Data Analytics & Artificial Intelligence

A global initiative to help our clients build trust and confidence in AI within the organisation, and accelerate future innovation by empowering them to address the risks and challenges associated with AI in a proactive manner.

The Responsible AI Toolkit is an accelerator, currently built to be a modular solution that can be deployed or modified for existing workflows and, may be extended or enhanced as needed. The Czech AI Team played a key role in defining the Toolkit architecture and interface as well as by leading the AI development process.

We have designed the reporting layer and developed the Interpretability and Explainability module of the toolkit.Together with the Czech Application Development team, we brought life to the platform.

Continue to the website.

A large bank was undergoing a major digital transformation. One of the critical parts of the client acceptance process, the risk assessment, was still a paper-based weeklong exercise. Also, this scoring model was continuously losing on performance and stability.

Starting from scratch, the project delivered a fully functional scoring system, running on production on client premises. In total, we engineered 500+ features, 10 models and ran a series of workshops for fine-tuning the model and scoring system. We continue to manage the service and support the client post-deployment.

We reached out to industry data providers (telco), tapped into public (online) and government (registers) data sources and augmented these with PwC’s proprietary Geo-Data mart to maximize the level of insights and value possible.

For a large Czech car manufacturer, we implemented an illness prediction automated process. The goal was to lower the impact and costs of illnesses through improved workforce planning. On a daily basis, anonymised data from ERP is transformed and loaded onto Cloudera.

Outputs of the predictive models are stored and historicised in the same database. Aligned with client's HR reporting standards, the results are presented as Power BI dashboards to end-users.

When illness levels begin to rise, this enhanced monitoring and early-warning system enables our client to intervene and take action at a lower cost compared to a reactive approach.

 The objective has been to build various NLP solutions to support reinsurance related use cases. This allowed the client to further differentiate itself among competitors by fully utilising data and data analytics results for action-driven insights and strategies.

Oil & gas supply chain use case

• Improving the performance of existing text analytics solutions for the detection of risks connected with supply chain companies (through the application of advanced machine learning and AI). The deployment of deep learning models replacing rule-based and regular expressions solutions.

General use case

• Improving textual data preparation pipelines (embedding and tokenization).
• Entity extraction (e.g. insured name, address, SIC code) from various text sources.
• Entity linking - improve fuzzy matching logic.

Tech Stack: Tensorflow, spaCy, fastText, NLTK

Capabilities: Embedding, Tokenization, Named-Entity Recognition

One of the largest online retailer and technology providers: Identifying fake reviews

• Purchase decisions are highly affected by product reviews. It is sometimes too easy to downplay the quality of a product using fake reviews.
• Using analytical and statistical tests as well as NLP, the low-quality reviews with a high risk of being fake were identified.

 Internal: Detect the changing of job skills over time

• Skills are written in the advertised job description allowing a candidate for employment to excel in a particular job.
• The solution detects the changes of the required skills over time.
• Uses text mining to group the skills in each specific job and compare it with the previous month/year.

Tech Stack: Tensorflow, spaCy, fastText, NLTK

Capabilities: Embedding, Tokenization, Latent Dirichlet Allocation.

• Estimating potential customer value in an area to optimise branch location, closure or expansion.
• Insights based on socio-economic data from censuses, client segmentation, customer value (calculated using client internal transaction and interest data), footfall and dwell time.
• Client internal data used to develop a customer value prediction model (2 steps: classification of customers with non-zero value and estimating the value itself for those customers).
• Use of local census data to create synthetic population.
• Areas with high potential are identified by using the synthetic population features and customer value model (trained on the client internal data).

Adding new dimension

• Client challenge. A growing Czech bank was looking for the means to improve performance.
• PwC Role. Using the PwC GeoDataMart, create a database of hundreds of geospatial features.
• How we delivered. The bank almost doubled its client portfolio between 2016 and 2018 observing a significant drop in the default rate due to improved scoring.

Telco scoring pioneering

• Client Challenge. One of the largest Czech banks was lacking a model which would effectively assess the riskiness of new customers.
• PwC Role. Designing and arranging the ecosystem where banks and the telco company can share customer data.
• How we delivered. We defined the environment for sharing telco customer data to support the application process in a bank using score cards developed by us.

Improving performance

• Client challenge. A Czech branch of a global top 1000 company was challenging the performance of its own application scoring models.
• PwC role. We partnered with the telco operator and on the data of its customer developed model supplying the behavioural scoring.
• How we delivered. Our score card added more than 20% accuracy on top of the bank's scorecard.

Our client, a large global insurer operating across 10+ countries, made a decision to internalise data science capabilities in order to preserve their competitive advantage. They were looking for a reliable partner to take them through the incubation of their own, internal, data science team.

The challenge: This was not the first time our client tried to build a DS capability. In fact, they had already spent several years training their staff only to eventually see them leave.

What we did:

Trained the team: We hired and trained the data science team comprising of data scientists, business analysts and data/deployment engineers.

• Set up the infrastructure: We kick-started a future proof technological environment for the team to easily create ML&AI algorithms.
• Proved the value: Almost as a side product, we generated 4x the value of the project just within the first three months of cooperation.

Ultimate benefit: After a period of 13 months, we delivered a sturdy and fully sustainable DS team capable of servicing all their territories with a backlog of 30+ shared use cases.

Reason of success: The advantage they had, was strong commitment from the C-level managers across territories to build the internal capability.

A global PwC project building a cloud platform which delivers an end-to-end ecosystem for building an AI document annotator by any user.

A modern Azure platform

• Build on the latest cloud stack, leverage Azure Kubernetes Cluster and microservices architecture.
• (re) Deployment of a new workspace in just a few minutes.
• Fully scalable and performance optimal.

AI approach

• PwC artifactory of ready to deploy pretrained models.
• Deep learning models using cutting-edge NLP algorithms.

Continuous learning

• Unique AI framework where model performance can be continuously improved.

User-friendly experience

Goal was to create a simple and modern experience for users to retrieve any information from various structured (Excel files, database) and unstructured (PDFs, Word files) datasets. Build an easy to use tool which returns valuable predefined reports.

GenAI approach

We have developed several LLM-based solutions using an OpenAI API. Our solution supports summarization, vector database searches, structured database queries and Python coding. We have come up with strategies to prevent the model from hallucinating answers when data is unavailable.

Powerful insights

You can use the Copilot to extract insights from financial, auditing, policy documents, or internal knowledge base. We always provide direct links to the cited resource when useful. There are multiple ways users can interact with the tool: from a list of well engineered static prompts generating several reports, to a well known chatbot experience. Results can have multiple forms as well: text summary, or structured table in form of Web Application supporting Word/Excel file download.