I took the briefing “AI Leader: Generative AI & Agentic AI for Leaders & Founders” by Donner.
I initially expected to learn more about how to build an AI startup, including how to sell, market, and advertise an AI product. Instead, the briefing focused more on the business perspective of AI: how to evaluate AI opportunities and make better decisions about AI initiatives.
Overall, it helped me structure and connect my existing knowledge of Machine Learning, Artificial Intelligence, Deep Learning, and Large Language Models (LLMs).
AI risks
One of the important topics was the different types of risks associated with AI initiatives:
- Technical risks — problems with data, models, infrastructure, accuracy, reliability, etc.
- Operational risks — problems integrating AI into existing processes and making it work reliably in production.
- Strategic risks — investing in AI initiatives that do not create meaningful business value or do not align with the company's strategy.
We also discussed common traps when starting an AI initiative:
- Solution looking for a problem — building something with AI simply because AI is available, without having a real business problem to solve.
- Not enough data — the available data may be insufficient in quantity or quality.
- No clear success metric — not knowing how to measure whether the initiative is actually successful.
- Model drift — model performance can degrade over time as real-world data and conditions change.
RAG vs. Fine-tuning
I learned more about the difference between Retrieval-Augmented Generation (RAG) and fine-tuning.
RAG
RAG is a way of providing an LLM with relevant external context at inference time instead of changing the model itself.
Typically, information is stored in a retrieval system such as a vector database. Relevant information is retrieved and added to the prompt before the LLM generates an answer.
Advantages of RAG:
- No model retraining is required.
- Knowledge can be updated relatively easily.
- It can provide the model with up-to-date or private information.
- It is often easier to control and maintain than fine-tuning for changing knowledge.
Fine-tuning
Fine-tuning means further training an existing LLM on a domain- or task-specific dataset.
Compared with RAG, fine-tuning:
- requires additional training;
- can be more expensive and time-consuming;
- can make the model better at specific tasks, formats, or behaviors;
- may increase inference efficiency for certain use cases because relevant behavior is learned by the model itself;
- does not work as a simple replacement for RAG when the main requirement is access to frequently changing knowledge.
The key difference I took away is:
RAG changes the context given to the model; fine-tuning changes the model itself.
Agents and workflows
Another important topic was the difference between AI agents and workflows.
An AI agent is a system where an LLM has some control over how a task is completed. It can decide which actions or tools to use and determine the next steps based on the current situation.
Some characteristics of agents:
- the LLM controls part of the workflow;
- they can make multiple LLM calls;
- they can use external tools;
- they have a certain level of autonomy;
- they can decide what to do next instead of following only a fixed sequence.
A workflow, on the other hand, generally follows a predefined path implemented by code.
Examples include:
- Prompt chaining — passing the output of one LLM call to another.
- Routing — selecting different models or processing paths depending on the task.
- Other predefined sequences of AI and non-AI operations.
The main distinction is who controls the process:
In a workflow, the developer defines the path. In an agent, the LLM has some control over the path.
Guardrails
I also learned about guardrails.
Guardrails are constraints and controls designed to make AI systems and agents behave safely and reliably within defined boundaries.
They can help control:
- what an AI system is allowed to do;
- what tools it can access;
- what information it can use or expose;
- what types of outputs it can generate;
- when human approval is required.
This becomes especially important as agents become more autonomous.
Three types of AI
The briefing also provided a useful way to categorize AI systems:
1. Predictive AI
Used to predict outcomes based on existing data.
Examples:
- predicting demand;
- predicting customer churn;
- predicting stock prices.
2. Generative AI
Used to generate new content.
Examples:
- text;
- images;
- audio;
- code;
- video.
3. Agentic AI
Used to build systems that can autonomously perform tasks and execute processes by making decisions and using tools.
Examples:
- automating business processes;
- researching information;
- interacting with other systems;
- completing multi-step tasks.
AI initiative prioritization
One of the most useful frameworks for me was thinking about why a company should invest in a particular AI initiative.
An AI initiative should address a real pain point and create measurable business value.
Three major value drivers are:
Automation → Efficiency → Cost savings
AI can automate repetitive work and make existing processes more efficient, potentially reducing costs.
Augmentation → Doing more → Revenue and/or cost savings
AI can help employees do more rather than completely replacing them. For example, one employee may be able to handle more customers or complete tasks faster.
Differentiation → Innovation → Revenue
AI can also create new products, services, or customer experiences that differentiate a company from its competitors and create new revenue opportunities.
AI roadmap
Another useful framework was an AI initiative roadmap, which I think of as a continuous cycle rather than a one-time process:
- Plan — define the initiative, timeline, resources, and expected outcomes.
- Research — conduct R&D, evaluate approaches, and curate the necessary data.
- Build — develop the model, AI system, and product.
- Deploy — put the solution into production and drive adoption.
- Measure — track KPIs, monitor performance, learn from the results, and improve the system.
The important part is that the process does not really end after deployment. Measurement and feedback should lead to further improvements and potentially another planning cycle.
Important lesson: communication
One of the important points I took away is the importance of clear communication.
An AI initiative is not only a technical project. Engineers, product managers, business leaders, and other stakeholders need to understand what problem we are solving, why AI is appropriate, what the expected value is, and how success will be measured.
What I liked about the briefing
What I particularly liked about this briefing was that it was not only theoretical. It introduced practical tools, such as Google Sheets templates, for organizing and evaluating AI initiatives.
Instead of just learning concepts, I could see how these concepts could be turned into an actual decision-making process.
Final thoughts
I expected this briefing to teach me more about building and selling an AI startup, but I ended up learning something different and, in my opinion, more useful for my current career transition: how to think about AI from a business perspective.
I think this knowledge will be useful when I work as an AI Engineer, especially when I need to understand business problems, propose AI initiatives, and communicate their potential value to non-technical stakeholders.
I can also use these ideas when I eventually create my own AI startup. Instead of starting with “What can I build with AI?”, I should start with:
“What real problem can I solve, and why is AI the right solution?”