The Origins of Generative AI
The idea behind generative AI is older than today's popular AI assistants.
Early artificial intelligence systems were largely designed to classify information, recognize patterns, predict outcomes, or make decisions. Generative models introduced a different objective: learning the underlying patterns in data well enough to create new content that resembles the data on which the model was trained.
Several research developments contributed to today's generative AI systems.
Generative models before large language models
In 2013, researchers Diederik Kingma and Max Welling introduced the Variational Autoencoder (VAE), an approach for learning useful latent representations and generating data from them.
In 2014, Ian Goodfellow and researchers at the University of Montreal introduced Generative Adversarial Networks (GANs). GANs used two competing neural networks—a generator and a discriminator—to improve the quality of generated samples. The approach became influential in image generation and other generative applications.
These developments established important foundations for modern generative modeling.
The Transformer changed language AI
A major turning point came in 2017 with the publication of "Attention Is All You Need."
The researchers introduced the Transformer architecture, which relied on attention mechanisms rather than the recurrent architectures commonly used for sequence processing at the time. The architecture was more parallelizable and became the foundation for many subsequent large language models.
In 2018, OpenAI demonstrated the potential of combining Transformers with unsupervised generative pre-training. Its GPT research showed that a language model trained on large amounts of unlabeled text could subsequently be adapted to multiple language tasks.
This research direction eventually contributed to the modern foundation-model ecosystem, including today's large language models and multimodal generative AI systems.
From Chatbots to Enterprise AI Platforms
The first wave of enterprise generative AI adoption was dominated by experimentation.
Companies built internal chatbots, document summarization tools, marketing assistants, coding assistants, and question-answering systems. However, organizations soon discovered that simply connecting an LLM to a business application was not enough.
Enterprise systems have additional requirements:
Private and sensitive data protection
Access controls
Reliable information retrieval
Integration with existing applications
Auditability
Monitoring
Human oversight
Cost management
Consistent output quality
Protection against incorrect or unsafe actions
This has led to a new generation of enterprise architectures involving RAG, AI agents, tool calling, vector search, model gateways, evaluation systems, and enterprise data platforms.
The question is therefore no longer simply, "Can an AI model generate an answer?"
The more important question is:
Can the AI system reliably perform useful work inside a real business environment?
Real-World Applications of Generative AI in Enterprises
Generative AI can be applied across almost every knowledge-intensive business function.
1. Customer Service and Contact Centers
Customer service is one of the most mature enterprise applications.
AI systems can summarize customer conversations, retrieve information from knowledge bases, draft responses, classify requests, and in some cases conduct entire conversations.
Air India provides a large-scale example. According to Microsoft, its AI.g system handles approximately 40,000 customer queries each day across more than 1,300 types of questions. Microsoft reports that the system has resolved more than 13 million conversations with a 97% success rate.
This demonstrates how generative AI can move beyond a simple chatbot and become part of an operational customer-service platform.
2. Enterprise Knowledge Management
Large organizations have enormous volumes of documents: policies, manuals, contracts, technical documentation, reports, emails, and internal procedures.
Employees may spend significant time searching for information that already exists inside the organization.
RAG-based systems can connect an LLM to approved enterprise information sources. Instead of relying solely on information encoded in the model during training, the system retrieves relevant internal content and uses it as context for generating an answer.
Sun Life provides an example. Its "Sun Life Asks" system uses generative AI within the company's secure environment to answer employee questions. AWS reports that it resolved more than 600,000 internal queries during its first 11 months and now handles more than 10,000 queries per week.
3. Financial Services
Banks and financial institutions deal with large volumes of regulations, policies, financial documents, customer communications, and research.
Generative AI can help employees summarize documents, retrieve regulatory information, prepare reports, support investigations, and automate repetitive knowledge work.
Raiffeisen Bank International, for example, developed an internal AI system using Azure OpenAI Service and Azure AI Search. Microsoft reports that the system was designed to help employees with repetitive tasks such as documenting intelligence and summarizing legal, regulatory, and banking documents.
The important architectural principle is that enterprise AI should operate within controlled data and security environments rather than simply exposing sensitive corporate information to an uncontrolled public chatbot.
4. Healthcare and Life Sciences
Healthcare organizations can use generative AI for clinical documentation, knowledge retrieval, research support, administrative workflows, and patient-service applications.
Life sciences companies are also exploring AI for drug discovery and scientific research.
Genentech, for example, developed a generative AI system called the gRED Research Agent to automate parts of the process involved in analyzing scientific information for drug discovery and biomarker validation. AWS says the system is expected to save nearly five years of manual effort in biomarker validation across therapeutic areas.
Such applications demonstrate an important distinction: generative AI is not limited to writing text. It can also serve as an interface for complex information-processing workflows.
5. Employee Productivity
Another major enterprise application is the AI workplace assistant.
Employees can use generative AI to:
Draft emails
Summarize meetings
Create reports
Analyze documents
Generate presentations
Search internal information
Write and explain code
Prepare customer communications
Pierre Fabre developed an internal enterprise generative AI platform called PLA.I.GROUND using Microsoft Azure and Azure OpenAI Service. Microsoft reports that the platform was being used daily by more than 50% of employees and was designed to streamline work and create more time for innovation.
The value in these applications is often cumulative. Saving a few minutes on thousands of repetitive tasks can create substantial productivity gains across a large organization.
Enterprise AI Case Study: Novo Nordisk
Novo Nordisk illustrates another direction for enterprise generative AI: giving employees controlled access to AI so they can develop solutions for different business problems.
According to AWS, more than 25,000 Novo Nordisk employees have used its generative AI environment to create applications covering more than 2,500 use cases. The company began experimenting with large language models and generative AI in 2023 and subsequently focused on making experimentation available within a secure enterprise environment.
The lesson from this approach is that enterprise AI adoption does not necessarily require one centralized application.
Instead, organizations can create a secure platform on which business teams can develop multiple use cases while maintaining common security, governance, and infrastructure controls.
Enterprise AI Case Study: Oriserve and Multilingual Customer Engagement
Generative AI becomes particularly interesting when businesses operate across multiple languages and regional markets.
India-based Oriserve develops generative AI agents and conversation analytics for customer engagement. AWS reports that its solutions support more than 120 languages and regional dialects.
AWS also reports that Oriserve reduced its cost to serve by 50%, reduced model fine-tuning time from weeks to days, and achieved 90% speech-recognition accuracy across multiple dialects.
One example described by AWS involves Vodafone Idea (Vi), where an AI agent engages customers during the porting process and attempts to address reasons for potential customer churn. AWS reports that the solution increased customer retention rates by 75%. These figures are company-reported results and should therefore be evaluated in the context of the specific implementation.
From Generative AI to AI Agents
The enterprise AI market is increasingly moving from systems that generate information toward systems that can perform actions.
A traditional generative AI application might answer:
"What is the status of this customer's order?"
An agentic system could potentially:
Identify the customer.
Retrieve the order.
Check inventory.
Review shipping information.
Determine the appropriate action.
Update the relevant system.
Inform the customer.
This creates significantly more value—but also introduces greater risk.
An AI system that only generates text can produce an incorrect answer.
An AI system connected to an ERP, CRM, payment system, or operational database could potentially make an incorrect business transaction.
That is why enterprise agentic AI requires stronger controls around permissions, validation, transaction handling, logging, and human approval.
Why Generative AI Projects Fail to Scale
The technical ability to build a prototype is no longer the primary challenge.
The difficult part is productionization.
Organizations often encounter problems with:
Poor-quality enterprise data
Incomplete retrieval systems
Hallucinations
Inconsistent responses
High inference costs
Slow response times
Security and privacy requirements
Legacy-system integration
Insufficient evaluation
Lack of monitoring
Unclear ownership
Employees not knowing when to trust AI output
McKinsey's 2025 research illustrates the gap: while 88% of respondents reported AI use in at least one business function, only 7% said AI had been fully scaled across their organization.
This suggests that enterprise AI maturity is not simply about access to models. It is about redesigning workflows, integrating systems, establishing governance, and measuring business outcomes.
What a Production-Ready Enterprise GenAI Architecture Looks Like
A mature enterprise implementation generally contains several layers.
Data layer: Enterprise documents, databases, applications, and knowledge repositories.
Retrieval layer: Search, embeddings, vector databases, metadata filtering, and RAG pipelines.
Model layer: One or more foundation models selected according to cost, capability, latency, privacy, and task requirements.
Orchestration layer: Agents, workflows, tool calling, business rules, and context management.
Integration layer: CRM, ERP, HR, finance, customer-service, and other enterprise systems.
Governance layer: Authentication, authorization, monitoring, evaluation, auditing, privacy, and security controls.
This architecture allows enterprises to treat generative AI as software infrastructure rather than as a standalone chatbot.
The Next Stage of Enterprise Generative AI
The next stage of enterprise adoption is likely to involve increasingly integrated AI systems.
Organizations are moving from:
Chatbots → Copilots → RAG systems → AI agents → Multi-step enterprise workflows
At the same time, the role of humans remains important. AI can generate recommendations, summarize information, identify patterns, and execute approved actions, but organizations still need appropriate human oversight for high-impact decisions.
The strongest enterprise implementations therefore focus less on creating impressive demonstrations and more on measurable operational outcomes.
Conclusion
Generative AI has evolved through several decades of research, from early generative models and neural architectures to GANs, Transformers, large language models, and today's enterprise AI agents. The Transformer architecture introduced in 2017 was particularly important in enabling the modern generation of language models.
Today, enterprises are applying generative AI to customer service, financial research, employee productivity, healthcare research, knowledge management, software development, and business automation.
The real opportunity is no longer simply generating text or images. It is connecting generative models to an organization's data, processes, applications, and people.
The case studies from Air India, Sun Life, Novo Nordisk, Oriserve, Genentech, and Pierre Fabre show different approaches to this transition—from customer-service automation and employee knowledge systems to scientific research and enterprise-wide AI platforms.
For enterprises evaluating generative AI in 2026, the central question is therefore changing from "What can generative AI do?" to "Where can generative AI create measurable value, and what infrastructure is required to deploy it responsibly at scale?"
This article was originally published on Perceptive Analytics. At Perceptive Analytics our mission is "to enable businesses to unlock value in data." For over 20 years, we've partnered with more than 100 clients — from Fortune 500 companies to mid-sized firms — to solve complex data analytics challenges. Our services include AI consulting services and Power BI consultants, turning data into strategic insight. We would love to talk to you. Do reach out to us.
Top comments (0)