Artificial Intelligence in the Enterprise: A Real-World Perspective
Bob O’Donnell, President and Chief Analyst
Goal, Agenda and Methodology
technalys isR E S E A R C H
Determine AI application usage in US businesses
What applications
What tools are used
Where AI apps are deployed
Goals and challenges
Online survey of 504 US-based businesses building and/or using AI applications
Medium and Large Businesses
Over 3,700 in initial sample
AI in Business is Very Real
• 18% of companies using AI apps• Within that group, 56% in full deployment
• Average number of AI apps per company is 9.9
50%61%
72%
56%
0%10%20%30%40%50%60%70%80%
Medium Business Large Business Early Adopters Total
AI in Full Deployment
technalys isR E S E A R C H
But Doubts RemainFor Those Who Don’t Use AI
50%
35%30%
28% 28%23%
18%
12%10%
7% 6%3%
0%
10%
20%
30%
40%
50%
60%
Costs Intrigued,but notready
Lack of in-house
expertise
Don’t know enough
Applicability Negativeimpact onpersonnel
Negativeimpact oncompanyoperation
Negativeimpact on
society
Don’t see the value
No need AI isoverrated
Other
Non-AI Usage Concerns
AI Applications Focused On Real World
71% 69% 67% 67% 66% 66% 65% 63% 62% 61% 61% 60% 60%56%
51%
43%
0%
10%
20%
30%
40%
50%
60%
70%
80%
Most Common AI Applications • Back office applications are most common
• New variations on “big data” analytics also strong
• Futuristic applications, like computer vision and voice, are growing
technalys isR E S E A R C H
36.9%
35.6%
27.5%
AI App Development Stages
Production Pilot Development
AI App MaturityVaries Greatly
Top 5 AI Apps in Production
Data Security
Spam Filtering
Network Security
Device Security
Call Center/Chatbot
Top 5 AI Apps in Pilot
Business Intelligence
Voice UI/Natural Language Processing
Image Recognition
Web/Social Media Analytics
Data Security
Top 5 AI Apps in Development
Robotics Manufacturing Efficiency/Predictive
Maintenance
Call Center/Chatbot
IoT Analytics
Physical Security
Tech Companies Dominate Current AI User Base
27%
13%
10% 9% 8% 7% 7%4% 4%
2% 2% 2% 1% 1% 1% 0% 0% 0% 1%
0%
5%
10%
15%
20%
25%
30%
INDUSTRIES USING AI
Inferencing vs. Training• Most companies (73%) are
building their own AI models and training them
• But over ¼ are relying on other models and only doing inferencing
26.9%
28.3%
44.8%
Inferencing and Training Usage
Inferencing Training Both
Chip Architectures for AI More Varied Than Expected
27.1%
22.8%13.8%
11.9%
7.0%1.6%
15.9%
Chip Architecture Used For Inferencing
CPU Dedicated AI ProcessorGPU Dedicated Vision ProcessorFPGA OtherDon't Know
28.7%
26.3%16.6%
10.3%
2.5%
15.7%
Chip Architecture Used For Training
CPU Dedicated AI Processor
GPU FPGA
Other Don't Know
Google Winning in theAI Cloud
But 25% Not Using Cloud For AI
26.6%
19.8%
19.2%
9.7%
24.6%
Cloud Platform Usage
Google Cloud Amazon AWS Microsoft Azure Other Not Using
Cognitive Services Used by 80% of AI App Developers
0%10%20%30%40%50%60%70%80%90%
100%
Cognitive Services by Application
Computer Vision Speech Natural Language Knowledge Other Not Using
technalys isR E S E A R C H
Data Sources for AI Models a Challenge
32.5%
23.3%
22.1%
14.3%
1.7% 6.1%
Data Sources for Models
Existing Internal Data Paid 3rd Party
Newly Collected Internal Data Public Domain
Other Don't Know
technalys isR E S E A R C H
Interesting AIApplication Facts
AI Application Users• 84% Internal• 54% External• 38% for Both
Cloud Functions• 38% Inferencing• 30% Training• 29% AI as a Service
Container Usage• 37% Currently Using• 32% Planning to Use• 31% No Intention of Using
Split Across 10 OptionsFRAMEWORK USAGE VARIED
63% 31% Don’t Know
AI FUNCTIONS
55% See AI as aFeature
Top 5 AI Frameworks
TensorFlow
ONNX
MXNet
ResNet
Theano
Top 5 Apps Where AI is Core
Call Center/Chatbot
Image Recognition
Robotics
Voice UI/Natural Language Processing
Network Security
Improving Efficiency is Top Goal for AI
Improveefficiency
Speedup/automate
tasks
Increasesecurity
Analyticalinsights
Cost savings Processimprovements
Newcapabilities
Revenuegeneration
Mitigating risks Reducecomplexity
Reduceheadcount
Increaseheadcount
Other primarygoal
AI Goals• Taking into
consideration both the frequency of selection and the rankings, it’s clear companies are hoping to gain efficiency and speed from deploying AI
• Security applications are also important
Complexity and Cost RemainTop Challenges for AI
Technologycomplexity
Implementationcomplexity
Costs Getting accessto high-quality
data sets
Uncertainty ofimpact
Limited internalskill sets
Pace of change Organizationalpolitics
Overwhelmingnumber ofpotentialsolutions
Definingpurpose,
business case
Impact onheadcount
Overwhelmingnumber of
vendor choices
Other challenge
AI Challenges • Access to good data to train models and uncertainty of the full impact of AI are also big concerns
Private Cloud and Public Cloud Usage for AI Apps Nearly IdenticalMobile and Edge Still Small
20.9%
20.7%
19.6%
12.1%
9.4%
7.0%
10.3%
Deployment Locations
Private Data Center Cloud
PC Client As a Service (Many Locations)
Mobile Edge
Don't Know
technalys isR E S E A R C H
IT and Engineering Lead, but AI Apps Built By Many People
32%
27%
15%
9%
6%
5%4% 2%
AI App Builders
In-House IT In House Engineering/Programming
Contractors Tech Vendor
Large SI AI Specialty Developer
Specialty SI Other
technalys isR E S E A R C H
IT, Operations and Engineering Top AI Department Users
• Large gap in usage vs. valuefor IT
• Support and customer service getting best value for their efforts
85%
50% 48%39% 38%
19% 19% 15%9%
5%
57%
43% 42% 39%
28%
13%18%
10% 6%2%
0%10%20%30%40%50%60%70%80%90%
Using vs. Benefitting
Using Benefitting
Conclusions
• Most early AI efforts are more practical than futuristic
• Security and efficiency improvements are key
• For many, AI applications are next generation analytics tools
• No easy AI solution• Technology requirements and deployments vary
greatly by application
• Legitimate concerns exist, but there is a lotof excitement for the technology and its potential impact
• Costs and complexity remain big concerns
ContactBob O’DonnellPresident and Chief AnalystTECHnalysis Research, LLC1136 Halsey Blvd.Foster City, CA 94404
[email protected](650) [email protected]
For additional information and complete survey results, a 178-slide version of this report is available for purchase.