I watched a guy named Mark in the operations department build a working web application yesterday morning. Mark is the kind of guy who clic...
I watched a guy named Mark in the operations department build a working web application yesterday morning. Mark is the kind of guy who clicks Reply All to company announcements. He absolutely does not know how databases work.
Yet somehow he was dragging text boxes around his screen, and within ten minutes we had a live URL showing inventory levels pulling directly from the company Snowflake account. I asked him how he handled the authentication tokens. He just stared at me. He had literally just typed "grab the inventory numbers" into a text box.
OpenAI pushed this Codex update out last week and it completely changed the dynamic of how things get built around here. They finally stopped pretending this tool is only for software engineers. For the longest time, if you could not write Python or JavaScript, Codex was totally useless to you. Now, they added these things called role plugins.
This means the AI natively hooks into software like Salesforce, Canva, and Figma. You pick your job title from a list, and the system automatically bundles the tools you use every single day. Mark picked the Data Analytics role. He did not write a single line of database query language. He just told the computer what he wanted to see, and the AI reached into our systems to fetch it.
Trading Code for Webpages
The bizarre part is what happens after you hit enter. You do not get a snippet of code back.
In the past, even if the AI gave you perfect code, you still had to figure out how to run it. You needed a server. You needed a domain. You needed to understand deployment.
Now, you get a fully functional webpage. They call the feature Sites. Mark typed his messy prompt, and OpenAI servers silently spun up a hosting environment, rendered the frontend interface, and spat out a private link. You just copy the URL and drop it into an email. Anyone in the company can click it and use the app.
For years, trying to build something complex with AI meant getting stuck in a nightmare loop. You would get an app that mostly worked, but maybe the padding on a sidebar was slightly off. You would type "fix the sidebar." The AI would panic. It would completely forget what framework it was using, and rewrite your entire backend logic for no reason. Your app was ruined, and you had to start completely from scratch.
They finally fixed that. Now you just take your mouse, highlight the broken sidebar, and type "make this wider." It only patches that exact spot. They named it Annotations. I genuinely hate to admit it, but it works flawlessly. You can click on a chart, ask the AI to change the colors from blue to green, and the rest of the page stays perfectly intact.
The Trap Door Nobody Sees
Nobody is really thinking about the massive trap door sitting underneath all this.
You do not actually own the software you are building. Try to export the code. Try to take that shiny new dashboard and migrate it over to your own company servers so you have a local backup. You cannot do it. It is completely landlocked inside your ChatGPT workspace. It only works for people logged into that exact enterprise environment.
And it will inevitably break. An API will change its data structure, or the AI will invent some weird piece of routing logic that collapses under load. Mark will not be able to fix it. Our IT department is eventually going to receive a frantic email about a critical broken app they did not even know existed. It will be written in machine generated spaghetti code and hosted on a server they have zero access to.
Security and compliance teams are going to lose sleep over this. IT departments spend years trying to kill shadow apps that employees use to get around slow company processes. Now, OpenAI has essentially given every single employee a shadow app factory. Anyone can spin up custom software connected to live company data just by asking nicely.
I tried explaining this to Mark. I told him hooking the live customer database into an unverified script was probably a terrible idea. I pointed out that if his new dashboard pulls the wrong pricing data during a client call, he is the one on the hook for the mistake. He just shrugged, copied the URL, and emailed it to the regional director anyway.
The director loved it. He even asked Mark to build another one for the sales team by Friday. So I guess we just wait for the crash.
References:
OpenAI officially detailed this shift in their recent post on Codex for Every Role, Tool, and Workflow.
