Bij Teleperformance helpen we klanten waarde toe te voegen aan het klanttraject. We gebruiken Data Science voor onze Omnichannel-klantinteracties om de behoeften van de klant te voorspellen, zodat we het beste antwoord kunnen geven.
3. Through our omnichannel
customer experience capabilities,
voice, e-mail, chat, click-
to-call, social media,
video chat, automation,
face-to-face, and other
channels that your
customers use.
we interact every year by
VOICE
EMAIL
WEB FORM
CHAT W/
LIVE
AGENT
SOCIAL
MEDIA
MOBILE
APP
SMS
INSTANT
MESSAGING
AUTOMATION
FACE-
TO-FACE
VIDEO KIOSK/
VIDEO TELLER/
VIDEO CHAT
CLICK-
TO-CALL
4. 4
Over the past 40 years we
have worked
tirelessly to build the
largest Customer
Experience
company on the planet.
We know how to create
great & trusted service
moments for the
customers of our clients.
223Kpeople
We are a team of
350facilities
76countries
Present in
160markets
Serving
265
We provide
service in
languages and
dialects
In 2017
Revenue of
$4.720bn
€4.180bn
5. Tom Mac-Kenzie
Digital Project Manager,
Teleperformance BLX
Tom.Mackenzie@Teleperformance.com
+31 (0)6 514 653 89
• Current: Digital Project Manager, Bus. Development
• Previous: Senior Manager, Google BLX Project
Sales Manager, Google BLX Project
Manager New Teams, Google BLX Project
• Based in Tilburg, NL
• 7+years experience in outsourcing industry
• Created Messaging-Playbook based on experiences with local and global
partners. Created Competitor-Matrix to assess key-players in Messaging
Software. Created and pitched client-specific Messaging Pitches to more than
20 Clients/Prospects. Researched and created main differences narrative
between Chat & Messaging.
• Specializes in providing effortless Customer Experience for enterprises through
Instant Messaging & other innovations within customer contact
• Extensive background within online Ad Sales, SEA, Sales & People Management
• Specialties: Instant Messaging | WhatsApp Business | Facebook Messenger |
Customer Contact | Sales | Effortless Customer Experience | Innovation |
Online Marketing | e-Commerce | WhatsApp | Social Messaging | In-App
Messaging
Helping companies deliver an effortless customer
experience through Instant Messaging Solutions
A passionate online professional with a clear focus on achieving positive ROI
and an effortless Customer Experience across all channels.
Over 7 years proven experience in Customer Contact, Instant Messaging,
Effortless Customer Service, Advertising, e-Commerce, Online Marketing &
Sales Management.
Teleperformance Experience: 7 Years
Relevant Industry Experience
9. 9
THE CHALLENGE
The costs of handling the channel Social Messaging are too
high. The growth of the messages is massive and because of
that, the workforce grows accordingly.
A large number of questions asked by customers have
been answered before, using this data would reduce time
and effort for the agent, but also increase the satisfaction
of both the customer and the employee.
10. 10
Goal
Reduce number of interactions and
time spent to get from question to
answer
Maintain customer satisfaction and
save time.
THE CHALLENGE
Question
Answer
12. The data science life-cycle
Assumptions were made to define question and answer patterns
Question
Answer
Data formatting
§ Define question and
answer
§ Move from “message”
format to “conversation”
format
13. The data science life-cycle
Language complex to model, but we have found solutions
Data cleaning
§ Product tagging
§ Text cleaning
§ Stop words
§ Synonyms
§ Spelling mistakes
Wasmachine Wasapparaat
washingmachineTAG
Kapot Stuk
Text vectorization
§ Bi-Grams
§ Term Frequency – Inverse
Document Frequency
Dimensionality reduction
§ Principal Component Analysis
15. Initial clustering
Once the product is know, we cluster the specific questions
Product tag
clustering
Question
level
clustering
Cluster
checks
Model
answers
Cluster
filtering
All
Smartphone
Screen
display
Broken
screen, fixing
cost
Other
TV Cut-off
28%72%
26%
Product tag
clustering
Question level
clustering
74%
36
1361
16. Initial clustering
The clusters are checked, to validate that there is only one question in them
Product tag
clustering
Question
level
clustering
Cluster
checks
Model
answers
Cluster
filtering
22. Self-learning algorithm
Matching in cluster
Database containing historic
conversations
New conversation
Agent options
Give model answer
Use answer of most similar question
Reject and write custom answer
23. Self-learning algorithm
Select model answer
Database containing historic
conversations
New conversation
Agent options
Give model answer
Use answer of most similar question
Reject and write custom answer
24. Self-learning algorithm
Extend existing cluster
Database containing historic
conversations
New conversation
Agent options
Give model answer
Use answer of most similar question
Reject and write custom answer
25. Self-learning algorithm
Select answer of most similar question
Database containing historic
conversations
New conversation
Agent options
Give model answer
Use answer of most similar question
Reject and write custom answer
26. Self-learning algorithm
Start new cluster
Database containing historic
conversations
New conversation
Agent options
Give model answer
Use answer of most similar question
Reject and write custom answer
27. Self-learning algorithm
Reject and write custom answer
Database containing historic
conversations
New conversation
Agent options
Give model answer
Use answer of most similar question
Reject and write custom answer
28. Self-learning algorithm
(optional) remove from cluster, add observation to history
Database containing historic
conversations
New conversation
Agent options
Give model answer
Use answer of most similar question
Reject and write custom answer
31. Self-learning algorithm
Matching outside cluster
Database containing historic
conversations
New conversation
Agent options
Use answer of most similar question
Reject and write custom answer
32. Self-learning algorithm
Matching outside cluster
Database containing historic
conversations
New conversation
Agent options
Use answer of most similar question
Reject and write custom answer
33. Self-learning algorithm
Start new cluster
Database containing historic
conversations
New conversation
Agent options
Use answer of most similar question
Reject and write custom answer
34. Self-learning algorithm
Matching outside cluster
Database containing historic
conversations
New conversation
Agent options
Use answer of most similar question
Reject and write custom answer
35. Self-learning algorithm
(optional) remove from history, add observation to history
Database containing historic
conversations
New conversation
Agent options
Use answer of most similar question
Reject and write custom answer
36. Self-learning algorithm
Overview
MatchingNew conversation
In cluster
Outside cluster
Use answer of most
similar question
Reject and write custom
answer
Agent options
Give model answer
Use answer of most
similar question
Reject and write custom
answer
Learning
Extend cluster
New cluster
(Optional) remove
match, add to set
New cluster
(Optional) remove
match, add to set
37. Road to production
The developed solution can be used in other customer service situations
Collective
memory
AgentBotFacebook
38. Road to production
The developed solution can be used in other customer service situations
Collective
memory
AgentBotFacebook
Phone
Agent
Website
(self help)
Tweets and
forums
Commercial
actions
39. SUPERVISED LEARNING
RAW DATA DATA SCIENCE ALGORITHM SUGGESTED
SOLUTION
OPTIONS
AGENT SELECTS
BEST SOLUTION
ALGORITHM
IMPROVED
AND IS AN
OPPORTUNITY TO TRAIN
YOUR MODEL!
40. THE CHALLENGE
The costs of handling the channel Social Messaging are too high. The growth
of the messages is massive and because of that, the workforce grows
accordingly. A large number of questions asked by customers have been
answered before, using this data would reduce time and effort for the agent,
but also increase the satisfaction of both the customer and the employee.
THE SOLUTION
Teleperformance created a “Collective Memory Database” containing
Historical Conversations. These conversations have been clustered with
algorithms using Data Science Techniques. Model answers have been
created for these clusters and 3 options are pushed to agents when a
customer query comes in. These options are “Give model answer”, “Use
historic answer of most similar question” and “Reject and write own
answer”.
SUMMARY
Provide Suggested Answers to Agents
IMPLEMENTATION TIME: 3 Months THE BENEFITS
ü Cost Saving: Reduced 12% costs in PoC
ü ROI: 6 Months
ü Improved Response Time with 1,5 minutes
ü Manual effort reduced by 14%
ü Higher standardization of processes
ü Expected increase of Employee Satisfaction
ü Expected reduction of Agent Onboarding
NEXT STEPS
ü Integrate with Self-Service Portal
ü Integrate in Voice Channel