Since then, we have witnessed several tectonic shifts both in terms of technological advancements and how the larger ecosystem of model providers, system integrators and industries responded and adapted.

Through the pandemic era, we saw businesses adapting to become more resilient by building their digital core and accelerating their move to the cloud, digitalizing their core processes to best serve their customers. This shift paved the way for the upcoming AI led disruption, as organizations already had the foundation in place.

With the Gen AI revolution, businesses were far more agile and adaptive in swiftly evolving their digital core to an AI powered cognitive core to help them unlock efficiencies at scale, amplify human potential and deliver exponential impact. They have moved from simple proof-of-concepts to full scale production grade deployments in knowledge management, AI augmented software development, service management, documentation and collateral generation etc.

These organizations are further building on this momentum and setting themselves on a trajectory to be “AI-first”, where AI is bringing both parts of the enterprise brain together, by launching new products and services, enhancing existing offerings, improving operational efficiencies, and building a sentient ambidextrous organization that connects multiple signals to generate the right insights at the right time. Like the previous cloud and digital revolution, enterprises that had the proper strategy were able to leapfrog and gain a distinct advantage.

Below I have listed 10 major aspects of a successful AI-first strategy:

1. A platform-based, poly-AI approach is the ideal approach to stay future-proofed and democratize effectively:

Models are becoming perishable. In the last year, as per the Stanford HAI index report, a total of 149 foundation models were released, more than double the amount released in 2022. We are already looking at beating that number in the first half of 2024 itself. As newer models with more efficient design, trained with better quality data and having enhanced capabilities are released, enterprises would need the flexibility to switch between models and deployment architectures. It is important to build an abstraction layer that allows enterprises to select and integrate AI providers, models, micro-AI platforms, and tooling that best suit their unique requirements.

A poly AI approach helps leverage the best models for the right task, ie having specialized models for code generation, one for summarization, one for report generation and another for customer service etc. This is best done by creating a flexible enterprise grade platform that has capabilities of rapidly building, finetuning and deploying models. For a major telecom player, we built a similar platform with features like semantic search, summarization, conversational AI, and text to code that catered to 50,000+ users and delivered millions in savings.

2. Structured Discovery Approach to unlock value:

A structured discovery approach is required to unlock possibilities and enable us to identify the right use cases where AI can make an impact. Strategic AI value map analysis identifies high business impact areas rather than siloed use cases. At Infosys, we do this by leveraging our verticalized blueprints, industry catalogues, AI canvas. We consolidate and prioritize use cases with maximum impact and ROI leveraging AI Radar and refine and detail out the use cases using AI & Automation Canvas, which are our specialized assets. We have created playbooks for industries that clearly lay out a structure to embed and mature Gen AI into core processes and operations. Increasingly, Gen AI is getting embedded in all aspects of day to day life, so no industry can afford to overlook infusing it in their core products and services.

In our own IT services landscape, we have applied the same value map analysis to reimagine our services and offerings, and transformed how we approach application development and maintenance, IT operations, service management, legacy modernization etc. For example, in application development, we are using Gen AI for code generation, test case generation, documentation generation, Project Planning & Analysis, User Stories & Backlog Development, Refactoring. In IT infrastructure maintenance we are using Gen AI for automated resolutions and self-healing. It’s also used in migration and modernization of activities including data migration tasks by automating data cleansing, transformation, and mapping, analyzing code, documentation, specifications, and user manuals associated with legacy systems.

3. Human + AI Approach i.e. AI Assistants for everyone:

At Infosys, our primary aim is to amplify human potential, aligning with our company’s purpose. To realize this, we have developed multiple AI assistants tailored to various roles. For developers, a code assistant enhances productivity in tasks like coding, testing, and documentation. A consultant will have a knowledge assistant to help him retrieve the best knowledge assets with the least turnaround time. A personalized learning assistant supports continuous learning as per unique needs, while a sales assistant consolidates collective knowledge for client-facing teams.

4. Domain Adaptation is the key, specialized models are outperforming general models:

AI adoption is evolving from personal to specialized and custom AI applications using closed models and APIs. As Gen AI gets democratized, enterprises will create narrow Gen AI by fine-tuning the foundational generative AI open models on specific enterprise data to create custom applications. The focus will shift to industry-specific AI applications using specialized pre-trained models to deliver exceptional accuracies in specific domains or tasks.

There are two different approaches to using large language models (LLMs). One is to scale up the model size and increase the performance of general-purpose models that can handle various tasks. Large companies and AI startups are competing via this approach to build the biggest and most efficacious models such as GPT-4 with over a trillion parameters. The other approach is to scale down the model size and fine-tune open access models for specific domains and tasks via finetuning methods like PEFT. Despite its current lack of popularity, Infosys has used this approach which we term as the “narrow transformer” successfully and believes enterprises will follow suit for customized and cost-effective solutions with the requisite data privacy and security. Both scale-up and scale-down approaches have their respective advantages and objectives. They both work with a base of closed-access models and open-access models. Big and powerful models, often proprietary, are good for retrieval-augmented generation (RAG) and are used in business applications such as dialogue systems, semantic search, question answering, and summarization. They do not require any model adaptation. However, for specialized tasks where customization and cost are important, fine-tuning of open access models is the path to success. Open-access models are best suited for auto-completion tasks such as code completion and machine translation. These models are much more efficient and effective when they are fine-tuned with instructions using supervised learning or through extended pre-training with self-supervised learning. At Infosys we have deployed 7+ specialised fine-tuned code models for Oracle, SAP, Finacle etc in production that are able to deliver 80%+ accuracy. We believe that organizations that build these specialized models by augmenting and finetuning with their organization’s knowledge, shall sustain a long-term competitive advantage.

5. The Bimodal approach:

There is a need to innovate at speed, and also to innovate at scale. To tackle both these, we have our bimodal approach of AI foundry and AI factory. Establishing an AI foundry to experiment and incubate new technologies and develop new patterns and use cases will help the enterprise innovate at speed. The AI-factory-like approach will help bring in extreme automation and productization of learnings from the AI foundry. This approach will help balance and manage the risks associated with AI evolution while scaling its adoption within the enterprise. Both these approaches have their own platforms, tools and accelerators for implementation.

6. Regulations are evolving, but enterprises need to have a strong focus on Responsible AI to minimize risks:

The regulatory landscape is transforming rapidly, and different provisions of the EU AI Act is coming into effect this year, further catalyzing the growth of regulations around the world. There are myriad risks also like bias, privacy, security vulnerabilities, lack of transparency, copyright infringement etc. Organizations need to address this via a three-pronged approach. Firstly, they need to build automated technical guardrails, that intelligently detect and mitigate these threats in the input and generated output, with mechanisms to explain the rationale behind it. Secondly, they need to build process guardrails to ensure Responsible AI by design, ie embedding ethical consideration throughout the AI lifecycle, from data preparation and training to testing and inferencing. At Infosys, we took this project on earlier than most and we are the first organization to be certified in ISO 420001:2023 for AI management systems. Thirdly, enterprises need to institute and streamline AI governance via a centralized point of accountability to ensure safe use of AI complying with laws and regulations by conducting reviews, assessments and audits, continuous monitoring and developing and enforcing our AI policies. Infosys has launched the Responsible AI Office which is an inter-disciplinary body cutting across legal, privacy, security and AI professionals for this.

7. It is an ecosystem play, deep integration within the AI value chain is needed:

The AI value chain consists of multiple hardware (specialized CPU and GPU’s) and platform providers, like NVIDIA, Intel, etc, hyperscalars like Microsoft, Google, AWS etc. Increasingly specialized AI computing hardware and software stack are coming as tightly coupled ie, the models are optimized to deliver best performance on the specialized hardware stack. There are also multiple startups that are building specialized solutions for particular business problems, as well as accelerators for securing and accelerating generative AI deployments. It is imperative for AI first business to have a rich ecosystem of partners across the value chain. Through the Infosys Innovation Network we are able to collaborate with a rich ecosystem of start-up partners working in different arenas which increases the business value of our AI projects. Additionally, as the AI landscape is evolving rapidly, and new services, offerings are being released, we need a listening post or watchtower function in enterprises that will monitor the market and continuously scan for the best-fitting solutions.

8. Robust Data Management is the backbone:

The enterprise data architecture has not evolved for several years, and potential value remains locked in a hybrid of cloud-based systems and there is still a large footprint in legacy monoliths. Organizational knowledge remains locked in documents, systems, and an aging workforce. There is a need for a robust and responsible data governance, and data strategy for value maximization. A proper data architecture and platform enables rapid accessibility, discoverability and builds a robust backbone on which Gen AI delivers impact.

9. Executive sponsorship:

Executive sponsors have the authority and influence to champion the AI initiatives. Their endorsement lends credibility and encourages broader acceptance and adoption of AI initiatives within the organization. They also need to be involved in building a roadmap, fast-track resource acquisition, overcome hurdles by securing stakeholder buy-in, resolve conflicts during transition, and building strategic partnerships. To build an AI-first enterprise, AI pursuits can no longer be solely driven by data scientists and ML engineers in siloes. AI projects are increasingly becoming multi-faceted; use-cases are cross-cutting with critical social, legal, privacy, and IP considerations. Executives can ensure harmonious cross-functional, inter-disciplinary, multi-level collaboration to build resiliency and scale beyond the PoC/pilot phase.

10. Having a robust AI Talent strategy is key to getting the most out of the investments:

AI-first approach includes our talent strategy as an important pillar, with focus on three levels of enablement and upskilling.

• Level 1 is called AI Aware, wherein we are working on making everyone aware of generative AI technologies, basic operations like prompt engineering, responsible usage and how AI assistants can help them be more productive and relevant to clients.
• Level 2 refers to AI builders who can reimagine experience and processes to build industry-specific AI-led solutions.
• Level 3 refers to masters who understand the under-the-hood workings of ML (machine learning), DL (deep learning) and LLMs.hey are working on harder problems like fine tuning, pre-training, runtime optimization and responsible AI. With the right investments in training and enablement, we are able to drive adoption and employees are excited about the potential of technology, as it will help amplify their potential.