AI Startup: Getting started and potential roads ahead

(1)

AI Startup: Getting started and poten- tial roads ahead

Case: Fibelius

Andreas Salonen

Bachelor’s thesis December 2020 School of Business

Degree Programme in International Business

(2)

Author

Salonen, Andreas

Type of publication Bachelor’s thesis

Date 21.12.2020

Language of publication:

English Number of pages

59 Permission for web publi-

cation: (x) Title of publication

AI Startup: Getting started and potential roads ahead Case: Fibelius

Degree programme

Degree Programme in International Business Supervisor(s)

Saukkonen, Juha Assigned by Konemieli Oy Abstract

Artificial intelligence has opened a broad range of opportunities for startups to develop unique products for their target market. Startups face a variety of different challenges during this development process and for this reason, the modern business development frameworks have been introduced to assess the startups and provide guidelines to follow in different stages of the startup’s evolvement. The objective of this study was to find potential roads to follow for a Finnish AI startup Fibelius with the help of selected business development frameworks. Fibelius is focusing to provide Natural Language Processing technology solutions to a specific service context within a public sector organization.

A qualitative approach and a case study strategy were selected to develop a comprehensive understating of the subject. The primary data of the study was collected through a Google Forms survey tool. The conducted survey gathered 22 responses from the profes- sionals of the public sector organization and Finnis investors. The collected primary data was then analyzed and compared to specific features of the selected frameworks to provide an understanding of the occurred themes and find similarities between them.

The research findings suggest that all the selected frameworks can offer potential roads for Fibelius to follow. As some have a broader overview of the principles while others offer a more detailed path description, they all can act as a benchmark for Fibelius to develop its operations. The results also indicated that Fibelius should pursue to the public sector market with NLP technology product and emphasize the importance of acting on the feedback from the customers’ point of view. The modern business development frameworks can help Fibelius in the future product development process by providing the correct tools and metrics to analyze the response from the customer and begin to process of learning as fast as possible to find a correct market solution fit.

Keywords (subjects)

Artificial Intelligence, Natural Language Processing, Fibelius, Lean startup, Startup genome, Zero to one, Startup

Miscellaneous (Confidential information)

(3)

Kuvailulehti

Tekijä

Salonen, Andreas

Julkaisun laji Opinnäytetyö

Päivämäärä 21.12.2020 Julkaisun kieli:

English Sivumäärä

59 Verkkojulkaisulupa myön-

netty: (x) Työn nimi

AI Startup: Getting started and potential roads ahead Case: Fibelius

Tutkinto-ohjelma

Degree programme in International Business Työn ohjaaja

Saukkonen, Juha Toimeksiantaja Konemieli Oy Tiivistelmä

Tekoäly on avannut monia uusia mahdollisuuksia startupeille kehittää ainutlaatuisia tuot- teita kohdemarkkinoilleen. Startup -yritykset käyvät läpi monenlaisia haasteita kehityspro- sessin aikana ja näiden haasteiden läpi käyntiä pyritään ohjaamaan liiketoiminnan viitekehyksien avulla, jotka tarjoavat liiketoiminnan arviointiin tarkoitettuja työkaluja ja ohjeita startupin eri kehityksen vaiheisiin. Tutkimuksen tarkoituksena oli löytää mahdollisia suun- tia toimintaansa aloittavalle suomalaiselle tekoäly-startupille valittujen liiketoimintaviite- kehyksien avulla. Fibelius keskittyy tarjoamaan luonnollisen kielen prosessointiteknologia- ratkaisuja tiettyyn palvelukontekstiin julkisen sektorin organisaatioille.

Tutkimus toteutettiin kvalitatiivisella, eli laadullisella menetelmäsuuntauksella ja tutkimuksen strategiana oli tapaustutkimus, joiden avulla pyrittiin kartoittamaan mahdollisimman kattava kuva tutkimuksen aiheesta. Tutkimuksen data kerättiin Google Forms -tutkimus- työkalun avulla, joka keräsi 22 vastausta. Kyselyyn vastanneet edustivat julkisen sektorin organisaatioiden ammattilaisia ja mukana oli myös suomalaisia sijoittajia. Kerätty data analysointiin ja verrattiin valittujen viitekehyksien erityispiirteisiin, joiden avulla pystyttiin luo- maan käsitys kohdatuista aihealueista ja niiden yhtäläisyyksistä viitekehyksiin.

Tutkimustulokset osoittavat, että kaikki valitut viitekehykset tarjoavat mahdollisia seuraa- via askeleita Fibeliukselle; Osa laajemmin ja osa kohdennetummin. Näin ollen viitekehykset voivat toimia luotettavina vertailukohteina Fibeliuksen toiminnan kehityksessä. Tulokset osoittavat myös, että markkinoilla on tarvetta luonnollisen kielen teknologiatuotteelle ja tässä tuotekehityksessä on korostettava asiakkaan palautteeseen perustuvaa reagointiky- kyä. Liiketoiminnan viitekehykset voivat tukea Fibeliusta tulevassa tuotekehitysprosessissa tarjoamalla oikeat työkalut ja mittarit asiakaspalautteen analysointiin ja oppimisprosessin aloittamiseen oikean markkinaratkaisun löytämiseksi.

Avainsanat (asiasanat)

Tekoäly, Luonnollisen kielen käsittely, Fibelius, Lean startup, Startup genome, Zero to one, Startup Muut tiedot (luottamuksellista tietoa)

(4)

Contents

1 Introduction ... 3

1.1 Context and motivation for the research ... 3

1.2 Fibelius – A Natural Language AI Startup ... 4

1.3 Research objectives and questions ... 6

2 Literature review ... 7

2.1 Startup as a company category – Definitions and Specificity ... 7

2.2 Artificial Intelligence ... 11

2.3 Specific nature of AI-Startups ... 13

2.4 Natural language processing (NLP) ... 16

2.5 Frameworks of modern business development ... 18

2.5.1 Lean startup ... 19

2.5.2 Startup genome ... 21

2.5.3 Zero to one ... 25

2.6 Summary of the Literature Review ... 28

3 Methodology ... 30

3.1 Research approach and strategy ... 30

3.2 Data Collection ... 31

3.3 Data Analysis ... 32

3.4 Plan for research quality and ethics ... 34

4 Results ... 35

4.1 Product demand ... 36

4.2 Customer-specific features ... 40

4.3 Availability of the resources ... 44

(5)

5 Conclusion and discussion ... 46

6 References ... 50

Appendices ... 54

Appendix 1. Survey invitation letter ... 54

Appendix 2. Survey questions ... 55

Figures Figure 1 Funding of Finnish startups (Finnish Venture Capital Activity, 2019) ... 10

Figure 2 Techonology Adoption Life Cycle (Moore, 2002) ... 15

Figure 3 Five steps in computational processing (Dale et al., 2000, 10) ... 18

Figure 4 Build-Measure-Learn Feedback Loop (Ries, 2011, 79) ... 20

Figure 5 Key Challenges by Stage (Marmer et al., 2011b, 23) ... 24

Figure 6 Product demand ... 38

Figure 7 Customer-specific features ... 42

Figure 8 Availability of the resources ... 45

Tables Table 1 Overview of Results (Marmer et al., 2011a, 7) ... 23

Table 2 Examples for inconsistency (Marmer et al 2011b, 11) ... 25

Table 3 Primary tenets of the frameworks ... 29

Table 4 Research stages and timeline ... 35

Table 5 AI Startup: Getting Started ... 36

(6)

1 Introduction

1.1 Context and motivation for the research

The continuous rise of AI technology has opened limitless possibilities for technology startups to achieve hyper-growth targets which is the ultimate goal of a technology startup or at least it should be. The startup is a company designed to grow fast and the growth is the thing that strives startups forward (Graham, 2012). But predicting the high growth potential can be challenging or even impossible if the product/service is determined to find a solution for a new occurred problem. The challenges and continuous uncertainty often cast the shadow which surrounds the entire operation, people, product or the market. The modern business development frameworks are developed to help the startups to clear out the uncertainty and to understand if the startup is heading in the right direction. These frameworks can provide the necessary tools for a startup to learn how to steer, when to turn and when to preserve-and grow a business with maximum acceleration (Ries, 2011). The frameworks also present guidelines on how to introduce your product/service to the market and gather valuable data of the response to learn and reduce all the unnecessary waste around this process to save resources.

According to Berryhill et al. (2019, 3), artificial intelligence is an area of research and technology application that can have a significant impact on public policies and services in many ways. In just a few years, it is expected that the potential will exist to free up nearly one-third of public servant’s time, allowing them to shift from a mun- dane task to high-value work. The need for adaptation within the changing environment is crucial for the public sector which has opened a broad range of beneficial opportunities for AI solutions. According to a survey conducted by Organization for Eco- nomic CO-operation and development (OECD), 50 countries have launched or are planning to launch a national artificial intelligence program (Berryhill et al., 2019, 7).

The bigger companies have already started to offer AI systems to public sector organizations, but could the smaller startups have something to offer that the bigger companies cannot?

(7)

The motivation for this research comes from the case startup and from the other AI startup’s perspective to gain knowledge of artificial intelligence and its possibilities in public sector organizations. The research for artificial intelligence from a business perspective is very needed because the technology is used to an increasing extent and the business field is still open in a sense that there are relatively few technology providers compared to market potential. Although, the artificial intelligence is a significant phenomenon for this research, the focus is more on the business perspective. For the high technology startups, the business and customer perspective tend to come after the technology on a prioritization list which could lead to the situation where the startup introduces the long-produced innovation and there is no real need for the product/service. This research intends to prevent this possibility from occur- ring in the case company operations.

1.2 Fibelius – A Natural Language AI Startup

Fibelius was incorporated in Helsinki, Finland in July 2020 as Konemieli Oy. Fibelius is an early-stage enterprise AI startup helping businesses and government organizations to make their customer interaction dialogues and internal communication more intelligent and more automated by adding NLP-AI into the communication channels that are feeding the enterprise robotic process automation systems. The company has got its first government customer and is now working on its first customer funded project. Natural language processing (NLP) is a branch of artificial intelligence and it is used to aid computers to understand human’s natural language. The objective of NLP is to read, decipher, understand and make sense of the human languages in a valuable manner.

To be able to enter the S-shaped growth curve of that business as a first-mover Fi- belius aims to build a network of natural language processing and intelligence aug- menting dialogue robots with a specific focus on the natural and professional languages that are spoken in Finland. While the Fibelius vision itself, add Machine Intel- ligence to Human to Human and Human to Machine interaction and communication

(8)

is something that will most probably happen during the next 10-15 years, there are at least two basic scenarios of how that will happen.

The first one of those scenarios “AI is the game that only tech giants (or govern- ments) can play” underlines the fact that was recently demonstrated by OpenAI with its gigantic 175 billion parameter and 300 billion token (words) GPT3 language model and its API pricing scheme (Brockman et al., 2020). This scenario emphasizes the huge size of the initial investment one needs to train state of the art near-human level neural networks.

The second scenario is the long tail scenario that Andreessen Horowitz Venture Capi- tal firm A16Z also calls as the Decentralized Autonomous Hiveminds scenario. This scenario starts with the long tail data variation thesis which states that a large data corpus does not offer large tech giants a defensible moat against the competition (Yahya, 2020). Especially in enterprise AI a startup can start with a minimum viable data corpus and bootstrap models and a data journey by capturing the initial data corpus from many sources.

The source of sustainable competitive advantage

One way to measure the competitive advantage of business context focused language models is to measure the expressive power of the language model. Fibelius measures the expressive power of the text corpora and models and tests the intelligence of a trained language model by measuring if its expressive power gives the user the capability to perform the tasks and achieve the goals he is supposed to do when using the advisor user interface of the language model. Fibelius calls the intelligence that can be generated with this kind of neural network training Expressive Neural Network Intelligence or ENNI.

The ENNI is a neural network that takes as input expressions of natural and professional languages (text, speech, CAD pictures) and outputs text. The backpropagation, the reward mechanism and the neural network weight system of ENNI is focused on maximizing the expressive power of the output. The expressive power of a language is the breadth of ideas that can be represented and communicated in that language.

(9)

Considering the sustainable competitive advantage, the key finding here is: Fine- tuned ENNIes can be moated (the expressive power is a special type of bias the com- ponents of which are located within and between the neural network layers and nodes), they offer a sustainable source of competitive advantage and they offer a way to monetize the development and innovations by using a simple freemium-pre- mium-corporate customer contact us earning model.

1.3 Research objectives and questions

The objective of this research was to gain knowledge of how the modern business development frameworks could help to clear out the path from uncertainty and give tools to build the product/service that could help to stand out from the crowd and where bigger software companies are operating. It was also important for this research to understand the target market and their real need for the AI systems ena- bling to create strategies on how to enter the target market. For Fibelius the tactical and operative side of the business is in a very early stage and with this research, we want to find potential roads Fibelius to follow lighted by business frameworks. The research questions were formulated to guide the research process based on the research objectives:

• What could be suitable framework to follow for Fibelius?

• How could that framework help to develop AI-startup business like Fibelius?

• What would be the next steps for Fibelius to start implementing the framework?

The approach of the study is a qualitative case study to study complex and diverse phenomena within its real-life context to gather in-depth insights on the topic. The primary data of this study is collected through a google forms survey tool by which the public sector decision-makers and influencers gave their views on the subject.

The methodological choices of the study are presented in more detail in chapter 3.

Methodology. This study analyzes and compares three selected business

(10)

frameworks: Lean startup by Eric Ries, Startup genome by Max Marmer, Bjoern Lasse Herman, Ertan Dogrultan and Ron Berman, Zero to One by Peter Thiel. The presented frameworks share their vision of technology startups and their uniqueness. The startups are not a smaller version of larger companies and these frameworks are aimed to clarify this entirety and their stages of development (Blank, 2014).

2 Literature review

The literature review presents the key concepts that are covered within this research and also acts as a background for the research questions. The main sources of literature that have been used to provide the literature review are Google Scholar, Janet Finna and diverse selection of business articles. There are various types of literature from every concept that is presented in this literature review, but since the objective is to focus on a case company Fibelius the purpose is to give a comprehensive point of view to these concepts from a business perspective. The presented concepts are Startup as a company category – Definitions and Specificity, Artificial Intelligence, Specific Nature of AI-Startup, Natural Language Processing (NLP), Frameworks of modern business development and finally the summary of the Literature review which provides a summary of the key concepts and connects the presented literature to the research questions.

2.1 Startup as a company category – Definitions and Specificity

“A startup is the largest group of people you can convince of a plan to build a different future” (Masters & Thiel, 2014). This is how Peter Thiel has pictured the startup world in his book zero to one. He states that many companies are starting straight from stage 1, which means that something is already been invited. There is not going to be a new operating system by Bill Gates or a new social network by Mark Zucker- berg. Every moment in business happens only ones. Therefore, the future depends more on new innovations than on copying and extending old practices. World-changing innovations require startups to build the path there.

(11)

The term startup has multiple definitions and outlines that make it difficult to find one universal definition that connects them all. The term itself boomed in the late 1990s when the Internet became ready to be adopted outside its early military and university networks. For example, Peter Thiel founded PayPal together with Elon Musk when IT-companies started emerging at a record phase. This trend originated the strong belief and ideology that the Internet is the next world-changing techno- logical innovation. The most important of these innovations are what economists call general-purpose technologies a category that includes the steam engine, electricity, and the internal combustion engine which have more than 250 years to be the main drivers of economic growth (Brynjolfsson & McAfee, 2017). Investors saw new general-purpose technology and started investing in anything related to the internet. To- day this is known as the Dot-com bubble which lasted a relatively short time but will always be remembered as a Silicon Valley gold rush.

Even today the “change the world” mindset is strongly involved within startup com- munities. Words that are often heard in startup scenes are: It’s not about what’s- now but what’s-next. Startups probe for new possibilities, they examine what else needs to be done and then launch a path for that destination. Thinking like startup positions us to think aspirationally about change. It requires and rewards innovation and creativity (Brian, 2012, 4).

Eric Ries and Steve Blank are considered to be the most significant influencers in the emergence of the current startup culture and their publications are often referred to when having a conversation about startups. Eric Ries is an entrepreneur and author of the Lean Startup movement and Steve Blank is a consulting professor at Stanford University and also a well-known entrepreneur. According to Blank, the startup is a temporary organization designed to search for a repeatable and scalable business model (Blank, 2010). This theory of a temporary organization stage is arguably the most used definition from the term startup.

For a long time, the general idea was to treat and think of the term of a startup as a smaller version of a bigger business. Today the business world recognizes these as

(12)

two very different organizations whose operations need to be evaluated and developed in completely different ways (Blank, 2014). When comparing startups to a bigger company what makes a company a startup is that the startup does not yet know who their customers are or what their product should be (Ries, 2015, 9).

Despite the uncertainty, the startup is a company designed to grow fast (Graham, 2012) everything else follows from growth. Growth can be challenging for many startups, but growth is the thing that strives the startups forward. The need to grow surge from the funding base of such companies invested assets demanding a steeply rising curve of returns to balance the high uncertainty (Saukkonen, 2020, 67). How- ever, growth is also the thing that separates startups from the other “normal” companies that are established in the world. Startups are different by nature, for example, barbers or bakeries. It can be argued that these companies are not aiming for the rapid and exponential growth that startups aim for.

Ries in his book the lean startup states that a startup is a human institution designed to deliver a new product or service under conditions of extreme uncertainty (Ries, 2015, 8). A Startup that has faced extreme uncertainty during its development was Rovio. In 2009 after 51 games and a near bankruptcy the founders pursued to make one more game and have one more shot to build a successful mobile game. In eight months, they built Angry Birds. The game went viral and became one of the biggest mobile games and entertainment brand success stories of all time (About-Rovio, 2020). Despite the immense uncertainty what Rovio founders encountered they found themselves achieving something that many startup entrepreneurs are dream- ing to accomplish. The story illustrates the field of uncertainty and win or lose environments where startups operate.

Startups in Finland

Finland has a strong startup culture and ecosystem with accelerators, angel investors, VCs and strong innovation support by the government. The capital Helsinki ranks number one in the world in local connectedness among founders, investors and experts (Business Finland, 2020). Funding of Finnish startups has more than quadrupled in nearly ten years and Finnish startups are attracting more foreigner

(13)

investors around the world. The number of foreign investments has increased ten- fold since 2010. Figure 1 visualizes the growth of funding on Finnish startups and early-stage companies. The thing to pay attention to here is the rise of Foreign interest to Finnis Startups.

Figure 1 Funding of Finnish startups (Finnish Venture Capital Activity, 2019)

The Finnish startup scene can be considered to be high on the list when speaking of the most attractive startup scenes in Europe. This can be explained by the successful gaming and software industry led by companies like Rovio or Supercell (Business Fin- land, 2020). Successful industry spreads interest for other areas of expertise and raises interests to establish startups.

On the other hand, according to the global startup genome report (2020, 27) Finnish startup ecosystems are struggling to enter the top 30 at Global Startup Ecosystem Rankings and for example, Stockholm has made it into the top 10 on the list. This can

(14)

be used as a scale of where Finnish startups fit on a global market and how to improve Finnish startup ecosystems to get on the global rankings.

2.2 Artificial Intelligence

The same difficulty of term explanation applies to Artificial Intelligence (AI) as for the term Startup. Generally, there is no single definition of artificial intelligence due to the difficulty of the definition of intelligence itself (Raskulla, 2019, 247). The term can be divided into “artificial” and “intelligence”. In the words of Berryhill et al (2019) The artificial aspect of AI is quite straightforward: it refers to anything non-natural and, in this case, non-man made. It can also be represented through the use of terms such as machines, computers, or systems. Intelligence is a much more widely dis- puted concept, explaining why there is as yet no consensus on how to define AI, even among experts.

That’s why definitions of AI tend to sift based upon the goals that are trying to be achieved with an AI system. In this research, the artificial intelligence is used in a context where AI is defined as computer programs that enable machines, devices, programs, systems, and services to operate intelligently appropriately in a complex and partially unpredictable environment (VM 2018, 5). In this definition, the difference is thus the relative independence of artificial intelligence. Artificial intelligence is artificial intelligence if it has the ability to evolve from experience, as well as perform tasks without constant guidance (Raskulla, 2019, 248). These approaches are semi- supervised and unsupervised machine learning approaches where the AI program learns only by little human guidance or not at all.

A brief history of AI

The AI was first introduced to the world in 1956 by the American computer scientist John McCarthy. The 1956 Dartmouth Artificial Intelligence conference gave birth to the field of AI and gave succeeding generations of scientists their first sense of the potential for information technology to be of benefit to human beings in a profound way. The Field of AI was launched not by agreement on methodology or choice of problems or general theory, but by the shared vision that the computers can be

(15)

made to perform intelligent tasks. This vision was stated boldly in the proposal for the 1956 conference: “The study is based on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it ” (Moor, 2016). According to Warwick (2011) in 1980 and 1990s the AI saw a whole new approach, a short of bot- tom-up attack on the problem, effectively building artificial brains to bring about AI.

This completely opened up the possibilities and created a whole new set of questions. No longer was AI restricted to merely copying human intelligence – now it could be intelligent in its own way.

AI has significantly changed our everyday life. In the future, the role of artificial intelligence will only grow, and it will be utilized in increasingly challenging tasks, for example in the medical industry or in lawyer professions. We also know that AI is no longer based on a computer system as we know, but rather on a biological brain that has been grown afresh (Warwick, 2019, 10). However, the raise of AI is not that straight forward. The adoption of AI services can be a tedious task and expensive to implement. Especially for organizations that are driven by complex technologies and high amounts of data. The implementation process requires talent and resources from the organizations. It is also important to align the company structure and the culture to support the process because it can be long time when the tangible return of investment can be measured. Different restrictions and laws for example GDPR are not helping the implementation process either.

Modern AI approach

The rough comparison between Modern AI and Classic AI is that the classic AI in- spects the brain from outside and tries to mimic actions and performance in the AI system. This is successful when dealing with well-defined tasks from which a set of clear rules are appropriate. Whereas the modern AI focus more on defining how humans communicate with each other from the brain to another under circumstances in which each brain processes large, highly similar surrounding metal structures that serve as a common context (Nilsson, 2014, 2). It has been difficult to develop a computer system that can generate or even understand fragments of a natural language,

(16)

such as English for example. The thing which makes it even harder to mimic conversation is that the conversation has usually signals or expressions that are left unsaid.

Here it is needed to have an AI system that can jump into humans’ brain not try to imitate from outside.

If we are going to say that a given AI program thinks like a human, we must have some way of determining how the human brain works. The human brain has a basic cell a nerve cell that is also called a neuron. In words of Warwick (2019, 90) In a typi- cal human brain, there are about 100 billion of these. Each of these neurons are very small, usually being 2– 30 micrometers in diameter (one-thousandth of the size of a small coin). The neurons are joined to form an extremely complex network, each neuron having upwards of 10,000 connections. Each neuron consists of a cell body with a nucleus at its center. Several fibers, called dendrites, stimulate the cell body with signals from other neurons.

Each experience from outside of the brain functions differently depending on a situation and the pattern of signals that it is receiving. The brain adapts and for each situation, it is more likely to make a similar choice when it faces a similar situation and vice versa if the brain detects something that it needs to react differently it makes the adjustment and it is harder to happen again. This is a basis of biological brain growth and development for which the concept of artificial neural network (ANN) is built on. ANN is a powerful AI tool that can, for example, make cars drive autono- mously, read our minds and recognize our speeches.

2.3 Specific nature of AI-Startups

As mentioned earlier startups operate in an uncertain field where the ground shakes easily. The constantly changing AI-field and the uncertainty of a startup’s success combined can make a head scratch. Should the startup pursue venture capital or bootstrap the business by focusing on profitability from the beginning? It all starts from the need of the customer and how well the program can respond to that need.

In a nutshell, without the need, there are no businesses to grow for any startup.

(17)

AI-startups often fumble to move forward with a unique AI-program and to present it to the market in a way that attracts customers from the beginning. The cause of this can be in a dot-com bubble where the startups often operated in a “stealth mode” and refused to present their inventions to the market because it could alert potential competitors for a market opportunity (Blank, 2013). However, the faster the product is presented in the market the faster product awareness grows. Getting early user feedback and incorporating it to the next version is crucial for startups.

Geoffrey Moore in his book: “Grossing the Chasm” showcases the technology adoption lifecycle and what separates successful high-tech companies from the others.

According to Moore (2002, 5) every truly innovative high-tech product starts as a fad—something with no known market value or purpose but with “great properties”

that generate a lot of enthusiasm within an “in a crowd.” That’s the early market.

Then comes a period during which the rest of the world watches to see if anything can be made of this; that is the chasm.

Figure 2 showcases the technology adoption life cycle from the customers point of view and how the emerging technology affects the different customer groups in different ways. The model illustrates the first group to use the products as “innovators”

followed by “early adopters” next “early majority” and the “late majority” the last group to get hands-on the new product are the “Laggards”. Moore (2002) describes each group briefly:

Innovators pursue new technology products aggressively. They sometimes seek them out even before a formal marketing program has been launched. This is because technology is a central interest in their life, regardless of what function it is perform- ing.

Early adopters, like innovators, buy into new product concepts very early in their life cycle, but unlike innovators, they are not technologists. Rather they are people who find it easy to imagine, understand, and appreciate the benefits of a new technology, and to relate these potential benefits to their other concerns.

(18)

The early majority share some of the early adopter’s ability to relate to technology, but ultimately, they are driven by a strong sense of practicality. They know that many of these newfangled inventions end up as passing fads, so they are content to wait and see how other people are making out before they buy in themselves.

The late majority shares all the concerns of the early majority, plus one major additional one: Whereas people in the early majority are comfortable with their ability to handle a technology product, should they finally decide to purchase it, members of the late majority are not.

Finally, there are the laggards. These people simply don’t want anything to do with new technology, for any of a variety of reasons, some personal and some economic.

Figure 2 Techonology Adoption Life Cycle (Moore, 2002)

This model is important to understand when developing a unique AI product to

“cross the chasm” from the early market to a more mainstream pragmatic market.

Each of these groups can be used as an example to the next group and the idea is to keep the flow through the whole model from left to right. The current customer groups for fibelius are more in the left side of the model.

(19)

Major technical breakthroughs in AI have raised the awareness of AI-Startups. Ac- cording to Brynjolfsson and McAfee (2011), we are entering a “race against the machine” where we can’t win the race, especially as computers continue becoming more powerful and capable. But we can learn race with machines, using them as al- lies rather than adversaries. AI-startups are solving a real puzzle on how to use AI for large and important problems but at the same time dealing with ultimate uncertainty. This applies straight to case startup Fibelius. Fibelius is trying to find ways to enter the market and attract customers which generates cashflow.

2.4 Natural language processing (NLP)

Fibelius is focusing on one such AI sub-category called Natural language processing (NLP) which is increasingly used in business environments. It enables computers to understand, process and create natural language. Over the past few years, neural networks have re-emerged as powerful machine-learning models, yielding state of the art results in fields such as image recognition and speech processing. More recently, neural network models started to be applied also to textual natural language signals, again with very promising results (Goldberg, 2016, 1). NLP processing works through machine learning systems (ML). Machine learning systems store words, sentences and sometimes entire data sets. The computer systems that analyze the data find a pattern and make a prediction of what comes next.

Arguably the most used NLP tool is the chatbot which can automate the customer service and positively the business experience. Overall, technologies based on NLP are becoming increasingly widespread. For example, phones and handheld computers support predictive text and handwriting recognition; web search engines give access to information locked up in unstructured text; machine translation allows us to retrieve texts written in Chinese and read them in Spanish (Bird et al., 2009, 9). Every year Gartner Hype Cycle evaluates the latest AI inventions and provides a graphic representation of the maturity and adoption of technologies and applications. NLP has been involved multiple times on the list as one of the most important AI branches which will play an important role in the future.

(20)

According to Dale et al in their book: Handbook of natural language processing (200, 9) Any computational system that uses natural language input and output can be seen in terms of a five-step processing sequence (Figure 2).

1. The system receives a physical signal from the external world, where that signal encodes some linguistic behavior. Examples of such signals are speech waveforms, bitmaps produced by an optical character recognition or handwriting recognition system, and ASCII text streams received from an elec- tronic source. This signal is converted by a transducer into a linguistic representation amenable to computational manipulation.

2. The natural language analysis component takes the linguistic representation from step 1 as input and transforms it into an internal representation appropriate to the application in question.

3. The application takes the output from step 2 as one of its inputs, carries out a computation, and outputs an internal representation of the result.

4. The natural language generation component takes part of the output from step 3 as input, transforms it into a representation of a linguistic expression, and outputs that representation.

5. A transducer takes the output representation from step 4 and transforms it into a physical signal that, in the external world, is interpretable by humans as language. This physical signal might be a text stream, the glyphs in a printed document, or synthesized speech.

(21)

Figure 3 Five steps in computational processing (Dale et al., 2000, 10)

According to a report by market research company Tractica (Tractica, 2017), NLP is a fast-growing industry and the global market value of related software is estimated to increase from 136 million in 2016 to 5,4 billion by 2025. NLP software deployments will drive significant additional sales of hardware and professional services, bringing the total NLP software, hardware and service market opportunity to 22,3 billion by the end of 2025.

2.5 Frameworks of modern business development

To clear out the fog of uncertainty there has been developed Frameworks to guide and asses’ startups by measuring thresholds and milestones that startups face. Ac- cording to Ries (2011, 9), the old management methods are not up to the task. Plan- ning and forecasting are only accurate when based on a long, stable operation history and relatively static environment. Startups have neither. Therefore, Frameworks are there to save time for providing the starting point and basic guidelines to follow.

It is possible that there is no best framework for business to follow and it can be that combining them can bring the best results or neither. In the following sub-chapters frameworks, Lean startup, startup genome and zero to one are presented.

(22)

2.5.1 Lean startup

The lean startup methodology has its origin in the Toyotas production system (TPS).

The most important objective of the Toyotas system has been to increase production efficiency by consistently reducing waste (Ohno, 1988, 12). Toyota did not copy the normal mass production approach during its time but originated production system which has its basis in continuous improvement loop and just-in-time production system. Toyota discovered that small patches made their factories more efficient and generated less waste. In contrast, in the Lean Startup, the goal is not to produce more stuff efficiently. It is to as-quickly-as-possible learn how to build sustainable business (Ries, 2011, 192).

Eric Ries wanted to have a new approach to the building process of IMVU company due to his bad experiences with old-fashion strategies that produces useless waste.

He adopted the Lean manufacturing from Toyota and developed the lean startup framework which is based on the build-measure-learn feedback loop (figure 3). It measures which actions are creating value to the business which is wasteful. Lean thinking defines value as providing benefit to the customer; anything else is waste (Ries, 2011, 52)

(23)

Figure 4 Build-Measure-Learn Feedback Loop (Ries, 2011, 79)

The startup needs to have a hypothesis on how it can produce value to the customer and how it can grow before it enters into the loop. Ries describes these as a “leap of faith assumptions”

The feedback loop starts from the building process when ideas become reality. The build stage must be reached as quickly as possible from the ideas phase with a minimum viable product (MVP). According to Ries (2011), the MVP stands for the version of the product that enables a full turn of the Build-Measure-Learn loop with a minimum amount of effort and the least amount of development time. MVP is used to gather data from the things that matter the most and start the learning process as soon as possible. Its goal is to test fundamental business hypotheses (Ries, 2011, 97)

Measurements are a crucial point to determine if the development of the product is leading to real progress. Ries describes a tool innovation accounting that can be used to assess development accurately and objectively. Ries emphasizes the importance of analyzing the right metrics that are important for the whole process. To analyze every nut and bolt is not needed.

(24)

Leading to the last stage of the loop and the most important stage before exiting the feedback loop is to decide to proceed with the current strategy or chance the direction called the pivot. Ries describes the pivot stage as follows (Ries, 2011, 82) upon completing the Build-Measure-Learn loop, startups confront the most difficult question any entrepreneur faces: whether to pivot the original strategy or persevere. If the startups discovere, that one of the hypotheses is false, it is time to make a major change to a new strategic hypothesis.

The lean startup has increased its follower base and become widely known method to reduce waste surrounding startups and keeping the focus at the most important aspects of the development process. It is taught at more than 25 universities and also in almost every city around the world, you’ll find organizations like Startup Weekend in- troducing the lean method to hundreds of prospective entrepreneurs at a time (Blank, 2013).

2.5.2 Startup genome

The startup genome framework was introduced by Marmer, Hermann, Dogrultan and Berman to effectively assess startups. The framework was established as part of the bigger Startup Genome Report to which more than 650 web startups have partic- ipated to answer the survey and later over 3200. The main goal of the report was to increase the success rate of startups and accelerate the pace of innovation around the world by turning entrepreneurship into a science (Marmer et al., 2011a). Re- searchers found out that there is a lot of confusion among founders of the startups where to focus and when to pivot. Startup Genome authors recruited Steve Blank to help in the framework establishing process. For this reason, the framework has adopted influences from the Blank’s Customer Development framework.

Marmer et al (2011a, 6) had three core ideas that the report aimed to answer:

1. Startups evolve through discrete stages of development. Each stage can be measured with specific milestones and thresholds.

2. There are different types of startups. Each type evolves through the developmental stages differently.

(25)

3. Learning is a fundamental unit of progress for startups. More learning should increase the chances of success.

According to Marmer et al (2011a, 6), there are six different stages that startups face. From which the first four are loosely Steve Blank's 4 Steps Customer Develop- ment Model, but one key difference is that the Marmer Stages are product-centric rather than company-centric. The last two stages are focusing on life after the startup phase which are not covered within this report and this can be considered as one of the weaknesses of this framework. The stages and the average time to com- plete in the first four stages presented below:

1. Discovery (5-7 months) 2. Validation (3-5 months) 3. Efficiency (5-6 months) 4. Scaling (7-9 months) 5. Sustain

6. Conservation

At the discovery stage startups are aiming to discover whether they are solving a meaningful problem and whether anybody would hypothetically be interested in the solution. At validation stage startups get early validation that people are interested in their product through the exchange of money or attention. Efficiency stage startups refine their business model and improve the efficiency of their customer acquisition process. At the scaling stage startups step on the gas pedal and try to drive growth very aggressively (Marmer et al., 2011a, 14-15).

The startup genome report also identifies four different types of internet startups which are: the automizer, the social transformer, the integrator and the challenger.

Types of startups have all different characteristics and they move through the stages

(26)

differently. For example, the challenger startups take significantly longer to move through the stages than others (Marmer et al., 2011b, 30)

Table 1 Overview of Results (Marmer et al., 2011a, 7)

Table 1 illustrates the overview of the results from the first four stages that the startup genome report gathered. The stages are compared in six different sections which are the average months working, average funding raised, the average number of employees, average % user growth in last month, top competitive advantages and top challenges. The key to point out from the graph is the key challenges that

startups face during the stages (Figure 5). In the validation stage product-market fit seems to spike up and customer acquisition is the biggest challenge of them all. How- ever, in stages 1-3 startups shouldn’t be directly focused on customer acquisition.

Challenges like problem solution fit, Product market fit, and feature development are more actionable challenges that treat the root cause of the lack of customers

(Marmer et al., 2011b, 23). One thing also notable to consider is that hardly anyone of the startups believed that their product would be their main challenge.

(27)

Figure 5 Key Challenges by Stage (Marmer et al., 2011b, 23)

Marmet et al (2011b, 10) define startups in their report as a developmental organism that evolves along 5 independent dimensions: Customer, Product, Team, Business Model and financials. In this regard, researchers discovered a phenomenon described as a premature scaling. According to startup genome research, 70% of the startups scaled prematurely. The startup which can balance these five dimensions evenly are called consistent and the ones which cannot go through all of the stages and keep dimensions in a balance are called inconsistent. The Inconsistent startups are scaling prematurely and are less likely to gain success than consistent ones. Most startup failures can be explained by one or more of these dimensions falling out of tune with the others (Marmer et al., 2011b, 10). Examples of inconsistency are presented in table 2 below.

(28)

Table 2 Examples for inconsistency (Marmer et al 2011b, 11)

The startup genome framework is under construction and it will be developed over time to a high-quality framework. However, it already provides a holistic view of the different types of startups and various stages that startups face. The quantitative data provides startups the right tools to move forward and basic guidelines to pay attention to.

2.5.3 Zero to one

Zero to one methodology is about startups that can change the world. It has been developed by co-founder of PayPal, entrepreneur and venture capitalist Peter Thiel.

The co-author Blake Masters was a student at Sandford Law school in 2012 when he made detailed notes on Peter’s class “Computer Science 183” which became an internet sensation. According to his website (blakemasters.com), he is now a chief

(29)

operating officer at Thiel capital and the president of the Thiel foundation. Thiel has learned how to build companies that create new things through his past experiences as an investor in hundreds of startups including Facebook and SpaceX (Masters &

Thiel, 2014, 5). Thiel has noticed a consistent pattern from his experiences that successful people make value from unexpected places and they do this by thinking about the business not formulas.

The contrarian question that Thiel asks in his book is: what important truth do very few people agree with you on? (Masters & Thiel, 2014, 11). Steve Jobs did know the truth of how people would use smartphones in the future. He knew that we would not want to use the physical keyboard but everyone else disagreed with him at that time. This is a good example of a contrarian truth. The best contrarian truth will re- veal how people will act in the future (Masters & Thiel, 2014, 11). The business version of contrarian truth is: what valuable company is nobody building? The best startup is built on an idea that would not be possible three years ago.

The key is to bet in a contrarian truth that very few businesses are focusing and go from zero to one, to achieve a monopoly position. According to Masters & Thiel (2014, 26) monopoly is the condition of every successful business. Startups should not go to a market where is the “perfect competition”. Under perfect competition, in the long run no company makes an economic profit (Masters & Thiel, 2014, 19). In other words, they are aiming to go out of the business. The company that can focus on a niche market and creates a monopoly can help to generate enough cash flow to extend the business to another market over time.

Thiel describes a four-common characteristic for a monopoly:

1. Proprietary technology

Propriety technology makes the product difficult or impossible to replicate.

Google’s search algorithms, for example, return results better than anyone else’s (Masters & Thiel, 2014, 36). Google made a search engine that is at least 10x better than any other search engine.

2. Network effects

(30)

The network effects make a product or a service more useful the more it gathers customers. Facebook, for example, started from a university environment where Mark Zuckerberg wanted to have everyone in his class to have access to a social network. Then more people started using the same network and it is then likely that people will switch from the platform if their friends are there too.

3. Economics of scale

Amazon has a relatively low shipping costs compared to competitors because of their economic of scale. They are operating in such customer quantity that the shipping cost is not a large expense on the scale of the whole operation. A monopoly business gets stronger as it gets bigger (Masters & Thiel, 2014, 38).

4. Branding

Apple has acquired a strong brand that is associated with fashion and style.

Everyone wants to buy Apple products for this reason, and they are the first brand that people considered when they think of a fashionable brand. There- fore, it is hard for any other brand to compete in a market where Apple already has conquered the customer’s minds.

The basic of zero to one is to build a monopoly in a niche market and scale up after a monopoly position has accomplished. The product or service solution needs to deliver 10x speed, performance or convenience benefit over the status quo to shift con- sumer behavior within the market. the key to success is to avoid value competition at all costs. The startup can protect the monopoly position by focusing on common characteristics of a monopoly.

Zero to one is not providing a step-by-step manual, because every business is different or at least, every business should be different (Thiel & Masters, 2014). It gives a model to create something new, different and valuable that it creates a monopoly in a new market and goes zero to one.

(31)

2.6 Summary of the Literature Review

Startups are the heart of the development of our society. “A startup is the largest group of people you can convince of a plan to build a different future” (Masters &

Thiel, 2014). Despite the uncertainty, startups are meant to grow fast and go from zero to one to build a new route for humanity to follow. This will only work if

startups are aiming to build a monopoly position in a niche market. A startup is a human institution designed to deliver a new product or service under conditions of extreme uncertainty (Ries, 2015, 8). Without the plan and guidelines, startups are des- ignated to fail before the operations have even started to which this research seeks to prevent from happening.

There is no single definition of artificial intelligence due to the difficulty of the definition of intelligence itself (Raskulla, 2019, 247). In this research, the artificial intelligence is used in a context where AI is defined as computer programs that enable machines, devices, programs, systems, and services to operate intelligently appropriately in a complex and partially unpredictable environment (VM 2018, 5). AI is in a transition to overtake human capabilities and it is increasingly used in the business world which has opened possibilities for new technology startups to evolve.

Artificial intelligence can collect and organize data that are beyond human capabilities in normal manual processing. In the near future, between 70% and 90% of all initial customer interactions are likely to be conducted or managed by AI (Aquis, 2019).

Fibelius is focusing on one AI sub-category called Natural language processing (NLP).

In a nutshell, NLP enables computers to understand, process and create natural language. Different organizations are focusing to NLP artificial intelligence technology to understand the countless unstructured data sets. All of this is lead to better data processing which correlates to better customer experiences. However, the NLP technology is such a new technology and constantly evolving that there is a little amount of information how to use it in business scenarios. For this purpose, the frameworks of modern business development are presented to guide and assess case startup Fi- belius and try to seek the zero to one possibility within the target market. For this

(32)

research three business Frameworks were selected: Lean Startup, Startup genome and Zero to One. Table 3 illustrates the key elements of each of these ideologies based on the literature. The lean startup and startup genome both provide a step-by- step model to enhance the startups operation whether on product development with lean startups build-measure-learn feedback loop or in overall startup operations evolvement through stages by startup genome. The zero to one ideology concen- trates on the correct mindset when developing startups and how to develop a different future. The frameworks have in common the importance of learning and how it can be a measure of success and how it can help when converting a business idea into a real business operation.

Table 3 Primary tenets of the frameworks

(33)

3 Methodology

A research methodology builds the structure and guidelines to the research process which is important for a successful implementation of the research. Various tools and techniques were used to conduct this research and they are presented in the following subchapters. This chapter also describes the data collection, data analysis and the plan for the research quality and ethics process to give more insight into the research process and its timeline.

3.1 Research approach and strategy

The purpose of this research was to understand how the case startup could use modern business development frameworks to its advantage. It does it by seeking possibilities within the target market and to examine how the relevant literature and theory can help AI-startup move forward from product idea to operational efficiency phase.

The purpose of this research is to create more understanding of the particular phenomenon, not to give a straight yes or no answer to complex and diverse phenomenon. This can lead to making too many straight cuts in the wrong places. In the other words, this research explores to light up different paths that the AI startup can take, not to create the right path which is always the best path in every scenario. The research was conducted in a form of a single-case study to focus on a single company to give a holistic and meaningful picture of the contemporary phenomenon within its real-life context (Yin, 2009, 4). A case study research should be considered when the research intends is to answer “how” and “why” to a particular phenomenon.

The decision to create more understanding of a particular phenomenon by focusing on a single company gives this research direction to a qualitative approach to illus- trate how the theory behaves in a real-life business scenario. In this research, the qualitative approach is enriched by analyzing both non-numerical and numerical data. According to (Baxter et al., 2010, 554) unique in comparison to other qualitative approaches, within case study research, investigators can collect and integrate quantitative survey data, which facilitates reaching a holistic understanding of the

(34)

phenomenon being studied. In case study, data from these multiple sources are then converged in the analysis process rather than handled individually. The research could not have been fully quantitative because the interest of this research was not to give simple answers to a complex phenomenon but to create more understanding of the complex questions and themes that relates to it. Qualitative research is not, and should not be, able to produce generalizable information as quantitative research would (Sarajärvi & Tuomi, 2017, 20). Also, the amount of numerical data was not comprehensive enough to draw any straight relevant conclusions. This can be seen as a strength to this research since the versatile data gives more room for the research to analyze the data from different perspectives.

3.2 Data Collection

The most common data collection methods for qualitative research are interview, survey, observation and information collected from various articles (Sarajärvi &

Tuomi, 2017, 79). In this research, the primary data was collected in October-Novem- ber 2020 through a google-forms survey. The language of the survey was Finnish since the respondents were all Finns and I wanted to be as easy as possible to con- tribute and give valuable knowledge of how and when the AI technology is needed.

The author translated the responses into English according to main language of the study. The respondent’s pool consisted of Finnish Public-sector administration influencers, private equity investors and entrepreneurs whom I believed had the best views on how the Finnish public administration is transitioning to artificial intelligence, how companies can provide transition support and what suppliers, technologies and business models are needed in that transition.

The survey gathered 22 anonymous responses from a total of 160 possible respondents who received the survey. The respondents represented the following working titles: Chief Technology Officer, Director of AI and Artificial Ecosystem, Managing director, Development Manager, Service Manager, Lead Enterprise Architect, Senior Advisor, Financial Advisor, Human Resources Officer and other experts in the field.

The responses from venture capitalists and other investors were also important from

(35)

a research perspective and also from a case company perspective. The survey con- tained both multiple-choice and open-ended questions which were all designed based on the key elements of the modern business development frameworks presented in this research. Each question was thoroughly evaluated to ensure it can be attached to the frameworks. The aim was to conduct a relatively freely structured survey to which would give enough space for respondents to own interpretations and correlations of the phenomenon to which the researcher could use to make own interpretations of the phenomenon and tie them to the research questions.

3.3 Data Analysis

The data collected were partly numerical and partly non-numerical. This made it possible to research the AI startup field in a way, that would not have been possible if the data was only either non-numerical or numerical. It gives to the researcher with a few resources a more in-depth view of a particular field, without losing the possibility of listening to unrestricted sentences that the respondents emphasized with their answers to the survey. This choice of data collection is not unproblematic. It brings up a problem when trying to place this study to qualitative or quantitative segments.

When the choice of the segment placement is not clear, it makes it hard to follow the easy data analysis method. This entirety was considered and the problems that came with it were understood when deciding methods used in this research. The realiza- tion of the possible value, that this particular choice would give to this study, was enough to prevail the possible problems.

According to Dul & Hak (2008, 6), although in a case study quantitative data can be used to generate the scores to be analyzed, the interpretation of scores of the (small number of) cases to generate the outcome of the study is done qualitatively (by vis- ual inspection) and not statistically. The response pool was not large enough for full- scale statistical analysis and test hypothesis. Therefore, the results should be presented as one suggestive result and not as the best possible way to move forward. In the words of Silverman (2013), research questions can be thoroughly addressed by combining different methods, using qualitative research to document the detail of how people interact in one situation and using quantitative methods to identify

(36)

variance. The fact that simple quantitative measures are a feature of some good qualitative research future shows that the whole ‘qualitative/ quantitative’ dichot- omy is open to question.

The multiple-choice responses are analyzed and presented with a descriptive statistics method and compared to the presented frameworks. Descriptive statistics are seen as a quantitative data analysis method and it is used in this study to describe the basic features of the data in a study (Trochim, 2020). Descriptive statistics serve as a starting point for data analysis allowing to organize, simplify and summarize the basic data sets. There are three main types of descriptive statics: The distribution, which presents the frequency of each value, the central tendency which calculates the averages of the values, the variability or dispersion shows how spread out the values are (Bhandari, 2020). By using these descriptive samples, the author can eval- uate the case company’s current business model how is it matching the need of the target market and how it could be improved in the future.

Open-ended question analysis is done with thematic content analysis by identifying and describing themes that arise from the data. The researcher analyses the pres- ence, meaning and relationships of words and concepts to make inferences related to research questions. According to Guest et al. (2011, 10), thematic analyzes move beyond counting explicit words or phrases and focus on identifying and describing both implicit and explicit ideas within the data, that is themes. Keywords are then developed to represent the identified themes and applied or linked to raw data for later analysis. Due to limited responses, the aim was not to calculate the appearance of certain keywords within the data but rather to discover patterns which could lead to the creation of larger descriptive concepts. This is how the author finally formed three main themes: product demand, customer-specific features and availability of the resources. The collection of both numerical and non-numerical data will be presented and compared to the business frameworks in a results section.

(37)

3.4 Plan for research quality and ethics

Qualitative research is not a single unified research tradition, it contains several quite different traditions. Thus, it is obvious that there are different perceptions in the scope of qualitative research on issues related to the reliability and validity of the research results. According to Sarajärvi & Tuomi (2017, 155), validation stands for that the research examines what has promised and reliability means that the results of the research are repeatable if the study is done using the same data collection and data analysis methods. The reliability and validity terms have been criticized in qualitative research because they have been developed for the quantitative research world and they mainly concern only the needs of quantitative methods (Sarajärvi &

Tuomi, 2017, 155). In qualitative research, the basis of research results rests in different interpretations of the subject. The researcher has always various background fac- tors influencing interpretation, such as my education in international business, which modifies the choices made in the research. This poses a challenge to the reliability and validity of the study and, to some extent, to the relevance of the results of the study. Thus, attention should be paid to systematic clarification of the research process and stages.

Good research is guided by ethical commitment which affects the reliability and validity of the study (Sarajärvi & Tuomi, 2017, 144). The reliability of the research comes from the research methods and the credibility of the results. The validity is formed from the data collection tools and data analysis methods. Therefore, a clear reporting on how the data was collected and what data analysis methods were used to analyze the data is important to this research process. Also, having a responsibility to conduct a research for the case company it was important to stay independent in a relation to the results and conclusions so that the reliability and validity of the research do not suffer.

The researcher’s role to be fully transparent and trustworthy throughout the research to prove that the results be reproduced when the research is repeated under the same conditions is crucial. The author ensured that the research data from the survey and the case company were handled in an ethical manner. This meant in

(38)

practice that no names or any other sensitive information were used in this research.

The invitation letter and survey questions are presented in appendices section (Ap- pendix 1 & 2). The participants of the research were informed about the purpose, methods and possible uses of the research to confidently respect the respondent’s anonymity. The responses of the survey were treated confidentially from the beginning to the end and they will be discarded after completion of the research. The identity of the respondents cannot be identified from the completed thesis. The author also ensured that the data acquisition process was handled ethically, and other researchers work, and literature sources were cited properly. Next, the research process is presented in a form of a table to give insights into the building process.

Table 4 Research stages and timeline

Stage Timeline

Identified the research problem August / 2020

Reviewed the literature September / 2020

Specified the research purpose September / 2020 Selected the business frameworks October / 2020

Collected the data October – November / 2020

Analyzed the research data November – December / 2020

Finalizing the research December / 2020

4 Results

This chapter consists of three subchapters which one by one describes the results of the collected primary data and its data analysis. Modern business development frameworks from the literature review (table 3) are compared to the data to find correlations and potential roads to follow for the case company. Table 5 AI Startup: Get- ting started combines the themes and the data analysis methods under one table.

The subchapters follow the structure of the recurring themes in table 5. The purpose is that the reader can always come back to this table and see the recurring themes and collected survey results.

(39)

Table 5 AI Startup: Getting Started

4.1 Product demand

The data revealed the different needs for products in a public organization that can evolve operations at different organization levels. Responses sifted towards the four most popular product categories show in figure 6 Product demand. 77,3 % of the responses stated that in the public sector the focus of the artificial intelligence investments will be made for Intelligent Robotics RPA + ML systems. However, the deploy- ment of RPA + ML systems has already begun in multiple areas of public sector