Banking Rebuilt, Part 2: Show Me The Agent!
In this part, I propose a technical solution, drawing on years of systems integration and cloud architecture. Technical readers will have strong opinions on how to do it better. Good! That's the point.
As illustrated in the image above, the solution is:
- Oriented around AWS Step Functions - this is a defined flow of execution. In other words, "deterministic".
- Serverless - so it is super cost-efficient. You only pay for compute actually used.
- Uses the Investec Programmable API to make the payment.
- Stores credentials and tokens in AWS Secrets Manager - although I quickly put this together, I am a "civilised developer" even when I'm "prototyping".
Given my ease of access to the AWS Cloud, I've successfully prototyped this. However, ...
If I have to critically assess this "quick prototype on a Sunday afternoon during the F1", there are a number of glaring issues that would need urgent attention and remediation if you're even considering deploying as a real "Customer" solution. I've pre-empted your criticism and listed as many as I could think above.
It is also my intention to highlight to the "weekend prototyper" who I hope has made it this far that your -
Check out the "cool AI solution I built" is not anywhere close to reality!
While we're still on our AI magic carpet ride, let's explore a few more technical AI concepts so (i) I can show you a whole new world of AI services, (ii) propose how hybrid deterministic-AI solutions are possible and (iii) gently transition from the way we've been building solutions for decades to AI-powered ones.
Case in point - instead of hand-written extraction rules (e.g. with Amazon Textract) that only work for one supplier's invoice, I use the power of a Large Language Model (LLM) to process/parse/understand an invoice and extract the pieces of information I'm after. This allows us to throw different types of invoices at the LLM and let it use its intelligence to get the right information - from the right line or row or column.
You have undoubtedly heard about Agents and Agentic AI. To be completely transparent, I was only able to process it by conceptualising how it works. One concept that really helped was ...
The Agent Loop. Essentially what this means is that we give the Agent (i) a problem statement - in the form of context and instructions, (ii) tools that it can use to achieve the outcome and most importantly (iii) a brain, in the form of an LLM, that allows it figure out how to solve the problem.
As highlighted above, this problem becomes relatively trivial with an Agentic approach. Using a framework like Strands, I created an Agent, built a bunch of tools (in its simplest form as python functions), and an instruction in the form of a prompt.
At this juncture, I would like to highlight that the most important ingredient of this solution is the well-worded prompt or instruction.
Flashback to your Grade 11 English teacher who warned you that one day you will regret not putting more effort into your creative writing exercise to improve your writing!
In all honesty, my prototype started on my local machine - as all my projects do. I then deployed, iteratively, to Amazon Bedrock AgentCore to understand how the services worked. This helped me contextualise why AgentCore is the platform to build, connect, and optimise agents. And how it helps engineers ship agents fast, connect them to enterprise systems, and optimize them continuously. For the record, that is the official AgentCore messaging that I've tattooed on my arm in size 8 font!
You wanted to see the Agent? There it is - started on my local machine, now reading, deciding and paying on enterprise-grade infrastructure. Not a high school project anymore. Great for Architects and Agentic AI Engineers. But be honest: would your doctor build this? Your accountant? Your mom? Probably not, and they shouldn't have to.
In the final part, I tackle the real question: how would someone without 25 years of system integration and cloud experience use AI to make payments? I hope you will join me.
Note: I work at Amazon, but this is my own opinion.
