AI this, AI that, AI in my fridge, AI in my kids toy, everything is AI nowadays. Don’t get me wrong I love the fact that at least it has some real world impact and applications unlike blockchain which proved to be utterly useless. Jokes aside, I think AI has it’s time and place and nowadays I see it being used just for the sake of “being used”.
You see the tweets: agents coding overnight, productivity 10x, replace juniors tomorrow. Then reality hits when your project turns into a prompt debugging session that costs more than hiring a real developer. I've noticed this recently since a couple of potential new clients came to me after burning 3-6 months on "AI magic" that hallucinated features nobody asked for. Now they will probably pay me to do it all over again, but with a more balanced approach to building.
This isn't me hating AI. I run agents in the background while I doomscroll or play with my kid, which solves two of my problems - hanging out with the family more without being unproductive. But forcing AI where it doesn't belong is the fastest way to waste runway and look stupid to investors and to yourself.
Here are the rules I actually use before touching any AI tool on a project. Simple, no tech-bro lies. If it hits close to home, DM me and we can probably audit your mess in 15 min.
Rule 1: If it's your secret sauce, keep AI the out.
Your core logic is the thing that makes your startup actually different, at least I hope so. Pricing engine, matching algorithm, whatever proprietary magic or secret sauce you invented. AI is still basically a pattern-matcher trained on the internet. It will approximate your edge, maybe even "improve" it with some hallucinated shortcut, then leak the prompt logic or just dilute what made you special.
I've seen founders lose their moat because "Claude can do this better." No. Humans write the IP. AI executes the boring parts around it.
Red flag: If explaining your secret business logic takes a whiteboard and 30 minutes of back-and-forth, don't prompt it.
Rule 2: High-stakes high-pressure human work like negotiations, ethics, creative calls are yours to execute.
AI has zero empathy, zero skin in the game. Asking it to read client tone, handle objections, or decide "is this feature creepy?" is like asking your toaster for relationship advice.
Example: One client tried AI sentiment analysis on support tickets. It flagged "love this app!!" as sarcastic because patterns said exclamation marks = anger. Killed trust overnight. We switched back to a junior reading them and fixed this small nuance in a day.
If it involves trust, nuance, or not pissing off humans, do it yourself or with your team. AI accelerates bad decisions here.
Rule 3: Early stage when nothing is defined is not for the AI since it loves point blank instructions.
MVPs, pivots, fuzzy ideas? AI needs clear inputs to not suck. Give it vague briefs and it pattern-matches to the average software everyone else built.
You end up with a generic app that "feels right" until users hate it. Then you spend weeks fixing what a 1-hour team rant would've clarified.
My rule: Nail the "what problem are we solving" and rough flow on paper/whiteboard first. Only then let agents touch code or prototypes. Otherwise you're just vibe-coding with extra steps and a bigger bill. Please read up on specification driven development and do the specification properly. AI can work with you on it, but you are the driver here.
Rule 4: Anything touching real security or compliance.
Sensitive data, payments, regs? One hallucinated "just add this endpoint" and you're in a pickle consisting of lawsuits, audits, game over.
AI is great at summarizing policies or generating boilerplate docs. But writing actual safeguards? Nah. Humans review, audit, sign off.
I've had clients show me AI-generated auth flows with hardcoded test keys still in. Cringe. Skip AI here unless it's pure read-only helper stuff.
Rule 5: You can't measure if it actually helped.
If you add AI and can't track "time saved," "bugs down," or "cost per feature," it's an experiment, not a tool.
Too many founders go "feels faster" while spending double fixing drift and hallucinations.
My hack: Before/after metrics or don’t do it. No measurement of efficiency = no AI. Simple.
Bonus rant rule: If overseeing the AI takes more brainpower than doing it yourself, abort immediately.
Your ancient caveman brain wasn't built for endless "why did the agent do that?" loops. I've burned nights on that realization. Not worth it when the win is marginal.
Quick filter I run every time:
Is this repetitive boring work I hate? → Yes? Let AI grind it out.
Is this fuzzy, unique, high-trust, or my edge? → Humans only.
Can I measure the ROI clear as day? → Proceed. Otherwise skip.
AI is awesome at failing your bad ideas 10x faster (free lesson). Use it to accelerate the boring 80%. Skip it on the 20% that actually matters like your passion, your moat, your sanity.
If your project's already deep in "expensive experiment" territory, hit reply or DM me. I've unwrapped enough AI messes to spot the patterns quick. Happy to do a free 15-min sanity check. No sales pitch, just real talk.
What's the dumbest place you've forced AI so far? Drop it below and I'll roast or relate.