What's AI foundational models? How it can become the iPhone moment of AI development

Foundational models such as GPT-4, have been a seismic force in technology. Chatgpt’s sudden rise rekindled memories of the iPhone moment. It has caused tech companies worldwide to reshuffle their decks, putting AI at the top of their strategies. This is more than just a fleeting tech trend. Foundational models are reshaping the core principles of AI software development.

However, these changes are not yet well understood, and several teams continue to apply their old playbook to this new paradigm. In this article, we'll first examine the challenges with traditional AI. Then, we'll uncover how foundational models address those pain points, and finally, we'll explore how product leaders and other decision-makers can apply these insights to accelerate their AI roadmap.

Challenges with traditional AI product development
Previously, AI product development was a complex and cumbersome process, demanding vast resources. It required extensive data collection and preparation, as well as a profound technical understanding of machine learning (ML) to select the right model architecture and parameters. Finally, the model had to be tested, validated, and constantly monitored for real-world performance. All these steps necessitated a large team of data scientists and ML engineers.


For many companies, building this capability in-house has been a big challenge. The process was made even more daunting by a shortage of data scientists and ML engineers. Traditional AI software businesses face scalability and margin challenges. These stem from the need for significant, customer-specific work and input costs beyond typical support and success functions. Consequently, delivering AI solutions as full-fledged SaaS offerings was often difficult.

Many businesses scale their AI efforts on achieving the right ROI. For example, one of the top healthcare analytics companies in the US developed ML models to detect billing fraud. However, even after a year of effort, they were not able to achieve the desired accuracy. Still, they had to rely on nurses to verify medical coding data extracted by their ML models. As a result, they were unable to scale this initiative.

How foundational models are changing the rules
Foundational models are rewriting the rulebook for AI development in several transformative ways, directly addressing the pain points highlighted above:

Reduced need for custom models: Many customers wonder if they can train their own version of ChatGpt using their proprietary data. Not only is training a custom model excessively expensive, with costs ranging from $40 to $100 million, but it's also unnecessary. The good news is that instead of building custom models, you can now leverage existing foundational models by guiding them with specific data. Techniques such models or fine-tuning offer powerful ways to adapt general-purpose foundational models to your task-specific needs, without having to reinvent the wheel.

Streamlined development teams: As Silicon Valley's prominent venture capitalist Chamath Palihapitiya has said, "You no longer need a 40-50 person team to build minimum viable products (MVPs); you can now do it with just 3-4 people." Given the reduced need for building custom models, the overhead of creating AI products has been substantially lowered.

Saas scalability: AI's newfound capability has paved the way for delivering AI products as scalable SaaS solutions. Human involvement is minimized, enabling products to reach end-users more swiftly and directly. Consider healthcare, where medical documentation is a significant burden for providers. To address this challenge several companies have sprung up, claiming to use AI to translate patient-provider conversations into medical notes. However, all of them require human scribes for validation, a process that takes hours. Now, thanks to large language models, new solutions are emerging that can produce accurate notes within seconds after a patient encounter, all packaged in a convenient SaaS solution.

However, all of them require human scribes for validation, a process that takes hours.

Focus on speed to market: With lower technological barriers, the race to get products to market quickly is more vital than ever. Many companies invest a lot of time and effort in collecting data upfront and perfecting the model. Instead, companies should embrace a minimum viable product approach, one that fosters rapid development, iteration, and continuous improvement. Entering the market early allows them to collect valuable usage data, understand customer needs and preferences, and fine-tune the foundational model accordingly. This nimble approach provides a distinct advantage, setting them ahead of competitors.

Implications for decision makers in an enterprise
These shifts have several strategic imperatives for product leaders and other decision-makers in an enterprise, aiming to leverage AI in their operations or infuse it into their products and services:

Lead with a SaaS-first approach: Several companies in the business of AI operate like service businesses, creating custom solutions for their clients. However, foundational models, with their wide-ranging capabilities, enable the delivery of these solutions as highly scalable SaaS products. This requires a complete reevaluation of existing business models.

Embrace foundational models: Resist the urge to build from scratch. Utilize foundational models, fine-tuning them to meet specific needs. This approach can save both time and resources and help you scale your AI efforts using a SaaS approach.

Choose the right foundational model for your needs: Evaluate models based on their suitability for your particular application. While the most advanced foundational model may be ideal for general reasoning tasks like sentiment analysis, other models might be more effective for specialized tasks, such as medical image analysis
Invest in upskilling: As the AI landscape evolves, so too must the skills of your ML and data science teams. Investing in training in areas like prompt engineering, AI compliance, and AI product management will be essential to ensure continued success in your AI initiatives.

As we stand on the cusp of this new era, companies involved in using or selling AI products must wholeheartedly embrace these changes, reevaluate their strategies, and exhibit agility to capitalize on the abundant opportunities that foundational models offer. By doing so, they can position themselves at the forefront of this technological evolution.

(The writer is Senior Director of Cloud & AI Corporate Strategy at Microsoft; Former McKinsey & Company Engagement Manager)