This Week, I Used AI on Real Problems
Over the past week, my take on working with AI got a little more down-to-earth.
It didn't help me "build a product in one click." What it did was stick with me through the kind of problems you actually run into with real systems: a remote support workflow, the POS self-ordering experience, production bugs, recovering my personal site, and a longer-term code review plan.
None of this looks like the polished AI demos you see in videos. For a product that's already live, though, it matters a lot more.
Getting the remote support workflow working first
I spent part of the week testing a remote support setup. The goal was simple: the next time a store has a device issue, I don't want to talk someone through every single step. I want installing and connecting the remote channel to be as predictable as possible.
It didn't go smoothly. On the test machine I saw installs hang, scripts that didn't continue the way they were supposed to, and the program quitting right after the permission prompt was accepted.
AI didn't "magically fix everything" here. It was useful because it helped me break the problem down:
- Which steps depend on system permissions
- Which steps have to wait for a service to finish installing
- Which steps belong in the script
- Which steps need clear feedback for whoever is running it
I'm more and more convinced that AI is great for this kind of step-by-step troubleshooting. It can't see what's actually happening on the real machine, but it helps me miss fewer things.
Taking self-ordering from "works" to "works well"
This week I also focused on the UX problems in the scan-to-order self-ordering page.
For example, long item names got cut off. Tapping an item didn't give customers clear enough feedback. After picking something, it wasn't obvious how many you had or what was in the cart. And once you were in the cart, you still needed to be able to change quantities, check out, or cancel the order.
None of these are complicated features, but they decide whether a customer feels confident enough to keep ordering.
I had AI start by turning the requirements into user flows. Then we compared them against how staff take orders on the POS and listed the states that were missing: an item detail popup, a bigger item image, selected quantity, a cart breakdown, quantity +/-, checkout, and order cancellation.
This kind of work is the easiest to underestimate because it isn't "shipping a big new feature." But it's getting the product's fundamentals right.
Small bugs in production aren't small
A few bugs this week are worth paying attention to.
- One: the app launched on the secondary display. At a checkout counter, that directly gets in the way of using the system when it starts up.
- Another: with GST and service charge set to 0 in the admin panel, those charges still showed up at checkout.
- And after an item was taken off the menu, customers could sometimes still select it through scan-to-order.
All of these were reminders that edge cases in production systems have to be properly verified. Charges, whether an item is available, and whether the cashier side and the customer side agree are things you can't trust just by looking at one page.
Here AI acted more like a pair reviewer. It helped me keep asking questions: Which parts of the app does this setting touch? Do self-ordering and staff ordering behave the same way? After the fix, how do we verify it doesn't break normal orders?
Starting a longer-term code hardening routine
The system is already running POS checkout, which is a high-pressure setting, so I don't want security and stability to depend on one-off patches.
So this week I started working with AI on a phased code review and hardening plan. The ground rules:
- Keep every change as small as possible
- Start with the high-risk areas
- Test thoroughly
- Don't disrupt customers who are using it right now
- Gradually hold future development to the same standards
This one really matters to me. Building with AI agents definitely speeds things up, but once you have the speed, you have to add engineering discipline.
Wrapping up
My biggest takeaway this week: AI isn't there to make decisions for me. It's a working partner that helps me break down problems, fill in checklists, and look back at what happened.
Its best use turns out not to be "help me come up with a cool new idea." It's taking real, specific problems that are easy to miss and turning them, step by step, into changes you can act on, verify, and ship.
If you're using AI on real projects too, leave a comment. What's the most practical problem you've had AI help you with lately?