We recently hosted our fifth educational Fundamentals workshop for junior VCs, this time going back to basics to break down the AI tools everyone is talking about.‍

What is Fundamentals?

Fundamentals is a series of educational sessions by junior VCs, for junior VCs, ran by myself and Freya Wordsworth (Associate at Ascension). We wanted to build a safe space for others early in their VC careers to network and learn from experienced investors on topics where there might not be formal training. In turn, we hope this better equips us all to make the investment process as smooth as possible for founders.

Check out our previous sessions on Term Sheets, Cap Tables and Board Observing.

Back to Basics on AI

For this session we were joined by Rakesh Murria, COO at Ascension, who helped us unpack some of the language, tools and concepts that can make AI feel harder to access than it really is. The goal was to strip away the jargon and understand what people actually mean when they say API, RAG, MCP...

Once broken down, many of these concepts are just new ways of describing software capabilities that have existed for a while, now made much more accessible through advances in AI.

Key Terms

A few simple definitions covered in the session:

• Token: the basic unit models use to process text
• API: a way for one software tool to access another tool’s capabilities
• Agent: a system that has enough context and decision-making ability to take actions
• RAG (Retrieval Augmented Generation): giving a model access to relevant documents or information so it can answer based on that context
• MCP (Model Context Protocol): essentially an API with the instructions attached, so the model can understand how to use it
• Multimodal: models that can work across text, images and video, rather than just text
• Frontier model: the latest and most capable models, usually also the most expensive
• Vibe coding / no-code: building tools and products without needing to write traditional software from scratch

A useful way to think about AI agents

A simple framework for thinking about an AI agent as having four layers:

• The brain: the model itself and its reasoning ability
• The glue: the orchestration layer that decides what happens next
• The hands: the connectors that allow it to interact with tools, files and data sources
• The surface: where the user actually interacts with it, whether that’s Slack, a dashboard, a mobile app or something else

This framing is useful as it shows that the value of an agent is beyond the model itself - it often comes down to how well the intelligence is connected to context and integrated into workflows.

The Model Landscape

Rakesh took us through the current landscape and emphasised that different models are good at different things. Rather than asking which single model is best, the real opportunity is in building systems that can choose the right model for the task at hand.

The session also highlighted how quickly the surrounding AI tooling ecosystem is evolving, with new coding environments, no-code builders and agent platforms emerging at a rapid pace.

Example: the Ascension Knowledge Base

A major theme of the session was that model capability alone is only part of the story. What really unlocks value is giving models access to the right context.

At Ascension, the team has built an internal knowledge base that brings together a vast range of company information into one queryable layer.

Use of the knowledge base means that instead of manually feeding a model the one spreadsheet or file that matters in the moment, the system can automatically pull in surrounding context from across the business.

Practical Workflow Advice

For those earlier in their journey, Rakesh encouraged starting with accessible tools and gradually building more advanced workflows over time.

Some practical tips:

• Begin with user-friendly AI tools before moving into more advanced code or terminal environments which enable greater autonomy
• Understand the difference between subscription plans and raw API usage, as each can be more cost-effective depending on how you work
• Be mindful of how large files and datasets can quickly impact usage costs
• Create shared processes for managing reusable prompts, workflows and “skills” across a team
• Even in a non-technical team, GitHub can be an essential part of the workflow for managing version control on shared AI-built tools.

Final thoughts

The session reinforced that getting value from AI is less about chasing the latest model and more about understanding the fundamentals, experimenting with different approaches and building workflows that suit your team’s needs.