Cognitive SaaS: Building AI-native solutions with lasting competitive advantage
Cognitive SaaS integrates advanced AI, specialized knowledge, and learning capabilities to create solutions that adapt, evolve, and support user decision-making.
Ciao,
I took a longer-than-expected break from Radical Curiosity — since March 17, to be precise. During this time, I didn’t feel I had anything particularly valuable to share about my work, so I decided to take a step back, study, reflect, and focus on my projects.
A conversation with an investor friend about startups built on LLMs (Large Language Models) and a recent article on the Sequoia blog (Generative AI’s Act o1) inspired me to start writing again.
Let’s pick up from here.
Table of Contents
Cognitive SaaS: Building AI-Native Solutions with Lasting Competitive Advantage
Claudio Erba: from 0 to 200
Audiobooks: the power of the perfect narrator
Cognitive SaaS: Building AI-Native Solutions with Lasting Competitive Advantage
Recently, I had an insightful conversation with a friend of mine, an investor, about the growing trend of SaaS products built using large language models (LLMs). He expressed concerns that building SaaS on top of LLMs has become too easy and that many of these services lack defensibility because they can be replicated with minimal effort. This perspective is shared by many investors, who have grown increasingly skeptical of startups heavily reliant on LLMs. These startups are often dismissed as “AI Wrappers”—SaaS products that merely add a user interface or minor functionality around an existing AI model without contributing any significant innovation or proprietary technology.
This skepticism is understandable. The accessibility of LLM APIs has indeed made it simple for developers to create applications that superficially enhance existing models. These applications often lack the depth and uniqueness required to maintain a competitive advantage, raising concerns about their long-term viability.
However, it’s important to acknowledge that not all AI-powered applications fall into this category. While some may be simplistic wrappers, others incorporate advanced architectures that offer far more than just an interface to an LLM. Despite this, there hasn’t yet been a widely accepted term to distinguish these more sophisticated solutions from AI Wrappers. I propose the concept of Cognitive SaaS.
Anatomy of a Cognitive SaaS
A Cognitive SaaS involves a combination of components designed to assist and augment human decision-making and maintain human oversight in critical tasks.
Generic LLM (Large Language Model) for Reasoning: The heart of the framework is a foundational LLM that provides general reasoning capabilities. This model is a versatile engine for understanding natural language, making inferences, and generating responses. It acts as the essential cognitive substrate, enabling the system to engage in foundational reasoning and user interaction. The LLM doesn’t have specialized domain knowledge but is capable of general comprehension and reasoning tasks, allowing for flexible, broad applicability across domains.
Specialized Representation of Knowledge: This element augments the LLM with a specialized knowledge layer. It could be a database, ontology, or a structured representation tuned to specific domains, enabling the system to access precise and context-specific information. This specialized knowledge helps bridge the gap between general understanding and domain expertise, allowing for more informed decision-making. One common approach is to use Retrieval-Augmented Generation (RAG), which combines an LLM with a retrieval mechanism to gather relevant domain-specific knowledge in real time. Integrating a specialized knowledge graph or module, such as a RAG system, allows the cognitive software to become more adept at dealing with niche tasks beyond what a generic LLM can handle.
Multiple Agents: The cognitive framework includes specialized agents responsible for specific capabilities or tasks. These agents can be specialized LLMs, perception units, task executors, or data analyzers. By allowing agents to communicate and collaborate, the cognitive software system becomes more adaptable and capable of handling diverse tasks. The coordination between agents allows for distributed processing, where individual agents contribute to solving different aspects of a problem, thus fostering a more modular and scalable approach to problem-solving.
Chain Manager: A chain orchestrates how different components of the cognitive software interact with one another. It defines the sequence of operations, passing information between the LLM, knowledge representations, and other modules. The chain mechanism is essential for breaking down complex tasks into manageable steps, defining dependencies, and ensuring that tasks are executed in a coherent and meaningful order. It also enables the software to switch context and adapt its behavior based on the ongoing process, promoting structured and goal-oriented reasoning.
Capacity to Accumulate Data and Learn: Accumulation of data and continuous learning is central to the cognitive software framework. This element involves storing interactions, feedback, and outcomes in a memory system, which the cognitive software can access to improve over time. Learning mechanisms such as reinforcement learning, episodic memory, or fine-tuning the LLM can be used to adapt and enhance the software’s capabilities based on real-world experiences. By accumulating data over time, the cognitive software builds a more comprehensive model of its environment, user preferences, and task success strategies, leading to better performance and adaptability.
The Cognitive SaaS framework is rooted in the broader and more complex concept of cognitive architecture, designed to model human cognitive processes such as perception, memory, reasoning, and learning. Cognitive architectures aim to replicate human adaptability in AI systems, enabling them to reason, make decisions, and adapt across various tasks without human intervention. These ambitious systems integrate multiple components to continuously evolve and learn, functioning independently in various differentiations.
While the principles of cognitive architecture inspire Cognitive SaaS, it focuses on more specific, practical applications. Rather than aiming for complete autonomy and broad adaptability, Cognitive SaaS uses advanced AI tools and APIs to perform targeted tasks efficiently. Its primary purpose is to augment human capabilities, acting as a co-pilot that supports decision-making and task execution rather than operating as a fully autonomous system.
Example: A Cognitive CRM
To illustrate how Cognitive SaaS concepts come together in a real-world application, let’s look at an example of a Cognitive CRM. This example shows how each component discussed earlier can work harmoniously to deliver a powerful, user-centric solution beyond simple automation.
I asked ChatGPT to “imagine” a Cognitive CRM based on the framework outlined in this article. Here’s the response:
To create a CRM rooted in cognitive software principles, we must go beyond simply adding AI-powered features. This CRM would act as a co-pilot for users, augmenting their abilities with cognitive capabilities across various aspects of customer relationship management.
Generic LLM for Reasoning and Interaction: At the core of this cognitive CRM, a large language model (LLM) would handle natural language processing, allowing the system to interact conversationally with users. For instance, sales representatives could ask the CRM complex questions about customer history and the best time to follow up or get summaries of past interactions. This LLM would form the reasoning and basic interaction layer of the CRM, offering flexibility and deep comprehension in language-based tasks.
Specialized Knowledge Layer (Context-Aware CRM Data): The Cognitive CRM would integrate a specialized knowledge base, which includes customer data, industry insights, and real-time sales information. This knowledge base would be continuously updated, enabling the CRM to provide contextually relevant advice, such as recommending personalized customer engagement strategies based on past interactions, buying patterns, or industry trends. For example, the CRM could use retrieval-augmented generation (RAG) techniques to pull specific insights from vast knowledge bases, offering more than just generic suggestions when a sales representative asks for guidance.
Multiple Cognitive Agents (Specialized CRM Tasks): This CRM would employ several cognitive agents, each focused on different aspects of customer management:
Lead Scoring Agent: Analyzes leads to determine which will most likely convert based on historical data and behavior patterns.
Sales Forecasting Agent: Predicts future sales performance using past data, customer behavior, and industry trends.
Email/Communication Agent: Helps draft emails and suggests the best time to reach out based on customer sentiment analysis and previous interactions.
These agents would collaborate to enhance decision-making and streamline sales workflows.
Chain Manager (Task Orchestration): The Chain Manager would orchestrate interactions between the LLM, knowledge base, and cognitive agents. For example, when a sales representative asks for the best strategy to approach a customer, the Chain Manager would manage the sequence of operations:
Retrieve the customer’s past interactions.
Pull insights from the lead-scoring agent.
Draft a personalized message using the email/communication agent.
This structured reasoning ensures that complex, multi-step processes happen smoothly and coherently.
Continuous Adaptation (Learning Over Time): A core feature of cognitive software is its ability to accumulate data and learn over time. This CRM would track user interactions and feedback, continuously refining its recommendations and predictions. As it gathers more data on customer behaviors, sales cycles, and representative actions, it will improve its predictions and offer more personalized sales tactics. For example, through reinforcement learning, the CRM can learn from successful sales outcomes and optimize future suggestions. Over time, it would refine its approach to customer engagement, sales forecasting, and retention strategies.
In summary, this CRM wouldn’t just be an AI Wrapper that adds features to an existing platform—it would fundamentally enhance the human decision-making process, elevating sales and customer management at every level.
Conclusions
Creating a Cognitive SaaS is more than just building simple applications that provide minimal additional value compared to tools like ChatGPT Canvas or Anthropic Artifacts.
The Cognitive SaaS framework can act as a strategic checklist to ensure a product goes beyond the typical “AI Wrapper” by offering profound, value-driven innovation.
Does your product integrate a specialized knowledge layer that adds unique domain expertise? Are there multiple cognitive agents collaborating to solve complex problems? Is there a clear chain manager orchestrating the interactions between components to ensure seamless, structured execution? Most critically, have you built a system that accumulates data and learns over time, creating a moat of continuous improvement that competitors can’t easily replicate?
While many SaaS products in the market for years are adding AI and claiming to be AI-first, there is a significant opportunity to build a new generation of genuinely AI-native services. These Cognitive SaaS are designed from the ground up with AI at their core, unburdened by the technical debt and limitations of legacy systems that have been evolving for over a decade.
Claudio Erba: from 0 to 200
Claudio Erba, the founder of Docebo, shares his journey of building a company from the ground up, growing it to $200 million in revenue, and ultimately taking it public on NASDAQ. This free course, available on YouTube, is packed with invaluable insights and strategies, making it a must-watch for any aspiring founder or entrepreneur.
Audiobooks: the power of the perfect narrator
Since I started using Audible, the number of books I’ve read has dramatically increased. However, one thing that can make or break the experience is the narrator’s voice. A voice that’s too high-pitched, regional, flat, or overly dramatic can completely change how enjoyable a book is. Sometimes, I’ve stopped listening simply because I couldn’t connect with the narrator’s style.
I look forward to a future where text-to-speech technology, like ElevenLabs, evolves so that we can enjoy any book with multiple voice options.
The audiobooks I enjoyed in the last couple of months (all in Italian):
Margherita Hack, Nove vite come i gatti. This autobiography, written when Hack was ninety, reflects on the guiding principles of her life: a strong work ethic, persistence, civic and moral commitment, and confidence in herself and her ideas.
Margherita Hack, In piena libertà e consapevolezza. A manifesto on individual freedom and secularism in Italy, addressing topics like assisted fertilization, living wills, abortion, civil unions, free scientific research, and multiculturalism. Hack critiques the influence of the Catholic Church on these debates in Italy.
Isabel Allende, Donne dell’anima mia. With lightness and irony, the author reflects on her past to share the roots of her feminism. Growing up in a patriarchal environment, her instinct to rebel shaped her lifelong commitment to support the women still fighting for emancipation.
Salvatore Giannella, Michele Ferrero. Condividere valori per creare valore. Michele Ferrero, the creator of Nutella and many other beloved treats, was more than one of Italy’s greatest entrepreneurs. He pioneered a way of doing business that prioritized people, guided by the motto: “Work, create, give.”
Gino Strada, Una persona alla volta. This book captures the emotion, pain, struggle, and love of Gino Strada’s extraordinary journey—experiencing conflicts from the victims’ side and becoming a voice for change. Each page resonates with a profound and radical question: the abolition of war and the universal right to health.
If you were forwarded this email or if you come from a social media, you can sign up to receive an article like this every Sunday.
Thanks for taking the time to read this episode of my newsletter. I hope I’ve been helpful. If you think my sketchbook might interest someone else, I’d be glad if you shared it on social media and forwarded it to your friends and colleagues.
Nicola