Customized AI software development is one of the most powerful approaches to leveraging AI in business. It brings a true competitive advantage by implementing solutions tailored to the specific challenges and business needs of the enterprise. Undoubtedly, the main difficulty of this approach is the ability to successfully develop and implement AI projects. Cooperation with an experienced AI vendor, who is responsible for end-to-end delivery, is one of the most effective solutions, bringing with it a number of benefits.

Cost reduction

One of the main drivers influencing the popularization of AI software outsourcing is the significant cost reduction. The experience and know-how of the AI vendor ensure a transparent approach to project planning, implementation and related costs. Moreover, savings also come from reducing expenditure on hardware, workspace, training and employment. The client receives full support starting from needs identification to commercial deployment, as well as access to top tech talents.
At deepsense.ai we are flexible when it comes to the cooperation model. The first two are focused on a fixed price or time & material approach designed for customers who have clearly defined goals and requirements related to customized AI software development. In this case, the deepsense.ai team is responsible for the implementation of a specific solution or a component thereof. Where less specific projects are concerned – where in the first place the business use case needs to be identified and described in the data science language – the team augmentation approach is a better solution. It allows business owners to control costs at every stage and flexibly decide on the further development of the project. deepsense.ai’s people support the client’s internal team on a daily basis, as regular peers, and provide knowledge transfer. Such cooperation significantly increases the in-house software development capacity.

Flexibility

The diversity of AI and ML applications requires a plethora of experience and skills. Outsourcing provides a great deal of flexibility and confidence in testing new approaches. Even companies with in-house data science teams which are willing to execute multiple ML projects would need to spend a lot of time preparing and training for each implementation. Instead, they can work with an experienced vendor providing know-how about the latest technologies and possible solutions. At deepsense.ai, with over 100 world-class full stack developers, data scientists and software engineers on board, we provide effective support in the area of AI software development. Our people are equipped with the full tech stack required to help clients reach their milestones quicker, and ready to work hand-in-hand with internal business and technical teams.
Flexibility can be also provided by the agile approach to developing AI software. At deepsense.ai we work on the basis of CRISP methodology – a process model including six stages of running AI projects that leads to the right solution when repeated over and over again. This concept involves an adaptable attitude, where priorities and pathways change as project development progresses.

Access to top tech talents

The majority of enterprises are faced with a huge shortage of skilled AI talent. In particular, companies whose core operation is not related to AI may have a problem attracting world-class experts. Cooperation with an experienced AI vendor gives enterprises access to top tech talents without incurring the costs of recruitment and employment. Moreover, people with highly specific tech competencies can jump on board only when needed, for example just to streamline the process of data analysis or model training. At deepsense.ai we put great emphasis on acquiring and retaining the most talented AI/ML and data science experts. Our team members have backgrounds from top European technical faculties and have won international awards. Moreover, they develop tech competences not only through the implementation of commercial projects, but also through their involvement in the scientific development of AI/ML, academic activity and participation in prestigious data science competitions.

Focus on business value delivery

At deepsense.ai, the overriding goal of customized AI software development is not to deliver the software itself, but above all to provide the client with real business value. The continued success of a project depends on how deeply the given business use case is understood. That’s why at deepsense.ai we focus on close relations with business owners. From the very beginning of cooperation with clients, the entire team dedicated to the project participates in the discussion on business needs, possible solutions, and available data. This ensures that all team members have a broader awareness of the purpose, know the limitations, and are able to contribute both technically and conceptually. This approach makes it possible to efficiently determine the scope of cooperation, focus on low-hanging fruit first, and set up success metrics. The external team is independent and can objectively assess the chances of delivering business value in a given area. Such confidence combined with experience and agility in training ML models offers the best chance of success. As a result, clients remain focused on the core elements of their business and in the meantime get a customized solution tailored to their needs.

Faster results

While cooperating with an AI vendor, there is no need to take care of the necessary infrastructure or training, freeing up time for analyzing data, testing various approaches and building models. This significantly shortens the project implementation time. The vendor’s proven track record of building and implementing AI guarantees the delivery of first-class solutions. deepsense.ai’s portfolio includes more than 150 commercial projects for clients from the USA and Europe. Our commitment and know-how are valued by global companies including Nielsen, L’Oréal, Intel, Nvidia, United Nations, BNP Paribas, Santander, Hitachi and Brainly.

Summary

Close cooperation with an AI vendor allows enterprises to maximize the potential of state-of-the-art technologies and focus on the industry-related aspects of building a competitive advantage. We might go so far as to say that truly remarkable results are not possible without combining in-house and outsourcing models – at least not without investing a great deal of time and significant financial outlays. Cooperation with a reliable AI vendor ensures a number of benefits, and also brings a fresh perspective on the data which is being analyzed.

Introduction

In recent years, we have seen rapid development in the field of artificial intelligence, which has led to increased interest in areas that have not often been previously addressed. As AI becomes more and more advanced, beyond model effectiveness, experts are being challenged to understand and retrace how the algorithms came up with their results, and how the models are reasoning and why [Samek and Muller, 2019].

Such knowledge is necessary for many reasons: one is ensuring that the AI-driven solutions comply with regulations, particularly in finance. Another reason is a better understanding of how a model works, which translates into reducing errors and anticipating the strengths and weaknesses of the model, as well as avoiding unexpected behavior already in production. And last, but not least, it allows for the creation of models that are inclusive and eliminate the impact of social biases which can appear in the training data. The use of explainable AI (xAI) translates into increased trust and confidence when deploying an AI-powered solution.

The need for an explanation of AI models is especially noticeable in natural language processing tasks. A common part of many solutions in this domain is the use of vector representations of words. As it turns out, these representations also model human biases that are found in the data. Examples of such biases include: gender bias, racial bias, or social bias towards people with disabilities [Hutchinson et al., 2020].

Categorization of Explanations

The approaches used in explainable artificial intelligence can be divided in several ways. Some of the most functional in both practice and science are the divisions by what is being explained and by at which stage of model usage the explanation is happening.

The first one tells us what we are explaining:

• local explanation – presents an explanation of one particular decision, e.g. by showing which words in the input example were important
• global explanation – shows the entire model’s behavior. In this article, we focus on local methods.

The second possible split is based on how the explanation is created:

• explanation by design – the models that are intrinsically explainable, like decision trees
• post-hoc explanation – the model is a black-box; however, with post-processing methods it is possible to determine how the decision was made

This taxonomy allows us to organize and characterize the available methods.

Methods walk-through

LIME and SHAP

Let me start by describing the LIME [Ribeiro et al., 2016] and SHAP [Lundberg and Lee, 2017] AI explanation methods, which are examples of post-hoc local explanation algorithms. The idea behind LIME is to create a simpler, interpretable model that approximates the behavior of a complex model in the neighborhood of the analyzed example, and is visualized in the picture below:

The simpler model is trained on the data coming from the perturbations of the input data, where the ground truth is the result returned by the complex model. In the case of textual data, perturbations are usually made by removing words from the analyzed example. SHAP, on the other hand, leverages game theory and using Shapley Values determines the importance of perturbed examples.

They are very handy to use as they can work with any black-box model and with most data types. Another reason for the growing popularity of this approach is the great SHAP library (https://github.com/slundberg/shap) which is easy to use and provides well designed visualizations. Below, we can see an example of an explanation from SHAP documentation presenting why the model predicted the sentiment of a movie review to be positive.

Firstly, we initialize the pretrained model from Huggingface for predicting sentiment:

Next, we import the SHAP library and initialize the model explainer as well as give it an example to predict and analyze:

Finally, we use the visualization module to transform the resulting values into a graph showing the marginal contribution of each word (in simplified terms, we can think of this as feature importance):

And the result is presented below:

We can see here that the phrase ‘great movie’ had the highest contribution to predicting the phrase as positive and the model was unable to capture the sarcasm.

Unfortunately, no approach is without its flaws. Researchers point out that an explanation actually comes from another model, which may have fidelity problems in reproducing the actual model.

Gradient-based explanations

Another group of approaches used in explaining NLP model predictions are methods that analyze the gradient appearing in a neural network. More precisely, they analyze the change of the selected decision class with respect to the input example. Since one backward pass is enough to create a saliency map, and they do not use a surrogate model, they are free of the problem that plagued the perturbation approaches because they analyze the model explicitly.

A great tool that allows simultaneous decision analysis using multiple gradient approaches (and also the LIME algorithm described earlier) is LIT – Language Interpretability Tool (https://pair-code.github.io/lit). It also provides a comprehensive analysis of the model and dataset. The concept of this tool is a little different from SHAP. It works as a standalone web service, which can also be run in Jupyter Notebook.

Below, you can see the script that loads the dataset together with a pretrained model and runs the service in Jupyter Notebook:

We run the same example as in the SHAP section. The outcome is presented in figure 3:

The result clearly shows that the key phrases in the prediction were ’great’ and ’no taste’. Unfortunately, even in this example we can observe a divergence of the contribution of the word ’taste’, which seems to be relevant for correct prediction.

Recent research shows that explanations generated by gradient methods should be used with caution because the largest gradient tends to be concentrated in high frequency areas. In other words, where there is a lot going on in the input example (this is easier to understand in an image example where we have a background and an object in the foreground – the higher frequency will occur in areas of the object contours) [Adebayo et al., 2018]. As an alternative to such approaches, the intrinsically explainable models could be used, in which the decision mechanism is understandable by humans [Rudin, 2019].

Using attention to generate explanation

Recently, approaches that use an attention mechanism have become very popular for xAI in NLP. The attention indicates to the network which words it should focus on. It can be incorporated into many networks, e.g., by adding an attention module between the encoder and decoder in an LSTM, or one can build an entire network based on this mechanism, as in Transformer which is based on self-attention blocks. As it requires a specific architecture it will not work with every model, so it represents a group of methods in which the models explain themselves.

The Language Interpretability Tool also allows you to visualize all the heads and attention layers in the Transformer-based model (BERT) used in the earlier example. An example can be seen below:

Already in this single example, we observe that this mechanism pays a lot of attention to punctuation marks. Researchers also point this out and indicate that the attention mechanism, despite its great performance, does not always focus on data parts that are meaningful from a human’s point of view, and thus may be less plausible [Jain & Wallace, 2019; Wiegreffe & Pinter, 2019]. If you would like to explore a method that uses the attention mechanism but creates a saliency map like previous methods and simultaneously addresses the problem of too much focus on punctuation marks, we encourage you to check out the publication Towards Transparent and Explainable Attention Models by Akash Kumar Mohankumar et al.

Example-based explanation

The example-based explanations will be the last group of methods mentioned in this article. There are many approaches here, but what they have in common is showing a data point that is in some way relevant to explaining a particular example. The simplest example might be the Nearest Neighbor algorithm, which identifies the most similar example from the training set and presents it as an explanation. An extension of this approach are prototypes, where the most representative examples in the training set for each class are found and a prediction is made as in Nearest Neighbor by analyzing which prototype is most similar to the input example. On the other hand, there are counterfactual explanations, which attempt to change the input example as little as possible so that the classifier’s decision changes, i.e., they implicitly explain the decision boundary.

At this time, to the best of our knowledge, there is no production-level quality library for example-based explanation. The simplest approach would be using the Nearest Neighbor algorithm while encoding texts with the Universal Sentence Encoder [Cer et al., 2018] or other pretrained models. For more advanced examples, we should use more research-oriented methods.

An example could be ProSeNet [Ming et al., 2019]. The authors train the neural network to learn to find prototypes that best represent the entire dataset while being very diverse. The model makes predictions like the Nearest Neighbor algorithm, making it intrinsically explainable. For an example of explanation, see the image below:

Here we can see an input example together with most similar prototypes that were used to make a prediction.

The advantage of example-based approaches is that they usually work with numerous models and explain decisions very well. On the other hand, these methods are usually very simple and do not allow fits in complex pipelines, or in the case of models adapted to be explainable in such approaches (e.g., neural prototype finding like ProSeNet or concept learning), they offer poorer performance than their corresponding black-boxes.

Summary

Explaining artificial intelligence is still a developing area, and the available tools are not always ready for use in production. In each project we are driven by individual customer needs, and we tailor solutions to the requirements. In this article, I have shown several methods that are often used to explain NLP models. Each of them has its advantages, but also its limitations. Knowing them allows us to use the appropriate methods depending on our needs.

For those who would like to learn more about xAI, I recommend checking out the book Interpretable Machine Learning by Christoph Molnar.

References

Wojciech Samek and Klaus-Robert M ̈uller. Towards explainable artificial intelligence. CoRR, abs/1909.12072, 2019. URL http://arxiv.org/abs/1909.12072.

Ben Hutchinson, Vinodkumar Prabhakaran, Emily Denton, Kellie Webster, Yu Zhong, and Stephen Denuyl. Social biases in NLP models as barriers for persons with disabilities. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5491–5501, Online, July 2020. Association for Computational Linguistics. doi: 10.18653/v1/2020.acl-main.487. URL https://aclanthology.org/ 2020.acl-main.487.

Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. ”why should i trust you?”: Explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’16, page 1135–1144, New York, NY, USA, 2016. Association for Computing Machinery. ISBN 9781450342322. doi: 10.1145/2939672.2939778. URL https://doi.org/10.1145/2939672.2939778.

Scott M. Lundberg and Su-In Lee. A unified approach to interpreting model predictions. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS’17, page 4768–4777, Red Hook, NY, USA, 2017. Curran Associates Inc. ISBN 9781510860964.

Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, editors, Advances in Neural Information Processing Systems, vol- ume 31. Curran Associates, Inc., 2018. URL https://proceedings.neurips.cc/paper/2018/file/294a8ed24b1ad22ec2e7efea049b8737-Paper.pdf.

Cynthia Rudin. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 1(5):206–215, 2019. ISSN 25225839. doi: 10.1038/s42256-019-0048-x. URL https://doi.org/10.1038/s42256-019-0048-x.

Sarthak Jain and Byron C. Wallace. Attention is not Explanation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 3543–3556, Stroudsburg, PA, USA, 2019. Association for Computational Linguistics. doi: 10.18653/v1/N19-1357. URL http://aclweb.org/anthology/N19-1357.

Sarah Wiegreffe and Yuval Pinter. Attention is not not explanation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 11–20, Hong Kong, China, November 2019. Association for Computational Linguistics. doi: 10.18653/v1/D19-1002. URL https://aclanthology.org/D19-1002.

Daniel Cer, Yinfei Yang, Sheng-yi Kong, Nan Hua, Nicole Limtiaco, Rhomni St. John, Noah Constant, Mario Guajardo-Cespedes, Steve Yuan, Chris Tar, Yun-Hsuan Sung, Brian Strope, and Ray Kurzweil. Universal sentence encoder. CoRR, abs/1803.11175, 2018. URL http://arxiv.org/abs/1803.11175.

Yao Ming, Panpan Xu, Huamin Qu, and Liu Ren. Interpretable and steerable sequence learning via prototypes. In Ankur Teredesai, Vipin Kumar, Ying Li, R ́omer Rosales, Evimaria Terzi, and George Karypis, editors, KDD, pages 903–913. ACM, 2019. ISBN 978-1-4503-6201-6.

Many enterprises that consider AI implementation don’t know how to go about it successfully. Companies complain that they don’t have the necessary in-house knowledge or skills, and building a team from scratch is too time-consuming and risky. Because the effective implementation of AI projects requires building interdisciplinary teams – both from business and technology –  executing the process successfully is quite a challenge.

The answer to such a challenge may be to combine in-house and outsourcing models. In one respect, an in-house team provides crucial know-how and knowledge about their company’s operations. An AI vendor, meanwhile, provides cutting-edge knowledge of AI technology and a flexible, innovative approach to realizing the use cases. Ultimately, the vendor’s team becomes an integrated part of the client’s team.  Sounds promising… but before starting cooperation some key questions should be answered.

Where is the business value?

The first element that needs to be decided is what business case the cooperation with the vendor should address. While enterprises are more and more certain about AI’s potential there is still a lot of hype around, making it difficult to decide which use case to implement first. It is not only about not losing money on an AI project, but above all about identifying such business processes where automation will bring visible business values.

The unique approach and work methodology developed by deepsense.ai allows us to quickly identify the most relevant use cases. Our cooperation with a client always begins with an opportunity discovery phase that helps map and understand the business needs. Joint ideation sessions define the main processes and data sources that can become part of the implementation. This approach makes it possible to efficiently determine the scope of cooperation and set up success metrics.

Who from my team will be involved in the project?

Effective implementation of AI projects requires building interdisciplinary teams. Ones that command extensive expertise in available technologies and approaches to data analysis, while at the same time possessing a thorough understanding of the business processes taking place in the enterprise. They must be able to recognize the factors that influence the core operations. As business challenges go, this one certainly qualifies as daunting. However, it is important to consider who – within the organization – will be able to become a partner for cooperation with an AI vendor. This person, or persons, should not only take care of the smooth development & implementation of the project, but also have a full picture of the business situation within the analyzed use case.

What kind of  data can we share with the vendor?

The garbage in/garbage out principle is well known in data science. But, it should also be considered from a business stakeholders perspective. Certainly, most enterprises possess data that can provide valuable insights. However, it is important to review whether the dataset covers all data sources available in the organization as well as assess data quality and consistency. Usually the process of collecting and preparing data for analysis – to the customers’ surprise – takes more time than it was expected. To speed up the process, the deepsense.ai team always begins to review and analyze the data in close cooperation with the customer’s domain experts. At this stage, it is crucial to discuss basic questions, and check features and labels.

An additional element worth paying attention to is the data security. While working with an external vendor the knowledge of core business processes somehow goes beyond the enterprise. That’s why we always implement adequate security measures, such as encryption, network isolation, and strict data access policies.

How much time do I give to see results?

Valuable data analysis requires patience. Often, before reaching the final solution, several models are trained simultaneously using different perspectives and approaches. The duration of the cooperation depends on the specifics of the project. Usually it takes several months from the data audit and development plan to the operational rollout and commercialization of the entire project. The first working proofs-of-concept are usually presented within 4 weeks from the start of the project.

In what model do I want to work with a vendor?

At deepsense.ai we have the flexibility to work along two different models. The first one is focused on team augmentation. We work in cooperation with our customer’s teams to expand their technical skill-set within a broad set of roles. Our data scientists, data, and software engineers are ready to tailor to any specific needs. Our people are equipped with the full tech stack required to help clients reach their milestones quicker, and ready to work hand-in-hand with internal business and technical teams.

The second model provides end-to-end project delivery: from identifying the needs to commercial deployment. We also specialize in handling highly customized challenges for which ready-made solutions don’t exist.

Summary

The dynamic development of AI makes it difficult for companies whose core operations are not related with hi-end technology to keep up with the latest solutions. Close cooperation with an AI vendor allows them to maximize the potential of state-of-the-art technologies and focus on the industry-related aspects of building competitive advantage. We might go so far as to say that truly remarkable results are not possible without combining in-house and outsourcing models. At least not without a great deal of time and much larger financial outlays. Cooperation with a reliable AI vendor ensures flexibility, cost control, and also brings a fresh perspective on the data being analyzed.

There is no question that the enterprises that successfully leverage AI’s potential will be the ones to get ahead. While businesses are more and more certain about the goals they want to use AI for, how to execute and deploy it successfully remains a difficult question.

Enterprise AI – evolution not revolution

In today’s world, the dynamics and efficiency of operations are crucial factors, and data is the key to flourishing in the rapidly changing business reality. Data provides insights from various areas of business activity. Solutions based on artificial intelligence or machine learning not only support the analysis of huge amounts of data, but also provide a new approach to optimizing and automating core processes. Enterprises across the full spectrum of industries can build competitive advantage based on AI. Such a strategy need not come down to one-off AI implementations. Rather the most successful will take a holistic approach to treating data and high-end technologies as core organizational assets. From this perspective, enterprise AI can be considered as the next, natural stage in the digital transformation of enterprises.

New approach to AI team augmentation 

Many companies want to implement the enterprise AI approach, but don’t know how to go about it successfully. Companies complain that they don’t have the necessary in-house knowledge or skills, and building a team from scratch is too time-consuming and risky. In fact, the effective implementation of AI projects requires building interdisciplinary teams – ones that command extensive expertise in available technologies and approaches to data analysis, while at the same time possessing a thorough understanding of the business processes taking place in the enterprise. They must be able to recognize the factors that influence the core processes. As business challenges go, this one certainly qualifies as daunting.

The answer to such a challenge may be to combine in-house and outsourcing models. In one respect, an in-house team provides crucial know-how and knowledge about their company’s operations. The outsourcing partner, meanwhile, provides cutting-edge knowledge of AI technology and a flexible, innovative approach to realizing the use cases. Ultimately, the vendor’s team becomes an equal member of the client’s team. Such an approach seems to have a number of disadvantages, as the knowledge of core business processes somehow goes beyond the enterprise. However, the benefits of this approach far outweigh such disadvantages. It might go so far as to say that truly remarkable results are not possible without combining the in-house and outsourcing models. At least not without a great deal of time and much larger financial outlays. Cooperation with a reliable outsourcing partner ensures flexibility, cost control and it also brings a fresh perspective to the data being analyzed.

deepsense.ai’s holistic approach to enterprise AI transformation 

The unique approach and work methodology developed by deepsense.ai maximizes synergy. Our cooperation with the in-house client team always begins with workshops that help map and understand the business needs. Joint ideation sessions define the main processes, data sources and use cases that can become part of the implementation. The deepsense.ai team then begins to review and analyze the data in close cooperation with the customer’s team on a daily basis. This model was implemented in cooperation with our client, the retail research leader Nielsen. Working alongside Nielsen’s experts across the globe maximized the synergy between our two companies. Together we developed an advanced AI solution to automate extraction of content images of varying lighting conditions, viewing angle and quality. After implementing the solution we conducted technical workshops for Nielsen’s team to transfer the know-how and code. We also provided assistance in cloud readiness and scaling.

Summary

The dynamic development of technology makes it difficult for companies whose core operations are not related with hi-end technology to keep up with the latest solutions. Close cooperation with a technology partner allows them to maximize the potential of state-of-the-art technologies and focus on the industry-related aspects of building competitive advantage.

Innovative leaders from the insurance industry are starting to recognize the benefits of machine learning and deep learning applications. As business approaches to using AI-based solutions evolve, concerns about the transparency and explainability of AI models are also diminishing. While the trust in this type of solutions is growing, AI offers a new look at the product and service portfolio, as well as improve the sales process itself to best suit customer needs. As predicted by McKinsey experts, the technological revolution in the insurance industry has already started, and its progressive development will lead to a massive technological shift over the next decade.¹

AI-based customer service automation

One area where insurance leaders look to implement AI solutions is customer service automation. Chatbots and biometric identification are becoming permanent features of the customer service landscape.
Chatbots automate customer service, increasing customer experience and satisfaction. Behind the scenes, automation significantly reduces customer service workloads. As insurance industry leaders have indicated, over 90% of customer service engagements follow basic patterns, leaving excellent potential for automation while retaining high customer satisfaction.
Biometric identification supports explicit or unnoticed identity verification within remote channels. This can include voice identity verification in call centers or typing manner verification in online channels. Similar solutions are poised to gain traction, particularly in the post-covid world.

Back-office optimization

Another area where AI-based applications are highly effective is back-office optimization.  The documents and forms processed in the insurance industry are voluminous, so improvements in this area quickly bring noticeable results. Given that, deepsense.ai developed an insurance claims assessment system for one of the largest financial institutions in Eastern Europe.  Employing advanced computer vision algorithms, the system recognizes car parts and classifies which are broken, assesses the severity of the damage and estimates the cost of repairs. It allowed nearly real-time assessment of claim value based on image documentation sent by the client. With 70% of claims handled automatically, the manual work required in the process is significantly reduced, while the time needed to assess damage falls from days to seconds.

Automating text and image documentation analysis can be widely applied in insurance, including for cash operations, trade finance, insurance application processing as well as classifying incoming emails to go to the appropriate department or customer segmentation.

Customer insights

Applying deep learning to customer analytics makes it easier to combine insights from various data sources (e.g. transactions, online banking logs, call center interactions) as well as a 360^o customer view and personalized sales activities. This helps to better understand customers and to build personalized recommendations, making the business more responsive and efficient. For one client in the banking sector, deepsense.ai designed a machine learning model that created personalized recommendations, showing what should be proposed to a given client based on the type of customer they are. The model had a 93% recall rate, which is 60% more effective than traditional product recommendation techniques.
Thanks to accurate AI algorithms, we can also improve customer retention by predicting churn probability. This is important as customers often churn without obvious warning signs. The roots of deepsense.ai’s experience with anti-churn use cases go back to the banking industry. We developed an intelligent anti-churn system based on historical internal data and external databases for a leading financial institution in Central Europe. To identify risk groups that were more likely to churn, we leveraged multiple data sources and gradient boosting trees and also evaluated customer lifetime value. The results were used to prioritize communication with highlighted clients, 50% of which confirmed they were considering leaving the bank.

Insurance risk management

Machine learning enables nearly fully automated fraud detection, adapted to individual patterns and changing behaviors. It can be applied in areas where a large volume of events needs to be analyzed in real time. Our client, a leading CEE insurance company, suspected that some end customers were abusing access to private healthcare. deepsense.ai was tasked with analyzing data in a search for anomalies and spotting the data-marks of fraudulent transactions. With the knowledge it gathered, the team developed algorithms identifying common schemes and techniques of private healthcare abuse. The schemes included excessive medical diagnostics and exploiting flaws in billing systems. The team also identified potential fraud being committed by service providers abusing their agreement with the health insurance company. The model we delivered spots suspicious activities and has enabled the company to reduce losses by over 3M EUR annually.

Summary

A wide range of ML applications is increasingly being used to solve real business problems in insurance. As AI becomes more popular, these applications will become the market standard. Nevertheless successfully leveraging the new opportunities offered by cutting-edge technologies will require insurers to undertake a shift in corporate governance. However, if applied successfully, AI-based solutions will benefit both insurers and clients alike.

¹ https://www.mckinsey.com/industries/financial-services/our-insights/insurance-2030-the-impact-of-ai-on-the-future-of-insurance