WhenNOTtouseAI:7honest"don'tbuildthis"casesfromKurdistan

TL;DR: I have walked seven Kurdistan businesses away from AI projects in the last twelve months. Each one would have been a respectable invoice, and each one would have hurt the business that bought it. Here are the seven cases — anonymized — and the cheaper, more honest answer in each one.

Why this list exists

AI is the loudest part of my market right now. Owners read about it on LinkedIn, see a competitor doing something flashy, and decide they need it before they have decided what they need it for. Saying "no, not yet, here is what to do instead" is the single most useful thing I can sell, even though I cannot invoice for it. So this post is the public version of that conversation.

Case 1: 8-table cafe in Duhok wanting an AI menu chatbot

A small cafe with eight tables, two staff, and seventy regular customers wanted a WhatsApp bot to take orders. The owner had seen a viral post about a Lebanese restaurant doing this. We added up the math: a handful of orders a day, all from people the staff already know by name. The cost of building the bot was four months of revenue from those WhatsApp orders. The right answer was a printed menu with a QR code, and a saved WhatsApp Quick Reply for "today's specials." Total cost: $0. Done.

Case 2: small clothing brand wanting AI-generated product descriptions

A boutique with 80 SKUs wanted AI to write product descriptions in three languages. We tried a sample run. The descriptions were technically correct and totally interchangeable — every item read like every other item, because the AI had no actual information to differentiate them. Customers clicked away. The right answer was for the owner herself to record a 30-second voice note about each new piece (origin, fabric, why she chose it), and let me transcribe and lightly edit those into the three languages. The voice notes carried character; the AI descriptions carried nothing. Total: about $200 of transcription work for the season.

Case 3: pharmacy chain wanting AI medical advice

A four-branch pharmacy wanted a WhatsApp assistant that suggested over-the-counter medication based on customer symptoms. The legal exposure alone made this a no. AI hallucinations in pharma are not a chatbot bug, they are a public-health incident waiting to be reported. The right answer was an AI assistant that handled refills, opening hours, availability, and routing genuine symptom questions to the on-shift pharmacist by phone. Useful, safe, profitable. Built in two weeks.

Case 4: real-estate agency wanting AI lead-scoring

A 12-agent agency wanted an AI to score incoming inquiries by likelihood-to-buy. They had thirty leads a week. With volumes that small, no AI scoring is going to beat a senior agent eyeballing the list for ten minutes on Monday morning. AI lead scoring needs thousands of historical examples to train against — at thirty a week it would take a year of data before the model knew anything. The right answer was a clean inquiry form and a Monday-morning standup.

Case 5: school wanting AI teaching assistant

A private school in Erbil wanted an AI tutor that students could chat with for homework help. The hard problem here is not the chatbot — it is that the school's own teaching materials, lesson plans, and grading rubrics were not in any consistent digital form. Plugging an off-the-shelf model in without those would have produced a tutor that knew "math" but not "the way this school teaches math." Result: confusion and parent complaints. The right answer was a six-month digitization project for the curriculum first, then revisit AI. Boring, valuable, no AI invoice.

Case 6: NGO wanting AI to translate sensitive interviews

A local NGO wanted to use AI to transcribe and translate interview recordings with vulnerable communities into English for a donor report. Two issues. First, the people being recorded never consented to their voices going through a commercial AI vendor. Second, even with consent, the cost of an AI mistranslation in this context is not "embarrassing," it is "harmful." The right answer was a human translator bound by the organization's existing data-handling policy. AI had no legitimate place here.

Case 7: family-business owner wanting AI to "predict the future"

A textile importer wanted AI to "tell us what to buy for next season." Forecasting demand with AI requires three years of clean sales data, supplier delivery times, and inventory history. He had a notebook and a folder of WhatsApp screenshots. I would have happily taken the project and produced impressive- looking charts. They would have been astrology. The right answer was to spend the year cleaning up the back-office data, and re-quote the forecasting work next year when the inputs existed.

The patterns these cases share

Look across all seven and the same three signals show up:

1. Volume is too low for the AI build to payback. AI loves repetition. Hand-craft beats AI when the repetition isn't there.
2. The data foundation isn't ready. AI is the last 20% of the work, not the first. If steps 1–80 are paper and WhatsApp screenshots, an AI on top is decoration.
3. The cost of a wrong answer is high. Health, legal exposure, irreversible decisions. AI is great where mistakes are recoverable. Where they aren't, AI plays a supporting role to a human, never a leading one.

If you suspect your project is on this list

Send me a paragraph about what you are thinking of building and why. If your project belongs on this list, I will tell you, and I will tell you what to do instead. If it doesn't, I will quote the work normally. Free, takes about ten minutes. Use the contact page or WhatsApp.