From Prompts to Processes: My AI Automation Journey
Through conversations and interactions at conferences this year, I’ve noticed that most people in finance and actuarial work have the AI awareness piece—they know the tools exist and are actively using ChatGPT or CoPilot daily for quick tasks, such as drafting emails or refining text. What I keep hearing, though, is a desire to go further, paired with uncertainty about how to start moving beyond simple prompting into building more powerful automations. I felt the same way, and I wanted to upskill in this area—so I thought, why not document the process and share my journey?
I’m going to share my own journey of self-teaching, experimenting, and building test cases that mix traditional automation with AI-powered workflows. The aim is to give a practical view: not just the steps, but also the real-time commitment, the tools I use, the costs, and what I’d do differently if I had to start again. Hopefully, this helps others determine if this is something they can work on in their free time and on weekends, or if it requires a deeper time investment.
The Challenge
Between now and Christmas, I’ll be working on and sharing at least three AI automation projects. I've picked personal productivity use cases for tasks with outcomes that I value however consume much more time than I want to spend, or end up getting deferred or skipped altogether!
-
A multi-agent system with a parent AI agent that delegates and coordinates tasks across smaller specialised AI agents. Think of it like an AI project manager for day-to-day work.
I'll keep this one simple, starting with just agent tools for contact management, calendar management, emailing, and basic web research tasks. However, as I develop more workflows or 'tools', the intention is to connect them to expand this AI assistant’s capabilities exponentially. Some examples might include - tax invoice scanning and filing, reading and updating the family shopping list, or even finding and making dinner reservations.
-
A way to combine AI with traditional automation to triage emails, apply rules written in plain English, and give more guardrails and control over what gets filed, flagged, answered, or passed off to more complex or sensitive workflows.
While it would be quickest to create an autonomous agentic AI workflow there is a risk that it may share or delete important emails, whether unintentionally or through more malicious channels (i.e. email prompt injection attacks)!
The challenge will be finding a design that balances deterministic code vs ai to achieve a high level of security and accuracy, while still being able to handle new complex context/‘judgement’ based action without the need for high effort human intervention (i.e. re-coding). This one is going to be more challenging!
-
A tool that identifies and logs nutritional information from photos, voice, or text inputs; records it in a database; and tracks progress against goals.
The introduction of fitness tracking wearables and IoT devices has streamlined and automated the measurement of energy output, weight, body composition - however diet as the most important part of the equation has remained very manual to track. I thought to myself several years ago, "Wouldn't it be great if I could just take a picture of my food and it's logged in my food diary with all the nutritional metrics I care about?"
Well its finally time to try build this as a fun little side project! My minimum objectives for this process is to have multi modal inputs processing, with automated database logging and reporting. The goal is not to be 100% accurate, but rather to minimise effort (so I don't skip entries) with ballpark accuracy. The tool needs to be transparent and interactive – providing me with descriptive feedback that helps me evaluation the accuracy of the estimate, with an interface that remains easy to adjust where I think changes are necessary. The stretch goal is to build and connect a front end app quickly by leveraging AI coding assistance tools, and integrating some data and security protocols.
Each of these will get its own post later in the series. I’ll cover how I designed it, the build process, what worked (and what didn’t), plus the lessons learned along the way. I’ll also share any available learning resources I found useful—tutorials, platforms, open-source tools, and communities—as well as the actual time and cost involved.
Why I’m Sharing This
This isn’t about presenting polished products or a “perfect” way of doing things. It’s about showing the real learning curve, trade-offs, and stumbling blocks that come with trying to build practical AI automations while balancing a busy professional life. My hope is that by being open about the process, I can help others in the actuarial and finance community see what’s possible and where to start.
I’m also doing it to keep my continuous learning commitments on track – setting challenges with deadlines and making them public is a great self-motivation tool to keep going when life gets busy, or when I reach Dunning-Krugers valley of despair!
This is just the introduction to the series. Next up, I’ll dive into the first use case: building an agent-of-agents personal assistant, and what I learn from putting it into practice. As the first in series it will focus heavily on the educational resources I used to get started, as well as software choice.
Do you have a process you have wondered or wanted to use AI to help automate or improve but not sure how to get started? Let me know below and (time permitting) i'll add a 4th use case of the one with the most interest.
