aiStudio Tools Solving Real-World Problems
It starts by asking the right questions: What problem are we trying to solve? Is the intended use ethical an compliant? Is the case economically viable? Could it be solved just as well via simple rule-sets? Application is as critical as the technology itself: just because we can build it doesn’t mean we should. While we love working on AI, the aiStudio always puts client objectives first, not the technology. Yes, we are motivated by the challenge, even more so by the satisfaction in helping our clients find more effective solutions than ever before.
Design and implementation of an “AI Audit Agent” opens up deeper auditing possibilities
Uncertain exposure to policyholder claims related specific contract wording risked becoming a deal-breaker to the sale of a life insurance subsidiary. The seller required a swift risk assessment of over 300.000 policies within record time in order to salvage the transaction.
Deloitte designed an AI-powered “virtual auditor” based on the NLP (Natural Language Processing) technologies of TableMiner to scrutinize the entire portfolio, constituting over 21 TB after conversion into text-file format. The audit was performed below budget, ahead of schedule, and with unprecedented accuracy, proving that risk-based sampling can be substantially augmented through intelligent use of AI technologies.
Finding critical flaws in data that hinder accuracy of even the best model architectures
A large consumer lender in the UK had been using machine learning models for some time – and was well familiar with the risk posed by bias, when data does not fairly reflect the population it intends to represent. Detecting bias can be as difficult as predicting its effects: consequences range from inconvenient (rework, productivity) to severe (legal/regulatory risks, reputational damage), depending on the application.
Using Model Guardian, Deloitte quickly identified and quantified bias lurking within the client's credit scoring model, understanding and optimizing trade-offs. The final report explained biases initially found and tracked how subsequent iterations of the model effectively dealt with them, arriving at an acceptable compromise between accuracy and fairness. The project was able to provide insights quickly through Model Guardian’s wide variety of leading fairness metrics and convenient user interface, requiring no additional coding.
Sophisticated risk identification helps focus the scope of internal audit and maximize resources
The Internal Audit function of a large global public authority manually selects samples to further document during their audits. This process is time consuming & potentially skewed, as auditors investigate what has already been found to be deviant without necessarily checking for new deviant/fraudulent patterns.
Deloitte built an automated sample selection functionality on the basis of Consistency to detect anomaly (high-risk cases) hidden in the mass of data. The tool was integrated into the client’s web portal combined with advanced visualization capabilities from the Google Cloud Platform. It provided a list of multi-dimensional outliers – indicating either errors, acceptable process deviations or even potentially fraudulent behavior. The project enabled proactive insights in near real-time.
Intelligent data discovery reduces time and cost of data management, mapping & migration
A longstanding, multi-national bank manages a highly complex (and costly) landscape of systems and databases, placing it at a competitive disadvantage vs nimble fintechs. With IntelliMap, Deloitte greatly simplified the mapping process at the core of relational database architecture: rather than considering all possible combination, the tool analyzes field contents, name and description to propose the most likely matches.
Run on an anonymized database of bank accounts and account settlements, Deloitte processed high volumes of data (initial mapping output) in minutes instead of hours or even days. Up to ten-fold reduction in work time over manual database mapping. This freed capacity of highly specialized experts working on large data migrations, API connections, or data warehousing efforts.
Automating the data science pipeline allows AI solution developers to go farther and faster
A large South African home lender and leader in secured lending had reached a plateau in credit scoring accuracy. Incremental growth and competitive strength was greatly dependent on the predictive power of their credit scores. Focusing initially on secured lending clients, the company asked Deloitte to apply machine learning to the problem.
Using the rapid prototyper DNAi, the Deloitte team quickly found the optimal combination of features (drivers) and algorithms to arrive at a short-list of highest performing models. The team was subsequently engaged to fully re-model the scorecards, which not only resulted in more reliable models, but also easier maintenance. Default rates dropped 20% while able to grow the portfolio 40% within the same risk policy.
Greater accuracy in deep learning from automated quality control & correction of training data
LiDAR, radar, and computer vision are the key sensory components to AV perception of its surroundings. An automotive technology researcher involved in Autonomous Vehicles (AV) was concerned with the quality of image training data fed to the computer vision AI model. Labeling errors threatened the accuracy required for the AV to operate safely in favorable and adverse conditions, typical and unexpected situations.
Observed errors rates between 10-20% were unacceptably high. Quality assurance was prohibitively expensive and its random sampling procedure of limited effect. Deep Label swiftly solved both cost and quality problems, providing an automated means to both detect and correct mislabeled images. Model accuracy improved 11% while quality assurance costs were cut by over 80% (without the automated relabeling). Functionality has since been extended to suggest label corrections.
Cost-effective scrutiny prior to deployment ensures that AI models function as intended
Designed and implemented properly, AI-based models promise greater accuracy, handle larger volumes of data and adapt flexibly to new information, learning along the way. A compelling value proposition. There are, however, many pitfalls between business case and successful deployment. An abundance of open source libraries lure experts and debutantes alike, which can amount to glorious successes and perhaps equally spectacular failures.
A technology investor sought to acquire a specific AI-based solution, requiring an independent due diligence. Deloitte scrutinized the product – design strategy, architecture choices, development process, code review, documentation – based on the inspection detail behind the “Deloitte Trustworthy AI Framework” and especially the collection of tools from the aiStudio: Model Guardian to check for representative training data, Lucid[ML] to explain the inner workings of the blackbox algorithm, and DNAi to quickly assess vs challenger models.
Smart cities harness computer vision to enable rapid & reliable surveillance of road degradation
Two Belgian sewage utilities sought to leverage technology to both reduce cost and project execution risk for roadworks. Risk was a central concern: left unrepaired, road damage can extend below the surface to the sewage grid... or faulty sewage works could result in sinkholes forming on road surfaces. Scarcity of surveyors led to high-cost, infrequent road assessments.
Deloitte implemented the automated damage detection tool RoadRunner, enhancing it with a mobile phone “dash-cam” application to automatically record images as driving along the road. Combining the images with labeling by surveyor experts, the AI was trained to automatically detect damage along the driven route, fed back to a mapping visualization via geo-positioning of captured images.
Collaboration tools and AI-based forecasting improve complex management decisions
A large South African retailer faced increasing planning complexity with an ever broader product offering. Time-consuming traditional methods were unreliable, late and notoriously skewed. The retailer needed a new solution: a fast, data-based approach that could break free from failing heuristics and keep pace with product complexity and changing demands.
The breakthrough came from Deloitte’s creative solution, a collaborative platform that weaves together artificial and human intelligence. Artificial intelligence in the form of the Clairvoyance time-series engine provided the basis: forecasts for each country, each SKU, and each set of planning assumptions. The platform provided an interactive means for experts to jointly review and refine scenarios. Planning for over 8.000 country-SKU combinations was delivered on time and able to properly advise management decisions.