The Smartest Way to Adopt AI? Build an R&D Engine First

Johnson & Johnson (J&J) is a well-known medical equipment, medicines, and consumer packaging manufacturer, active since 1886. Recently, they gave us a glimpse of their AI strategy. Over the last two years, J&J launched nearly 900 generative AI projects across departments. This “let a thousand flowers bloom” strategy’s goal was to identify valuable AI applications within the organization. Now, after two years of experimentation, here’s what they found: 10% to 15% of these AI use cases delivered approximately 80% of the total value. They are now shifting their strategy to focus on the use cases that drive the highest value.

J&J isn’t alone in this strategic approach to AI. Visa’s President of Technology, Rajat Taneja, said in November 2024 that the company already has more than 500 generative artificial intelligence applications in use. “This is a time when I think we have to innovate very fast,” Taneja said in an interview.

Visa or J&J’s strategy provides a critical lesson for organizations aiming to remain competitive in the AI era: to adopt AI, you need to build an R&D-first mindset within your organization, experiment with thousands of use-cases, and then choose the ones that offer the highest ROI. Below, I will break down why this is the only strategic approach that you can take with AI.

Shift in Strategy

This R&D-first strategy that J&J and Visa deployed is very very different from how technology products were built in the past in larger firms: You typically began with a long planning cycle, built detailed requirement specs (PRDs), aligned teams around milestones, and deployed software or process automation that was meant to last for years. This approach worked well when each layer of the architecture was deterministic, where capabilities were well-defined, behavior was predictable, and outcomes could be tightly controlled.

However, with generative and agentic systems, the landscape has shifted. To understand this shift, you have to recognize these systems for what they are: a fundamentally new computational paradigm that we are witnessing unfold in real-time. This new paradigm touches upon everything from how users interact with machines to how we view unstructured data, how products make decisions, or how data is stored and retrieved.

This shift challenges long-held assumptions about how software is built, tested, and deployed. It redefines what a “product” is and demands a complete rethinking of the architecture.

What History Teaches Us

Every major technological shift has rewritten the rules of product development.

In computing, when transistors replaced vacuum tubes, it enabled the creation of smaller, faster, and more reliable machines, eventually paving the way for personal computing. Companies like IBM recognized this shift early and evolved from mainframes to PCs. In contrast, firms like Digital Equipment Corporation (DEC), which dominated the minicomputer era, failed to adapt. DEC dismissed the need for PCs, clinging to their existing model, and eventually faded into irrelevance.

The emergence of the internet provides many such examples. Amazon is a classic one – began as an online bookstore, became a logistics powerhouse, and later the largest provider of cloud infrastructure through AWS. Meanwhile, Borders Books, once a retail giant, outsourced its online sales to Amazon. By the time it realized the implications, it was too late. Borders filed for bankruptcy in 2011.

BlackBerry failed to see that touchscreens, app ecosystems, and user-centric design were going to bring a seismic shift to mobile devices. They insisted on sticking to keyboard-driven devices and enterprise markets, and eventually lost market share and had to exit the smartphone business. Their story has been captured in an award-winning film.

The fact of the matter is, predicting the future amid unfolding innovation is inherently challenging. We have our cognitive biases, and technological progress often unfolds in nonlinear, exponential, and unpredictable ways. The challenge of managing emerging technologies lies in a paradox – the Collingridge Dilemma. In the early stages, the consequences of a technology are hard to predict, making it difficult to guide or regulate effectively. Yet by the time those consequences become clear, the technology is so deeply embedded in systems, products, and society that changing course becomes extremely difficult and expensive.

However, despite the challenges, the risk of ignoring AI is far greater than the discomfort of uncertainty. What makes the current moment different is the pace and breadth of change. AI is not just altering a single layer of the stack, it’s reshaping the entire product lifecycle: from how we gather insights, generate ideas, design interactions, and deliver value, to how systems evolve post-deployment. To wait for AI to “settle” before taking action is to miss the wave entirely.

Why AI May Be Bigger Than the Internet

In a speech at The AI Action Summit, Google CEO Sundar Pichai described artificial intelligence as a “fundamental rewiring of technology”. He emphasized that AI represents the most significant technological shift of our lifetimes, potentially surpassing the impact of the internet itself. Pichai has consistently highlighted AI’s transformative potential. In a 2018 interview (four years before ChatGPT’s public release), he remarked, “AI is probably the most important thing humanity has ever worked on. I think of it as something more profound than electricity or fire.”

Yet, many underestimate the potential long-term impact of AI, thinking that ‘AI’ is limited to ChatGPT-like chatbot interfaces, or are disappointed when a claimed future AGI technology fails to solve a simple mathematics problem. One reason the modern iteration of AI is so hard to grasp is that it defies the mental models we’ve used to understand previous waves of technology. The internet connected people and information, but it largely operated within fixed interfaces and deterministic logic.

AI, on the other hand, is increasingly multimodal, and so its touchpoints are numerous. We can now build systems that use an ensemble of AI models to reason over text, image, video, audio, and sensor data (LiDAR, thermal imaging) simultaneously. With multimodal systems, we are not restricted to screens, buttons, and keyboards as our primary interfaces – AI systems can talk, see, sense, and act as a bridge between us and data.

This has implications for how we interface with technology. If the internet forced us all to type or use touch screens, AI is flipping that relationship. We’re moving toward systems that don’t sit on screens, but fade into the background, ubiquitous, context-aware, and woven into the fabric of everyday life.

In many ways, we have arrived at the future that philosopher Jean Baudrillard warned us of, the ‘Hyperreal’ – where representations replace reality. What does the world look like when the bridge that connects us to each other and to our data can imitate and extend human intelligence?

Reimagining the Future

If the rules of product development and human computer interaction are being rewritten, you must rethink how you engage with your users and what users get out of your system. You have to begin by asking fundamental questions:

• What is the unit of value in this new world? What do users expect?
• Is your product static, or a system that learns and evolves around the user?
• Is the user still the one issuing commands, or does the system and the user learn from each other?
• How does trust get established when systems behave probabilistically?
• What are the different interaction touchpoints and their modalities?
• How do I make use of the volumes of data around user interaction I have from the past?
• And most importantly: What role should your product play in the life of a user who is surrounded by intelligent systems?

When you bring LLM or vision AI systems into the mix, you will find that it require a fundamental rethinking of UX, value delivery, and long-term product evolution. And once you ask these questions, you will soon realize that the architecture, the team composition, and the tactics through which you introduce products to the market have altered.

I have seen businesses reach an ‘aha’ moment when their mindset shifted from “How can we use AI to improve what we already do?” to “What would this look like if we built it AI-first from the ground up?” It required them to reframe the problem, rather than search for a solution right away.

In other words, you have to first ‘unbundle’ your product stack, and then ‘rebundle’ it for an AI future. This requires going back to the lab and multiple iterations to get it right.

However, if you had to bring in a change in mindset within your organization, where do you start?

Organizations that thrive with AI view internal AI R&D not as a cost center, but as an IP-generation engine. The insights you gather, the models you fine-tune, and the datasets you build all become proprietary assets. They compound over time, forming the bedrock of competitive advantage that cannot be copied overnight.

This is why you will notice that increasingly higher volumes of research are now being published by corporations, and approximately 70% of individuals with a PhD in artificial intelligence are employed in private industries, a significant increase from 20% two decades ago.

You have to shift from predefined user stories, scoped features, and linear roadmaps towards innovation models where you build for emergence – where you don’t fully know the solution in advance, but discover it through experimentation, iteration, and learning in context.

A Playbook for R&D-First AI Development

Like every well-oiled machine, you have to create an innovation system within your organization. A successful AI R&D system is built on repeatable infrastructure, institutional memory, and leadership that rewards exploration, not just execution.

To build this system, the playbook is becoming increasingly evident:

First, start by identifying high-leverage points: places in your product or operations where AI can drive nonlinear impact. These typically include parts of your workflow that require repetitive decision-making (by humans or programmed rules), unstructured data handling, or touchpoints with rich customer interaction history.

Then, move on to prototyping for emergence. Instead of waiting for perfect clarity or fully scoped roadmaps, create fast, low-cost AI testbeds to run pilot projects. If you use modern tooling, you can put together workflows fairly quickly, albeit with lower accuracy or efficiency. With a prototype in place, test assumptions, refine prompts, and measure real-world impact. Create reusable assets like prompts, embeddings, and synthetic datasets that should be stored in internal AI libraries to enable compounding reuse across teams, parts that form the building blocks of an AI-native architecture.

Capture the learning from these initiatives and institutionalize them as proprietary IP. The models you fine-tune, the internal datasets you build, and the edge cases you solve become durable assets and create a moat that compounds over time.

Finally, productize the ones in which you see the highest impact. Treat the rest as learning that you can tap into later. Productizing AI brings in new pieces into your architecture, such as MLOps systems, queueing pipelines, or parallel ingestion systems, that you hadn’t envisioned before. You may also need to combine an ensemble of AI models, leveraging smaller models for simpler tasks and keeping the larger ones for tasks that require reasoning.

The faster your organization can cycle through experimentation, learning, and productization, the stronger your AI advantage will become. The eventual goal is to zero in on a few AI workflows and use-cases that can deliver massive value in the long run.

The Building Blocks

Suppose you or your organization has decided to embrace a culture of exploration – where do you start? How do you cut through the noise?

As I explained above, you should start with high-impact use cases. You also should evaluate the data you have at hand and the data you generate across interactions daily. Once you take a stock of the current architecture, then drill down into the technology stack.

Here are key AI technologies that forward-thinking companies are experimenting with today:

• Large Language Models (LLMs): Foundation models like GPT-4o, Claude, Gemini, and open-source alternatives like Mistral and LLaMA, to understand human language, reason over multimodal data, and perform actions.
• Fine-tuning, Knowledge Distillation, and Supervised/Unsupervised Learning: Supervised, unsupervised, and semi-supervised learning to tailor models for your domain. Fine-tuning large models to internal datasets or knowledge distillation to compress complex models into lightweight, deployable variants without major performance loss.
• Retrieval-Augmented Generation (RAG): Combining LLMs with internal knowledge bases (via vector databases (like Qdrant, Weaviate, or PGVector/pgai) or Knowledge Graphs (like Neo4j), or your existing data sources for advanced retrieval over unstructured data.
• Multimodal AI: To build systems that interpret and reason over text, image, audio, video, LiDAR, and thermal sensor data—e.g., CLIP, GEMINI 2.5, and Segment Anything (SAM) by Meta.
• Agentic Frameworks: Tools like LangGraph, CrewAI, and AutoGen to build multi-agent workflows where AI agents reason, plan, call tools, and collaborate to complete complex tasks.
• Vision AI, Edge AI, and OCR Pipelines: Tools like YOLOv8, Detectron2, and LayoutLM for structured data extraction and image/video analysis from video and image data. DeepStream SDK for scalability.
• Speech and Audio Intelligence: Models like Whisper (transcription), Bark (text-to-speech), and AudioCraft (sound generation).
• Time-Series + Predictive AI: Models built using Temporal Fusion Transformers (TFTs) or platforms like Nixtla and GluonTS to predict demand, detect anomalies, or forecast trends.
• Personalization & Recommendation Engines: Using reinforcement learning with feedback loops to train and build recommendation engines.

This list is not exhaustive – newer approaches are emerging every day, and you have to learn to accommodate them in the future.

For instance, context windows are rapidly increasing. Latest models (such as Llama4) now feature 10 M+ context window length, which means you can fit a lot more data within your prompt. Cost per token of LLM inference has also been dropping, which means that you can now leverage LLMs for a wide range of workflows. With longer context windows and lower inference costs, you can afford to bring in rich history, structured inputs, and domain knowledge directly into the prompt, turning LLMs into lightweight reasoning engines.

Future Notes

AI is not a one-time integration or a vendor decision. It’s a strategic transformation that requires you to rethink your architecture, retrain your teams, and reinvest in how your company learns.

In the end, building an R&D-first culture is not just about adopting AI, it’s about developing a dynamic capability to evolve with it. As models get smarter, context windows grow, and costs drop, the real differentiator won’t be access to AI; it will be your organization’s ability to experiment, adapt, and compound knowledge faster than your competitors.

The most strategic move you can make today is to build your internal AI R&D engine. Everything else will follow.