification) thesaurus structures to identify entity values

Determination of type of patterns that are being detected; not Unsupervised learning can be applied to virtually any
typically selected in isolation because learning algorithms are data source as the tools look for patterns; can be applied
embedded in many aspects of AI application; requires guidance of to unstructured and semi-structured content as well as
data science specialist knowledgeable in specific type of algorithm transactional data, images, audio and video, scientific
datasets, sensor data, social media sources, and so on

Similar challenges to unsupervised learning: typically embedded Same as unsupervised; the difference is having training
in AI applications rather than being applied as standalone tool sets that contain patterns sought
and requires specialized knowledge; can be used to find types
of images, content or data, purchase patterns, audiences, users,
combinations of products, or other patterns depending
on data sources
Most AI programs contain hybrid implementations of supervised Same data used for unsupervised learning can be used
and unsupervised and therefore the complexity of each approach as input for supervised learning after representative
can be compounded; requires research and experimentation to patterns have been identified
find correct combinations and fine-tune algorithms

that level of commitment, fund- organization’s knowledge landscape user problems. Customer account
ing needs to be allocated to extend and implementation of new gov- data and purchase history can be
proven approaches with emerging ernance, metrics, and data quality processed to look for similarities in
AI technologies. programs—governance to make buyers and predict responses to of-
The roadmap for AI transforma- decisions, metrics to monitor the fers; email response metrics can be
tion includes continuous evaluation effectiveness of those decisions, and processed with text content of offers
of payback and ROI and focuses data quality to fuel the AI engine. to surface buyer segments. Product
on short-term wins while pursuing Table 1 presents example appli- catalogs and data sheets are sources
longer-term goals. Most organiza- cations for AI technology. of attributes and attribute values.
tions are attempting to solve the Public references can be used for
problems described in Table 1 with Identifying Data Sources procedures, tool lists, and product
limited approaches, departmental- Training data can come from typi- associations. YouTube video con-
level solu­tions, standalone tools, cal knowledge bases, the more tent audio tracks can be converted
and insufficient funding. These highly curated the better. Call cen- to text and mined for product as-
problem classes are facing most en- ter recordings and chat logs can sociations. User website behaviors
terprises, and though progress can be mined for content and data re- can be correlated with offers and
be made with limited resources and lationships as well as answers to dynamic content. Sentiment analy-
siloed approaches, this would be an questions. Streaming sensor data sis, user-generated content, social
extension of business as usual. Tru- can be correlated with historical graph data, and other external data
ly transformative applications will maintenance records, and search sources can all be mined and re-
require an enterprise view of the logs can be mined for use cases and combined to yield knowledge and

computer.org/ITPro  63
 Data Analytics

user-intent signals. The correct data will occur in subtle ways—such J., 17 Nov. 2015; http://blogs.wsj.
sources depend on the application, as improved usability of applica- com/digits/2015/11/17/can-artificial
use cases, and objectives. tions and findability of informa- -intelligence-sell-shoes/.
Table 2 describes examples of tion. These will not necessarily 4. R. Miller, “DigitalGenius Brings
AI tools with representative ap- appear on the surface to be AI. Artificial Intelligence to Customer
plications, limitations, consider- Over time, AI-driven intelligent Service via SMS,” Tech Crunch, 5
ations, and data sources. Though vir tual assistants will become May 2015; http://techcrunch.com/
not meant to be an exhaustive more fluent and capable, and will 2015/05/05/digitalgenius-brings
list, and recognizing that one become the preferred mechanism -artificial-intelligence-to-customer
class of tool is frequently lever- for interacting with technology. -service-via-sms/.
aged in other tools and applica- Humans create knowledge, while 5. S. Earley, “Lessons from Alexa:
tions (an intelligent agent can use machines process, store, and act Artificial Intelligence and Machine
inference engines, which in turn on that knowledge. AI is applied Learning Use Cases,” blog, Earley
can leverage learning algorithms, human knowledge. Organizations Information Science, 24 Mar. 2016;
for example), the table articulates need to develop the foundations w w w.e a r l e y.c o m / blo g / l e s s on s
considerations for exploring one for advancing AI by capturing -alexa-artificial-intelligence-and
approach versus another. and curating that knowledge and -machine-learning-use-cases.
by building the foundational data 6. J. Brownlee, “How to Prepare Data
Defining Governance, structures that form the scaffold- for Machine Learning,” Machine
Curation, and Scalable ing for that knowledge. Without Learning Mastery, 25 Dec. 2013;
Processes those components, the algorithms http://machinelearningmastery.com/
AI and cognitive computing are have nothing to run on. how-to-prepare-data-for-machine
managed in the same way as many -learning/.
other information and technology References
governance programs. They require 1. J. Vögeli, “UBS Turns to Artificial Seth Earley is CEO of Earley Informa-
executive sponsorship, charters, Intelligence to Advise Clients,” tion Science (www.earley.com). He’s an
roles and responsibilities, decision- Bloomberg, 7 Dec. 2014; www. expert in knowledge processes, enterprise
making protocols, escalation pro- blo omb e r g.com /ne w s /a r t ic le s / data architecture, and customer experi-
cesses, defined agendas, and linkage 2014-12-07/ubs-turns-to-artificial ence management strategies. His interests
to specific business objectives and -intelligence-to-advise-wealthy include customer experience analytics,
processes. These initiatives are a -clients. knowledge management, structured and
subset of digital transformation and 2. C. Green, “Is Artificial Intelligence unstructured data systems and strategy,
are linked to customer life cycles and the Future of Customer Service?” and machine learning. Contact him at
internal value chains. Because the MyCustomer, 3 Dec. 2015; www. seth@earley.com.
objective is always to affect a pro- mycustomer.com/service/channels/
cess outcome, all AI and cogni- is-artificial-intelligence-the-future
tive computing programs are closely -of-customer-service. Selected CS articles and
aligned with ongoing metrics at mul- 3. E. Dwoskin, “Can Artificial Intelli- columns are available for free at
tiple levels of detail—from content gence Sell Shoes?” blog, Wall Street http://ComputingNow.computer.org.
and data quality to process effective-
ness and satisfaction of business im-
peratives—and ultimately are linked IT Professional (ISSN 1520-9202) is published bimonthly by the IEEE Computer Society. IEEE
to the organizational competitive Headquarters, Three Park Avenue, 17th Floor, New York, NY 10016-5997; IEEE Computer So-
ciety Publications Office, 10662 Los Vaqueros Circle, PO Box 3014, Los Alamitos, CA 90720-
and market strategy. Milestones and 1314; voice +714 821 8380; fax +714 821 4010; IEEE Computer Society Headquarters, 1828
stages are defined to release fund- L St. NW, Suite 1202, Washington, DC 20036. Annual subscription: $49 in addition to any
IEEE Computer Society dues. Nonmember rates are available on request. Back issues: $20 for
ing for program phases, each with members, $143 for nonmembers.
defined success criteria and measur- Postmaster: Send undelivered copies and address changes to IT Professional, Member-
ship Processing Dept., IEEE Service Center, 445 Hoes Lane, Piscataway, NJ 08854-4141.
able outcomes. Periodicals Postage Paid at New York, NY, and at additional mailing offices. Canadian GST
#125634188. Canada Post Publications Mail Agreement Number 40013885. Return unde-
liverable Canadian addresses to PO Box 122, Niagara Falls, ON L2E 6S8, Canada. Printed
in the USA.

A
I will no doubt continue Editorial: Unless otherwise stated, bylined articles, as well as product and service descriptions,
reflect the author’s or firm’s opinion. Inclusion in IT Professional does not
to impact every aspect of necessarily constitute endorsement by the IEEE or the Computer Society. All
our personal and profes- submissions are subject to editing for style, clarity, and space.
sional lives. Much of this impact

64 IT Pro May/June 2016