Artificial intelligence has become ubiquitous, featuring more and more in our business activities and everyday life. Yet the term itself covers a wide range of technologies performing different tasks. Each of them has a role to play in procurement, but it’s important to understand the differences between them to ensure that you apply the right tool for the right job. Ethical and regulatory implications might also come into play, as well as reputational and other risks to your business. In this article we focus on generative AI and how it relates to, and differs from, other forms of AI.
What is Generative AI?
Generative AI can be distinguished from other forms of AI by its capabilities and machine learning techniques, but for the moment we’ll focus on functionality and application. Generative AI (GenAI) refers to artificial intelligence that creates new content (text, images, music, code, etc.) based on self-supervised learned patterns. Examples include ChatGPT, DALL·E, MidJourney, Claude, and Gemini. GenAI is revolutionizing the creative industries, but it also assists in those aspects of procurement that involve the creation of new content, or the revision of existing content, such as RFQs and contracts.
Other (Functional) Types of AI
Predictive AI forecasts future outcomes based on historical data. Examples include demand forecasting to optimize orders and inventory levels to minimize costs while satisfying customer demand. Predictive AI is also used in stock market predictions, predictive maintenance and weather forecasts. Its accuracy depends on the quality of the training data and biases in the data can lead to flawed predictions. [Link to previous article.]
Agentic AI systems act autonomously to achieve goals, often with reasoning and decision-making. Self-driving cars use agentic AI, as do some personal assistants (e.g., AutoGPT, Devin AI). Agentic AI will play an important role in automating complex workflows in procurement but may lead to unintended behaviors if not properly aligned within specific parameters and guardrails. [Link to previous article.]
A fourth type of AI that we have not yet considered is discriminative AI, which can be defined as a type of artificial intelligence where the objective is to opt for the best decision or choose the correct class for the input data by learning from training data. Unlike generative AI, discriminative AI models are trained to differentiate the data. They learn the relationship between the inputs and the outputs or labels/categorical variables. Typical use cases for AI models include image and speech recognition, natural language processing, facial recognition, and medical diagnosis systems. Discriminative AI has significant existing and potential applications in procurement and supply chain management, particularly for classification, prediction, and decision-making tasks. For example, discriminative AI can be used to classify suppliers based on performance metrics (including delivery time and quality, pricing) and external data (including ESG scores and financial stability) to recommend the best-fit suppliers within constraints such as cost, location, or sustainability goals.
Discriminative AI can likewise be used in fraud detection. In such applications it classifies transactions or supplier behaviors as “fraudulent” or “normal” by identifying anomalies in invoices, purchase orders, or payment patterns. You could also use discriminative AI image recognition to classify manufacturing defects in goods received from suppliers. In automotive supply chains, AI inspects parts for defects, reducing recalls. In short, discriminative AI outperforms generative AI in classification tasks. But it too relies on high quality training data, which should also be audited for biases.
Other Ways to Classify AI
Narrow, general or superintelligence
We have dealt with functional differences between types of AI, but there are other distinctions. The classic AI classification is by capability. Narrow AI (or weak AI) is artificial intelligence that is focused on one particular task. Siri is a good example of narrow AI, and many recommendation engines fall into this category.
By contrast, general AI (or AGI, artificial general intelligence) refers to systems that can perform many different kinds of tasks as well as a human, or even better. To do this, AGI must be able to handle many types of tasks, such as solving problems, understanding new information, or learning new skills, at a level similar to humans.
While generative AI excels at creating new content like text, images, and video based on learned patterns, it lacks the broader understanding, reasoning, and adaptability that characterize general AI. Generative AI is thus normally understood as narrow AI.
Artificial superintelligence (ASI) is a hypothetical software-based system with intellectual powers beyond those of humans across a comprehensive range of categories and fields of endeavor. An ASI system has yet to be developed and is only currently theoretical. ASI’s superior capabilities would apply across many disciplines and industries, and would include cognition, general intelligence, problem-solving abilities, social skills and creativity. There are concerns that ASI could one day present an existential threat to humanity.
Learning approach
A third way to classify AI is by the learning approach taken: supervised, unsupervised, reinforced, and self-supervised learning. AI that takes the supervised approach learns from labeled data (input-output pairs) to make predictions or classifications. It could do this with the classification of discrete labels, as is done to detect spam emails. Or it could use regression analysis of continuous values, for example in sales forecasting. Supervised learning is the backbone of discriminative AI in applications such as classifying defective versus non-defective products in manufacturing. The main technology tool for this purpose is a convolutional neural network (CNN), which is designed to learn and extract features from data automatically.
Supervised learning is also used to flag fraudulent credit card transactions and to predict customer churn (when customers will leave a service, such as a mobile telecom operator).
By contrast, unsupervised learning finds hidden patterns in unlabeled data without predefined outcomes. Thus, it is central to predictive analytics. Methods used include clustering (grouping similar data points) and dimensionality reduction (simplifying data). Key applications include market segmentation for targeted marketing, anomaly detection for predictive maintenance and cybersecurity, and supply chain optimization, clustering suppliers by reliability or geographic risk. Unsupervised learning doesn’t necessarily replace supervised learning but enhances it by uncovering latent patterns that make predictions smarter.
Reinforcement learning (RL) is ideal for sequential decision-making (for example, in autonomous systems). In daily life, we are most likely to experience this in dynamic pricing. Its core technology is deep-Q networks (DQNs). These use a neural network to approximate the Q-function, which indicates the expected cumulative reward for a given action in a given state. In other words, it trains agents to make decisions in complex environments through trial and error. By receiving rewards or penalties for their actions, RL agents can learn to make decisions that maximize their cumulative reward, improving their performance over time. Another application in business is warehouse robots optimizing pick-and-place routes.
Contrastive learning enables models to distinguish between similar and dissimilar data points and is used to retrain vision models for medical imaging.
Finally, there is self-supervised learning. This is central to generative AI.
Generative AI and Self-Supervised Learning
Self-supervised learning (SSL) trains models to predict hidden parts of input data (such as masked words in text or patches in images) without human labels, creating its own supervisory signals. SSL dominates generative AI for three main reasons. First, cost-efficiency: SSL eliminates the need for expensive labeled data. Second, scalability: SSL leverages abundant unlabeled data (such as internet-scale text/images). And third, transfer learning: SSL-pretrained models can be fine-tuned for specific tasks with minimal labeled data. SSL leverages several core techniques such as masked modelling, which randomly masks portions of data and trains the model to reconstruct the missing parts. For example, BERT (introduced by Google in 2018) predicts masked words in sentences, learning bidirectional context, which has application in language translation and document summarization.
Autoregressive modeling is widely used in generative AI to predict the next token in a sequence, the best-known example being GPT-4, which generates text word-by-word.
Advantages and Concerns about Generative AI
The biggest advantage of GenAI in functions such as procurement is organizational productivity. Gen AI can answer questions in a human-like style, reduce effort on tedious and monotonous tasks, provide accessible summaries of complex topics, produce automatic translations and transcriptions, etc. Most industry sectors and workplaces will be using a form of GenAI to enhance and optimize their work. It’s already integrated into much of our daily activity, for example Copilot within Microsoft Office.
Nevertheless, there are disadvantages. Care must be taken to guard against copyright infringement, misrepresentation or misinformation. Concerns have also been raised about so-called “deepfakes”. These can be used to disrupt business activity through invasion of privacy and erosion of trust. For example, cybercriminals can use AI-generated voice or video clones of executives (such as a Chief Procurement Officer!) to trick employees into authorizing fraudulent wire transfers or sharing sensitive data. A Hong Kong bank lost $35 million after an employee transferred funds to fraudsters who used a deepfake audio clip impersonating a senior executive.
GenAI is not, and is never likely to be, perfect. It can generate results that are false, misleading, or nonsensical, even though they appear to be plausible—so-called “hallucinations”. For example, it could fail to detect irony and sarcasm in customer feedback, leading to poor decisions. There is always need for human-in-the-loop (HITL) interventions.
Training GenAI requires huge amounts of power, which indirectly generates huge amounts of carbon. This has important consequences for climate change. An estimate of the electricity needed to train ChatGPT-4 is between 51,772 and 62,318 megawatt hours of electricity. This generated thousands of metric tons of carbon dioxide emissions.
GenAI Use Cases in Industry
The use of GenAI has provoked most interest—and controversy—in the creative sectors. But as we have already seen, GenAI brings many benefits to manufacturing in areas such as design optimization, process improvements, personalization, quality control and predictive maintenance.
Other sectors benefiting from GenAI include:
Healthcare/drug discovery: GenAI contributes to generating new molecular structures and estimating their characteristics, which speeds up the search for new drugs. Tools such as Microsoft’s DAX Express draft clinical notes from doctor-patient conversations.
Education: GenAI acts as a learner assistant, creating customized study guides or quizzes based on individual student performance (for example, ChatGPT is used to generate practice questions for biology students). GenAI tools such as Scribbr can also identify plagiarism in student reports.
Public sector: GenAI synthesizes public consultation feedback (for example, summarizing thousands of submissions into actionable insights). GenAI chatbots handle FAQs on topics such as tax queries and travel guidelines, and streamline processes such as visa applications, reducing wait times.
Finance: GenAI is widely used in fraud detection, analyzing transaction patterns to flag anomalies (e.g., JPMorgan’s COiN platform); automated reporting to draft financial statements and forecasts (e.g., Accenture’s pilots for F&A clients); and customer support with AI-powered chatbots to resolve banking queries (e.g., Bank of America’s Erica).
Transport & logistics: GenAI models predict traffic patterns to optimize routes and improve logistics efficiency. Generative models can be used to simulate driving scenarios to train self-driving systems. AI manages inventory and predicts demand spikes.
The Future of AI: Convergence
The future of AI probably lies in the convergence of generative AI with other AI paradigms (predictive, discriminative, agentic) and technologies (IoT, automation), creating integrated systems that amplify capabilities beyond what any single approach can achieve.
GenAI’s creativity will combine with predictive AI’s forecasting power for dynamic decision-making. In healthcare, for example, generative models will design drug candidates, while predictive AI will forecast clinical trial outcomes.
The combination of GenAI and agentic AI opens up exciting possibilities. GenAI powers content creation such as software code, while agentic AI autonomously executes the tasks for which the code is designed, such as scheduling and data retrieval. In future, multi-agent systems (such as AutoGPT) will orchestrate complex workflows, such as product design, which feeds into regulatory compliance and then marketing.
Emerging trends include neuromorphic AI: future LLMs may integrate brain-like computing for energy-efficient, real-time generation. Multimodal fusion will see models such as GPT-4 and Gemini processing text, images, and video simultaneously, enabling holistic problem-solving. Smaller language models (SLMs) like Meta’s Llama 3 will make hybrid AI accessible on edge devices, leading to greater democratization of GenAI.
Conclusion
Generative AI can help us achieve great things, but many, especially in creative sectors, believe that it comes at an ethical, social, environmental and human cost. This is less of an issue in other sectors and in functions such as procurement, but it is still important to consider these factors and support ethical frameworks around GenAI, as well as certain other forms of AI. Above all it is always vital that humans should check that the content produced is accurate.
