Disclaimer
I am NOT a developer, frontend, backend, full stack, or short stack. I do not pretend to be and will be the first to tell you I don't really understand how some of the code works when it comes to websites and some other things. Do I want to learn - ehhh. Do I have time to dive into modals, html code, ruby, etc - nope.
My crutch and my weakness
Enter AI - Claude, ChatGPT, Codex, Gemini.
My Crutch - because I can have it do the things I want it to do without having to learn the full process. Need an internal website to give to Product Owners/Managers for a couple hundred thousand customer profiles that include their relational access to other Products, their account profile, and their registered MFAs? AI's got me. Here's the data, make something fancy and flashy to make the non tech people ohhh and ahhhh.
My Weakness - because I don't fully understand software dev languages. Do I understand what I am doing with AI tells me to run deno task dev? At a high level, yes. Could I run you through the steps behind it? Probably after a bit of looking. However, the question is do I actually understand everything its doing and, more importantly, the WHY it's doing it? Absolutely not.
It's this weakness part that you have to watch out for. Could you troublshoot and fix a production issue with what you put out or are you relying FULLY on AI to handle things for you?
What have I done with AI?
To build something new
I've built a multi org switcher for PowerShell to be used in conjunction with the Okta.PowerShell module from Okta found here: okta-multiorg. Could I have written this out myself? Definitely. However, the question is how LONG would it have taken me? Likely a couple of weeks to tweak everything into how I wanted it to be. However, I spent 15 minutes crafting a statement to Claude, having it ask me questions about how I wanted to use it and some other back and forth. I brought this into work to test for real, found a bug around the link popping up on the default browser that was causing issues with the account it was attempting to link, and then asked OpenAI to fix the issue and allow me to copy the URL instead of automatically opening it.
I rewrote our entire team's module, approximately 6,000 lines of code from using a homegrown PowerShell module to using Okta.PowerShell from Okta. Could I have done the conversion myself? Yup. Would it have taken me longer to write, test and iterate my code? Yup. Is AI the perfect companion for going from an old thing to a new thing? It most certainly is.
To troubleshoot HAR files
Do you need to figure out quickly where your SSO is failing in Production? This is the PERFECT use case for using AI. Give it the HAR file, explain the site urls that you know are in play in the flow, and explain what each is to AI and tell you to break down where its failing.
Hey, I need you to analyze this HAR file for me, its generating an xyz error.
The urls in play are as follows
www.abc.com -> the url of the application
xyz.okta.com -> the url of the okta instance that the application is using for SSO
efg.com -> this is the customer IDP that is integrated into xyz.okta.com
Analyize this HAR file, and tell me why its generating xyz error. Additionally ask me additional questions to better answer this request.
This is a good way to parse the file, have it ask you additional questions and get to the root of the issue.
To understand, quickly, code you inherited
Inherited some ancient code that is running somewhere and is a couple hundred lines long with hardly any comments or a synopsis? AI is perfect to review the code, add comments, add a synopsis. You could just have it break it down to you in the response and you do nothing to update the code for the next guy - but we are not monsters here.
AI is on LSD
Hallucinations are real. As you go down the rabbit hole of AI usage, you can and will find that things get weird. So weird that you might think that the AI agent is on something REALLY REALLY good, as that could be the only reason it gave the answer it did. It's your duty to tell AI that its off track or that 123 is not configured that way and is not an issue. I've seen some interesting ones.
For instance, I asked ChatGPT if it could build me as code, an Okta Workflow flopack. It said it definitely could, and several iterations later - Okta would still not recognize it as a valid file. I've asked it to research a topic and after asking some additional questions around it, it ended up going down a completely separate rabbit hole. Is some of this response issues from myself? Valid possiblity, and quite likely the reason. However, you HAVE to realize when its off course and either get it to correct itself, or correct the code yourself. I will not accept any code from my team that was written by AI that they cannot walk me through functionality, explaining the logical transformations and output. You HAVE to know your code. How many companies have had issues with databases and other thigns being either permanently or temporarily deleted because of AI or code written by AI that was not vetted properly.
TL;DR
AI is a tool in your toolbelt. It is NOT your replacement and it is NOT infallible. You need to implement proper guardrails both on yourself and on the processes in which you adopt the output from AI.