Tải bản đầy đủ - 0trang
1Sales and Marketing Reloaded—Deep Learning Facilitates New Ways of Winning Customers and Markets
communication on the Internet, be it the website or in social networks,
today belong to a company’s everyday routine. The efficiency of a sales or
PR campaign heavily depends on the choice of companies and people to be
addressed. Are they interested in the subject? Is this a well-chosen point in
time? Has the company just concluded a contract with an innovative CMS
provider, or is an outdated stack still being used? Classical sales and marketing approaches define target groups by way of simple selections or segmentations. Companies are selected on the basis of commercial sectors and sales
margins and transferred into the sales process.
Prior to the first call by the sales team, little can be said about the probability of the conversion with this approach. There is neither data nor
a method available to make a forecast about whether the prospective customers can really be won over as a customer in the sales funnel. Yet, for an
efficient and agile sales process, having extensive and up-to-date data is crucial. The establishment and development of individual leads in issues of the
topics they focus on, their sales forecasts and their digitality are crucial for
successful communication. Accordingly, an ideal system should make a sure
prediction as to which prospective customer will be the next to sign a contract. This way, the sales team can achieve the maximum conversion rate.
The high complexity of the data and the high dynamics this data underlies are a typical field o application for deep and machine learning algorithms. In the following, I will illustrate how these are applied to the field of
automated lead prediction.
5.1.2Analogy of the Dating Platform
Tell us your customers with the highest sales and we will predict who your
next successful customers will be. (Kulpa 2016)
In principle, lead prediction can be easily compared with a dating platform.
In comparison with a simple assumption about which products go well with
a company, lead prediction learns new information from every new customer to, in turn, predict better customers. The predictions become more
reliable and precise from the interaction and the feedback resulting from it
In comparison with a sales rep, who avails of a subjective and limited view
of the companies in the sales pipeline and the market itself, lead prediction approaches use a wide spectrum of data from various sources, which is
merged to an ideal outcome in a highly topical and highly dimensional deci-
5 AI Best and Next Practices
Fig. 5.1 Analogy to dating platforms
sion-making process. The features used can be divided into different groups
and they consider various aspects and properties of the suspects.
Under many different aspects, a comprehensive picture of every potential
lead is generated. The task at hand is the complete recording of the current
state and an estimation of the development of the company. This covers the
classical master data, an exact classification of the activity and an estimation
of the development of the company in its sector (Fig. 5.2).
Firmographics contain traditional company data that is taken from the company registry (name, location, commercial sector) and extended by further
indicators such as turnover and number of employees. The commercial sectors are a classification of the activities of companies that were published by
the Federal Statistical Office in 2008.
Fig. 5.2 Automatic profiling of companies on the basis of big data
Thanks to the dynamic identification of the subjects from the website, in
comparison with the commercial sectors, lead prediction achieves a very
accurate thematic classification and localisation of the company. I addition,
these tags have high topicality and new trends quickly become visible. In
comparison with the commercial sectors, instead of commercial sector software development, a company is given the tags app development, big data
or machine learning.
Word2Vec is used for the thematic classification of companies. Word2Vec
was released by Google in 2013 and is a neuronal network, which learns
the distributed representations of words during training. These vectors have
astounding properties and abstract the semantic meaning in comparison
with simple bag-of-word approaches. Words with similar meanings appear
in clusters and these clusters are designed such that some word relations
such as analogies can be reproduced under the application of vector mathematics, as in the famous example: “king − man + woman = queen”.
Via the Word2Vec presentation of texts, operations can be mapped; the
relationship of Apple to smartphones is identical to the relationship of dell
5 AI Best and Next Practices
5.1.6Digitality of Companies
The digitality of a company shows how far the company has completed the
process of digitalisation. Various aspects of digitality are included in this
score: the technology of the website, the visibility of the company on the
web, the ad spending and SEO optimisation and the degree of innovation
of the business model. On the basis of this score, companies can be easily
segmented depending on the degree of their digitalisation. Both young startups and established companies in the e-commerce sector are distinguished
by a higher-than-average digital index, whereby more traditional business
sectors reveal a rather less distinct degree of digitality. Table 5.1 shows the
individual dimensions of the digital index (Fig. 5.3).
5.1.7Economic Key Indicators
Key indicators from the investor relations environment are determined for
• Development of the staff: A stable or a growing number is a sign of a positive development of the company.
• Consumer activity: What is the situation in the individual commercial
sectors and how is the development estimated?
• Does the company pursue technological trends?
Table 5.1 Dimensions of the digital index
How much traffic does the site generate? How many users see the
site? Unique visitors, page views
Mobile readiness: Are the offers also designed or optimised for
SEO & advertising: Ads spending and SEO optimisation available?
Social media comprises:
• Social media readiness: How many channels is the company represented on?
• Social media activity: how active is the company on the social
How well is the company networked?
How does the user perceive the quality of the website? How fast
does the site load? How well-written re the texts?
How innovative is the company’s business model?
Fig. 5.3 Digital index—dimensions
Base on this spectrum of data, which is available in high topicality, the lead
prediction generates a presentation that summarises all aspects of the company in a 360° perspective.
The characterisation of the entire companies that should be used for lead
prediction is an essential step. On the basis of this generic customer DNA,
further companies are identified that have the same DNA (Fig. 5.4).
5 AI Best and Next Practices
Fig. 5.4 Phases and sources of AI-supported lead prediction
5.1.9Prediction Per Deep Learning
Deep learning is a subject that is causing quite a stir at the moment. In principle, it is a branch of machine learning that uses algorithms to recognise
objects, for example, and understand human speech. The technology is in
principle a revival of algorithms, that were popular from the beginnings of
AI: Neuronal networks. Neuronal networks are a simulation of the processes
in the brain whereby neurons and the specific fire patterns are imitated. The
real innovation is the layering of various neuronal networks which, in combination with the essentially greater performance of current computers, led
to a quantum leap in diverse sectors of machine learning.
The classifier for the prediction learns a generic DNA on the basis of profiling the successful customer relations, which is projected onto the entire
company’s assets. The prediction of the optima leads can be understood as
a ranking problem. The lead with the highest probability of a conversion
should be in first place in the sales pipeline. In principle, it can be understood as a classic regression task where the probability of conversion is to
be predicted. Thus highly suitable is a gradient boosted regression tree, also
called random forest.
5.1.10Random Forest Classifier
The algorithm gradient boosted regression trees, also called random forests,
belong to the ensemble learning methods This classifier uses an ensemble of
weak regression trees that have a low hit quota when considered in isolation. The quality of the prediction can be improved significantly when various trees are trained with different parameters or samples. The results of the
individual trees are aggregated to a total result which then enables a more
balanced and high-quality prediction. The so-called bagging triggered a
boom of the traditional regression trees. As aggregation, either a majority
vote or a probability function is chosen (Fig. 5.5).
The lead prediction generates high-conversion leads because
• The entire spectrum of information available about a company is integrated into the decision-making;
• The data is highly topical and without bias;
• The random forest is capable of abstracting complex correlations in the
• The method learns iteratively from the interaction with the sales team.
Fig. 5.5 Lead prediction: Automatic generation of lookalike companies
5 AI Best and Next Practices
The choice of leads is the first step in the sales process; the second one is to
find the ideal point in time for addressing them.
5.1.11Timing the Addressing
The right addressing, the right occasion and the right point in time—good
content marketing demands each individual aspect to be as successful as
possible. Various studies have shown that important purchase decisions
are made at certain occasions in life. When marketing and acquiring new
customers, a well-chosen point in time of addressing them is essential for
success or prospect of conversion. In a sales team, this is typically done intuitively on the basis of experience. How is this decision made if this knowledge has not yet been gathered? We have developed our own approach.
We scan the Internet for signals and this way, we are informed about economic changes in companies. Any mention of companies is analysed and
the impact they have is evaluated and whether they reveal a positive or a
negative development. A rapidly increasing number of complaints to a nonresponsive customer service can be an indication of internal problems within
the company. News, blogs, social media and the website are a highly topical
source of information about the condition and development of a company.
Scheduled relocations, structural changes, expansion strategies or profit
announcements, for example, are quickly visible and are a sign of a positive
or negative development of a company. On the basis of these “early signals”,
statements can be made about how probable a company will react to being
addressed at the current point in time.
Alerting openly scans the Internet and crawls cyclically websites and social
media channels for content snippets containing information about a company. These snippets are the potential alerts that are filtered and aggregated
according to significance down the line. In the first step, the probability of a
company being mentioned in the given text is determined. Sequence learners, which make a decision based on the lexical similarity and the context of
the word as to whether the mention refers to a company or not, are used for
In the second step, a deep learner decides whether the validated snippets
on a company trigger an alert or whether they are a part of daily background
noise. To this end, a model is trained on the basis of historic text data and
corresponding share developments, to recognise correlations between snippets and the development on the stock market. The time lag between alert
and real change “lag” is automatically learned by the system. Recurrent neural networks, in comparison with other approaches on the basis of a “sliding
time window” in combination with a classical regression, do not have the
limitation of the finite number of input values.
Subsequently, the system is in a position to make its own predictions
about the profit development of a company. These indicators are used in
lead prediction to choose those companies among those with a very similar
DNA that, at the current point in time, are most probably interested in an
evaluation of the business activities.
5.1.13Real-World Use Cases
220.127.116.11 Company: Network Monitoring
The spectrum of customers of these companies is diversified. Many of these
companies are located in the environment of information technology and
offer server hosting, for example. O the other side of the spectrum somewhat exotic companies emerged, such as operators of large production
plants, silos, chemical production plants, etc. A manual evaluation of the
leads from the lead prediction turned out to be difficult, meaning that we
decided in favour of A/B testing. In the sales process, the leads that were predicted by lead prediction scored higher than average and generated a 30%
higher conversion rate.
18.104.22.168 Company: Online Shop for Vehicles Construction
Two predictions were made with this project. The first one was aimed at
the regular customers, the second one at customers that did not belong to
the sales team’s general target group, but were acquired by chance instead.
The aim was to increase the market of this so-called alien group, in order to
enter a market segment that had not yet been defined in detail. The conversion rate improved by 40% in the classical segment; in the new segment, an
increase by 70% could even be measured
5 AI Best and Next Practices
22.214.171.124 Company: Personnel Service Provider
A clear lift can be recognised with this prediction case. Via classical list generation, seven appointments used to be generated from 700 telephone calls;
that is a conversion rate of one percentage. On the basis of the leads determined per lead prediction, there were nine appointments from 300 telephone calls; that is a conversion rate of three percentage. That is a significant
increase. However, it should be kept clearly in mind that this is quite a small
5.2Digital Labor and What Needs to Be
Considered from a Costumer Perspective
Alex Dogariu, Nicolas Maltry, Mercedes-Benz Consulting
In this use case by Mercedes-Benz Consulting the experience from thousands of real-life customer/Digital Labor interactions as well as numerous
customer research studies on Digital Labor are summarised and the necessity
for a centralised platform approach towards Digital Labor is elaborated on.
The landscape for customer management, customer experience, CRM and
customer service is changing rapidly due to the evolution in AI and its growing adoption in real-life use cases. One particularly rapid growing application of AI are Chatbots and Digital Assistants in customer interaction. The
trend towards the automation of work and the associated savings potential
in terms of workforce creates also great concerns, skepticism and fear among
the workforce and thus also in the public perception. In addition to contributions on the opportunities of digitalisation, AI and automation, there
is always talk about demands for regulations and guidelines. For example,
trade unions speak of “job killing by artificial intelligence”. The integration
of human and AI-based labor is thus very important, but won’t be further
discussed in this use case.
“Mercedes-Benz Consulting analysis of various contact centre data in the
automative sector revealed that 80% of customer inquiries are repetitive and
are rather simple in nature”. These 80% can thus be automated through digital labor. We usually refer to this as the “Fat Head and Long Tail” approach
When looking at current customer service operations, we can clearly identify plenty of opportunities for digital labor. The ordinary customer service
opening hours range from 8 a.m. to 8 p.m. This frustrates customers, espe-
Fig. 5.6 Fat head long tail (Source Author adapted from Mathur 2017)
cially if they have a specific concern or a problem with a purchased product or
service and want to get it dealt with immediately. Analysis of for example live
chat data has shown that most live chat requests happen between 8 p.m. and
11 p.m. on a daily basis and mostly on weekends on a weekly basis. The use of
Digital Labor is very useful here as it is available 24 hours a day, 7 days a week.
As machine learning, natural language processing (NLP) and robotic process automation evolve, Digital Labor will also be able to take on increasingly
complex tasks. For example, answering more difficult customer inquiries in a
personalised way or perform business tasks. The idea behind this is that digital labor will recognise the needs of the customer in advance and based on
previous behaviour, decisions and existing preferences, proactively engage in
conversations to help customers and promote products and services.
Currently, the clear majority of customer interactions with digital labor
include the ability to escalate to real agents or customer service staff. The
already mentioned agents are a valuable, expensive and limited resource
whilst customer requests to contact centres are steadily increasing. Therefore,
automatisation and self-service turns out as a promising option for contact
centre managers. Software that can reply to requests in a natural way (e.g. a
Chatbot) is a strategically important chance to handle rising costs and customer expectations (Aspect 2017).