What happens when AI gets it wrong (and when it works brilliantly)
I "hired" an AI to do my freelance work. Did it work out?
Hi!
Last week I tried using Manus, an AI task assistant, to help with a client project — CMS migration from an old version to a newer one.
I fed it as much info about the project that I can, and asked it to generate all the migration code. And it did! It even did one step better — it provided code to run its own validations to make sure everything was correct and complete.
It’s wild seeing an AI assistant not just write code, but test it too. This kind of automation lets me move faster on freelance projects and spend more time on strategy and high-value tasks. If you’re hiring a dev, this is the kind of leverage you want on your side.
On paper, it looked ideal: Manus could write all the code and even validate its own output. A perfect developer assistant, right?
Not quite.
In practice, it didn’t work. Without full access to my project’s architecture, Manus had to make guesses. Those guesses were wrong. I spent hours debugging before giving up and doing it manually.
This highlights a key challenge when using AI for existing client systems — especially as freelancers, consultants, or agency teams. Real software is messy. Architecture is spread across codebases, dev servers, scripts, and undocumented workflows. If AI doesn’t have the full picture, it tries to “fill in the blanks,” and that’s where it breaks.
So When Does AI Work for Client-Facing Work?
If you’re like me — juggling multiple clients, building internal tools, or improving your own service delivery — here’s the good news:
AI works incredibly well when:
✅ You're creating new, self-contained features
✅ You’re building tools for internal use
✅ You’re working in clean, well-scoped environments
Here are a few AI wins from my other AI-build efforts:
🟢 Website Uptime Monitor – A tool I built for myself to keep tabs on client sites, and my own sites. If anything goes offline for more than a minute, I get alerted instantly. Simple, helpful, and 100% AI-built.
→ Try websitemonitor.live - it’s free!📨 Email Template Builder for Pop Invoice – A custom feature I added for users of my SaaS, Pop Invoice (which automates invoicing using Notion). AI helped generate a flexible system for email subject and body customisation.
→ Try Pop Invoice for free.🪑 Posture Monitor (screenshot below) – Not client-related, but very real: I built a webcam-based tool that tells me when I’m slouching. Years of consulting in front of a laptop = neck pain. AI to the rescue. I’ve not released this one, but would love some
guinea pigstrial users. Let me know if you want to try it out!
What’s certain is this: my 20 years of experience as a software engineer + AI is a winning combo. These tools now allow me to reduce project delivery time by up to 10x — without sacrificing quality.
I’d love to help other companies unlock the same advantage, whether that’s for internal tools, client-facing products, or streamlining your own delivery process.
If you’re curious what AI (+ engineer) could do for your next project, just hit reply to this email and write:
👉 "Let’s chat!"
– Farez