Signals
Q4 2024
Informed AI
Proactive AI
www.gatesnotes.com www.cnbc.com www.venturebeat.com www.venturebeat.com www.builtin.com www.wsj.com www.wsj.com www.mhlnews.com www.mhlnews.com www.pages.awscloud.com www.cio.com www.cio.com
www.cio.com
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
Sign up for Mastercard Signals
Challenges
Moderating expectations
Gen AI
Since generative AI surged into the mainstream in late 2021, it’s drawn bold proclamations from tech leaders. Bill Gates has described it as the "most important advance in technology since the graphical user interface,"1 while OpenAI CEO Sam Altman has claimed to be a “little bit scared” of the tech he’s done much to create.2 Three years on, the world is navigating between these sentiments, balancing immense opportunities with emerging challenges.
AI innovation is accelerating, with each new iteration of large language models pushing capabilities further and raising the stakes for businesses, governments and society alike. But the evolution isn’t linear. As the initial hype around generative AI fades, enterprises are shifting from early enthusiasm to more informed experimentation. They’re also confronting gen AI’s high development costs and energy consumption, its voracious need for processing power, its strict data requirements and the ethical questions it raises.
Faced with this complicated environment, companies are learning by doing — refining use cases and focusing on scalable solutions that offer obvious value. Success hinges on proving not just technical efficacy but commercial viability of the sort that, for example, characterizes AI-driven knowledge bots, which are already making organizations’ internal processes more efficient. As their reliability grows, such tools pave the way for broader consumer-facing applications.
This edition of Signals examines what lies ahead, beyond the hype and towards a new reality. We identify three rapidly developing areas of AI, analyze their likely use cases and assess their potential impact.
Informed AI
Informed AI uses fine-tuning and retrieval-augmented generation (RAG) to enhance LLMs for specific applications.
Fine-tuning involves training an LLM on task-specific data and customizing its responses to address unique needs. RAG connects an LLM to external data sources, allowing the AI to retrieve and utilize unstructured data on demand. A law firm might fine-tune an LLM on its case history to provide answers relevant to its operations or use RAG to connect the LLM directly to its case database. Which method to use hinges on the quality of the training data, on the necessity for real-time updates and, most importantly, on security considerations. This is particularly significant in sensitive fields like law and finance, where protecting proprietary information from unauthorized access or assimilation into broader AI models is paramount.
Return to top of section
Sources
What’s next in gen AI
Enhancing large language models (LLMs) with external knowledge sources tailors them for specialized use cases across banking, retail and more. These models transform into subject matter experts, providing precise, context-aware insights.
Informed AI: A trusted subject matter expert
Integrating data from cameras, sensors and machine vision devices enables gen AI to interpret environmental cues and even read human emotions. This form of AI adds new contextual understanding and reasoning, making interactions more intuitive and systems more responsive.
Perceptive AI: Intelligence that sees, hears and then “thinks”
These systems anticipate needs, make decisions and achieve goals with minimal human input, representing a more advanced form of intelligence capable of semi-autonomous operation.
Proactive AI: Systems that anticipate
and then act
A trusted subject matter expert
How it works
80%
of enterprises employ RAG to supplement LLMs 7
An existing LLM undergoes further rounds of supervised learning, assimilating distinct data sets that make it proficient in specific tasks.
Fine-tuning
Compute- and data-intensive, fine-tuning is useful in giving an LLM a particular “voice” or ensuring that it produces content in a certain format.5
The right voice
AI retrieves information from a database external to the language model and generates a response using information from it, improving accuracy and relevance.
RAG
RAG lets AI systems update their knowledge bases in real time from curated data sources.
Up-to-date knowledge
20%
of enterprises
employ fine-tuning 6
How it could be used
Informed AI can monitor for fraud and analyze customer transaction data, providing personalized financial advice and product recommendations.
Banking
Personalized financial services
Underwriters could use RAG to incorporate into their analyses more comprehensive data pertaining to local market conditions or historical property trends.
Mortgage underwriting
73%
of mortgage lenders see gen AI as crucial to boosting operational efficiency in lending 8
Informed AI can create personalized travel itineraries based on user preferences and budgets, plane and hotel inventories and historical data.
Customized itineraries
By integrating with various travel databases, AI can provide travelers with real-time guidance throughout a journey.
Real-time updates
Travel
Work assistants powered by RAG-augmented AI can automate routine tasks like customer onboarding or generating reports from multiple inputs.
Enhanced productivity tools
AI can retrieve documents, research and insights from vast corporate databases, streamlining often cumbersome processes.
Knowledge management
Office
Emerging players
Cohere offers RAG capabilities tailored to enterprises’ needs — aimed at the business market.
Unstructured provides technology that automates the transformation of unstructured data into formats appropriate for RAG.
Glean provides an AI work assistant that deploys RAG to surface timely information and feed it into LLMs.
Plausible future
By automating and simplifying data retrieval and analysis, informed AI can provide custom — and timelier and more relevant — insights to knowledge workers, freeing time for more strategic tasks. Tools like AI-powered work assistants will likely become indispensable in law, research, banking and consulting. Enterprises are expected to adopt informed AI to optimize operations, improve customer interactions and drive innovation. Informed AI also enables highly personalized experiences through RAG and other advanced techniques, boosting user engagement and satisfaction across the retail, travel and customer service sectors. As these capabilities evolve, it could streamline workflows and transform how businesses connect with their customers.
Perceptive AI models are designed to interpret the world in a way that mirrors human perception, leveraging multimodal systems that integrate and process text, images and unstructured data from diverse sources such as social feeds.
By utilizing multiple sensory inputs — like video streams and sensor data — perceptive AI enhances contextual awareness, enabling AI agents to understand human intent through cues like facial expressions and body language. Looking ahead, perceptive AI stands to benefit significantly from advances in “reasoning” LLMs such as OpenAI’s o1, which can “think” through the answers it generates rather than relying solely on the pattern recognition techniques of earlier models.
Perceptive Ai
Intelligence that sees, hears and then “thinks”
Perceptive AI uses generative AI models that can process and interpret data from various sensors, such as cameras, microphones and other devices connected to the Internet of Things (IoT). This data is then combined to create a contextual understanding of the environment in question, allowing more accurate and relevant AI-driven responses.
Gen AI + sensor data
20%
2023
How it works
These models integrate multiple data types (such as text, audio and video) to provide a holistic understanding of context. AI models can analyze users’ spoken words and facial expressions to determine their emotional states and intentions.
Multimodal models
By incorporating augmented reality (AR) and virtual reality (VR) technologies, perceptive AI can create more effective immersive experiences.
Potential integration
with AR/VR
50%
2027
WAREHOUSES USING
AI-POWERERD VISION TECH 9
Perceptive AI could power sensor tech that holistically tracks the conditions of patients not only in clinical environments but in their own homes, alerting medical professionals if intervention is needed. Such solutions could transform eldercare.
Continuous monitoring
AI can analyze multimodal data (such as data related to patient history, lab results and imaging) to help doctors diagnose diseases more accurately and quickly.
Accurate diagnoses
Healthcare
Humanoid robots will use machine vision and other modes of perceptive AI to navigate factory floors, hazardous facilities like mines and other industrial environments.
Robotics
Perceptive AI will boost the utility of the Artificial Intelligence of Things, that combination of networked connected devices and AI capabilities.
The AIoT
Industry
1/3+
Using multimodal data in healthcare can improve prediction accuracy by more than a third 10
In IoT-enabled supply chains, sensors and cameras can connect to digital channels, improving inventory management, reducing stockouts and optimizing product placement.
Store operations
Sensors can track in-store customer behavior, helping merchants understand shopping patterns and tailor marketing strategies.
Customer insights
Merchants
How it could be used
Emerging players
OpenAI o1, released in September 2024, is the first LLM to “reason” in its work, generating human-like chains of thought.
Archetype pulls in data from numerous sensors and changes it into plain language descriptions of physical world situations.
Covariant builds software for AI-powered industrial robots that can learn from camera and other sensor data.
Care.ai uses sensors deployed around a healthcare facility to keep track of patients, collecting real-time information analyzed by an AI engine.
Plausible future
Perceptive AI has the potential to revolutionize human-AI interactions by delivering more intuitive, empathetic and context-aware responses. Customer service and personal assistant applications stand to gain from more natural interactions, while smart homes and cities can harness context-aware automation to elevate user experiences. In industrial settings, perceptive AI enhances safety and productivity by generating real-time insights from multimodal data, enabling better decision-making on the fly. It also has the potential to reshape workplaces by fostering more effective collaboration between humans and machines, especially in tasks that require an understanding of context and behavior. Looking ahead, advances in multimodal processing, reasoning and contextual understanding, as well as the development of ethical AI frameworks, will be crucial for the widespread adoption and evolution of perceptive AI. As these technologies mature, they could become integral to how we work, live and interact with the world around us.
Return to top of section
Return to top of section
Plausible future
Sectors like banking, healthcare, logistics, travel and customer relations will likely experience the most significant impact as proactive AI systems begin to anticipate and address future needs with minimal human intervention. As proactive AI advances, it could enhance customer experiences by providing personalized, anticipatory solutions — from customized travel planning to optimized financial management. The ability to foresee and fulfill customer needs before they emerge could offer businesses a powerful competitive advantage. Looking ahead, we can expect substantial advances in proactive AI, with these systems becoming increasingly autonomous and capable. As AI models evolve and integrate with emerging technologies such as IoT, their influence and impact will only expand, reshaping industries and redefining how businesses engage with their customers.
Start-up H, staffed by ex-DeepMind scientists, seeks to build autonomous systems that can do jobs that humans have historically done. It’s working on “frontier action models” that could boost productivity.
Emerging players
Proactive AI can analyze customer data to identify preferences and purchase patterns, then automatically deliver personalized offers and rewards.
Personalized offers
AI models can book flight, hotel and other travel arrangements based on user preferences and past behavior, providing a seamless planning experience.
AUTOMATED BOOKING
Travel
AI agents can take over repetitive and time-consuming tasks, such as data entry, report generation, basic customer support and certain software development chores.
Task automation
Proactive AI can manage projects by tracking progress, assigning tasks and ensuring deadlines are met.
Project management
Office
Proactive AI can autonomously manage personal or business finances, optimizing savings, investments and spending based on a user’s goals and changing life conditions.
Automated financial planning
AI agents can monitor transactions in real time and flag suspicious activities.
Fraud prevention
Banking
How it could be used
10%
of large companies use AI agents currently 11
Proactive AI uses advanced machine learning techniques, including reinforcement learning and deep learning, to improve over time. Integration with IoT devices and other data sources enables these AI systems to make informed decisions and respond to real-time information. “Reasoning” AI models like OpenAI’s o1 will also undergird proactive AI.
Enabling technologies
Complex tasks are broken down into collections of smaller actions that individual AI agents can handle. These agents collaborate to complete intricate
open-ended tasks.
Multi-agent systems
Systems are increasingly connected to LLMs via API, allowing them to act regardless of whether there's a human being in the loop or not.
Models that act alone
How it works
Proactive AI models are designed to reason through user-defined problems, make decisions and take action to resolve them. While fully autonomous systems and artificial general intelligence (AGI) — AI with human-like intellectual and decision-making capabilities — remain on the horizon, current advances are paving the way for them.
Proactive AI systems use LLMs integrated with various software components or agents, enabling them to interact with external systems such as databases and APIs to analyze data and execute tasks. In the future, proactive AI will likely operate with minimal human intervention, anticipating needs and addressing issues before they even surface.
Proactive AI
Systems that anticipate and then act
50%
plan to use them in the next 12 months 12
75%
of companies are considering using AI agents for software development 13
Adept AI in March 2023 announced that it had raised $350 million in new funding, including from General Catalyst and Spark Capital. The San Francisco-based company makes solutions that automate software-based processes.
Imbue, which includes Amazon’s Alexa Fund and former Google CEO Eric Schmidt as investors, is building a model that will power “autonomous agents.” These systems will perform office work, such as setting up meetings or analyzing information, without constant human oversight.
Challenges to overcome
Data quality and relevance
Ensuring that data used for training and augmentation
is accurate, current, unbiased and relevant is essential. Poor data quality can lead to incorrect, misleading, biased and undesirable outputs.
Reliability
Even enhanced AI systems remain probabilistic and lack inherent understanding of the real world. Users must avoid treating AI outputs as infallible truths, recognizing the limitations of these systems.
Regulatory compliance
Navigating diverse global regulations around data usage, storage and processing can require constant vigilance and adaptation.
ETHICAL DILEMMAS
Proactive AI edges us closer to AGI, raising the question of whether such technology is desirable at all. Whether sentient AI agents should be entitled to rights, like humans, is another question that could emerge.
Data integration and
multimodal processing
Merging data from various sources (audio, video, sensors) into coherent insights requires sophisticated algorithms and processing frameworks, presenting technical hurdles.
Safety
Ensuring the safe operation of AI, especially as it becomes more autonomous, is critical.
Data security and privacy
Handling sensitive and proprietary data requires robust security. Balancing data accessibility with privacy is particularly critical in sectors like banking and healthcare. Fine-tuned and RAG-informed LLMs will need to operate on a zero-retention basis to protect data integrity.
Complex prerequisites
AI relies on large foundational models, real-time data processing, interconnected devices and infrastructure that can handle large-scale data and integrations. These prerequisites are costly and complex.
Model interpretability and transparency
Enhancing the transparency of AI models to avoid “black box” solutions is crucial for accountability and public trust.
Interdisciplinary expertise
Deploying perceptive AI demands expertise across AI, psychology, linguistics and sensor technology. Assembling effective interdisciplinary teams can be a significant challenge.
Contextual understanding
Developing proactive AI that accurately interprets complex human behaviors, intentions and contexts is challenging — even humans often struggle in this realm.
Consistency
Current AI models often struggle with nuanced, extended instructions and with maintaining consistency in unfamiliar environments.
Moderating expectatioNS
Generative AI is evolving from a speculative innovation into a powerful force poised to reshape industries. Informed AI is enhancing decision-making, perceptive AI is elevating human-machine interactions and proactive AI is emerging to deliver unprecedented autonomy.
We draw on over two decades of AI expertise to make every digital experience smarter, safer and more personalized. AI is integral to our solutions, safeguarding more than 143 billion transactions annually across our network. With a dedicated team of data scientists, AI technologists and AI governance experts, we are committed to integrating privacy and ethics by design into every innovation. Across our capabilities — data intelligence, open banking, identity, fraud protection and cybersecurity — Mastercard ensures that trust is at the forefront and that AI is used responsibly and ethically. Our long-standing investments in AI and data technologies have enabled us to uncover differentiated insights, enhance payments and outsmart fraud. Recent advances have led to cutting-edge solutions like Decision Intelligence and Shopping Muse, empowering our partners with unmatched capabilities, delivering tangible advantages in the competitive landscape. Looking ahead, we will continue to deepen our commitment to AI, recognizing its potential as a foundational element for commerce. As we innovate, we remain steadfast in our principled approach, ensuring that AI is used responsibly, benefiting both society and our customers. It's a journey we'll take together.
Mastercard and AI
Our world-leading Decision Intelligence is a real-time AI solution that helps banks score and safely approve 143 billion transactions annually. Using new generative AI technology, our new DI Pro solution will scan an unprecedented one trillion data points per annum to predict whether transactions are likely to be genuine or not. Initial modelling shows AI enhancements boost fraud detection rates on average by 20% and as high as 300% in some instances.
Decision Intelligence Pro
Through DY, Mastercard helps its customers deliver valuable, useful and enjoyable experiences at scale with smart personalization. DY uses AI to personalize every step of the consumer journey, across numerous channels, including web, mobile apps, email, SMS, digital assistants, kiosks, in-store displays and more. In 2023, DY's AI delivered 371 billion impressions of personalized info across 450 brands.
Dynamic Yield
Return to top
Perceptive AI
Gen AI Rising superpowers
IBM’s marketing department is using generative AI to produce copy and other content. As part of its suite of solutions the company is using multiple LLMs that it’s fine-tuning with internal content guidelines to make it a “brand brain.”3
The Recording Academy used RAG-enhanced gen AI to create social media and other content about recording artists during and before the 2024 Grammy Awards. The gen AI solution deployed RAG to tap into the Academy’s music and artist databases for relevant information.4
Google’s multimodal AI models are a key part of its search and assistant technologies. Google Lens uses multimodal AI to recognize objects, translate text and analyze images in real-time, enhancing user interactions by combining visual and textual data.
Tesla’s Autopilot and Full Self-Driving (FSD) systems rely heavily on multimodal AI, integrating data from cameras, radar and ultrasonic sensors to navigate and make decisions in complex driving scenarios.
Companies like OpenAI o1 and Adept are at the forefront of this development. OpenAI’s work with LLMs and AI agents has been foundational for proactive AI, and its new “reasoning” OpenAI o1 model is likely to promote advances in this space. Adept is creating systems capable of autonomously performing complex tasks across various environments and, notably, recently agreed to license its technology to Amazon.
Paving the way for proactive agents
Banks, merchants and digital players stand to benefit from these innovations, but the path forward has challenges as the technology matures. Banks can leverage AI for more precise risk management and personalized financial services but must also address data security and privacy concerns. Merchants can use AI to create dynamic, personalized customer experiences but must also navigate integration and gain customer trust. Digital players are well-positioned to harness AI's capabilities to drive innovation but will face challenging ethical and regulatory landscapes. While innovations traditionally progress in stages, the gaps between these stages are rapidly closing — advances that once took decades are now achievable in just a few years. As the pace of change accelerates, today's decisions will shape how effectively these technologies drive future growth, efficiency and resilience.
Tokenization is proving itself to be a transformative technology. It not only efficiently overcomes challenges around data security and privacy but also unlocks new opportunities for commercial and governmental players with the emergence of a token economy. As innovators continue to explore additional use cases beyond payments, tokenization will reinforce its role as an invisible enabler, underpinning secure and efficient digital interactions across various sectors.
Tokenization is proving itself to be a transformative technology. It not only efficiently overcomes challenges around data security and privacy but also unlocks new opportunities for commercial and governmental players with the emergence of a token economy. As innovators continue to explore additional use cases beyond payments, tokenization will reinforce its role as an invisible enabler, underpinning secure and efficient digital interactions across various sectors.
Tokenization is proving itself to be a transformative technology. It not only efficiently overcomes challenges around data security and privacy but also unlocks new opportunities for commercial and governmental players with the emergence of a token economy. As innovators continue to explore additional use cases beyond payments, tokenization will reinforce its role as an invisible enabler, underpinning secure and efficient digital interactions across various sectors.
At Mastercard, we are committed to leading this transformation, ensuring that as AI reshapes the landscape, it does so responsibly and with a focus on creating value for all.
Signals
Q3 2024
Decoding SME’s
Solving the SME puzzle
The fintech journey
Preparing for tomorrow
Challenges