How I Merged My Thoughts Into an AI Agent
Jul 26, 2025
By Carter Springall, CTO & Co-founder @ Artifact AI
In just one week, I figured out how to transfer my accounting expertise into our AI agent—essentially merging my thought process with its decision-making framework. The result? A bookkeeping system that not only automates complex accounting tasks but does so with the same reasoning and nuance I use in real life.
The Breakthrough
The key to this leap was a combination of tools that let me “upload” how I think into the AI. At the center of it all is a universal SDK layer I developed, which plugs directly into our technology stack.
This SDK isn’t just an integration—it’s a foundation for letting our AI agent approach financial decisions the way I do.
The SDK layer can be imported into any of our repositories, making this solution incredibly versatile and easy to deploy across our entire technology stack.
Why Cursor + Codex Made It Possible
The secret sauce was pairing two powerful systems:
Cursor: A tool for indexing entire codebases and understanding context across multiple repositories. I used custom Cursor Rules to train it to follow my accounting decision-making pathways.
Codex: Combined with specialized files and a lot of brainstorming, I documented my accounting expertise in a way the AI could internalize. This forms the basis for future model fine-tuning.
AI Memory Changed the Game
A key insight was to use Cursor’s memory feature:
Whilst implementing the Grok‑4 integration I found the agent would constantly suggest for me to use an incorrect model name or tell me that there was no such thing as grok 4. I saved to memory that “Grok‑4 is now a model” and, to my surprise, it worked perfectly!
I also shared web resources on integrating agents with Model Context Protcol Servers (MCPs) from providers like Bedrock, XAI, Google, Anthropic (Claude), and OpenAI. This gave our AI agent a constantly updated knowledge base to work from.
How I Transferred My Thought Process
Here’s what I did step by step:
Built a knowledge base of accounting principles and reconciliation procedures
Designed decision trees to mirror my classification logic
Added a scratchpad feature so the AI “shows its work” like I do
Implemented feedback loops so it can learn from mistakes and continuously improve
Results That Speak for Themselves
Our AI agent now:
Automates full-cycle bookkeeping with minimal oversight
Matches the accuracy of a new human bookkeeper
Cuts processing time from hours to minutes
Delivers consistent results across all transaction types
What’s Next
This is just the start. I’m now exploring how this thought-merging technique can handle more complex financial workflows. The universal SDK is ready to support far more than bookkeeping—it’s the foundation for a new way of combining human expertise and AI reasoning.
By merging human thoughts with our AI agent, We've moved past rule-based automation into something truly powerful: an AI that understands why, not just how.
By Carter Springall, CTO & Co-founder @ Artifact AI