Last month, I talked a client out of a $2M AI project. They wanted to use GPT-5 for real-time financial fraud detection. After two days of discovery, I recommended they stick with their rule-based system. They were shocked. Here's why I was right.
The Uncomfortable Truth About AI
We're in an AI gold rush. Every company wants it, VCs fund it, and engineers (myself included) love building it. But here's what nobody talks about at conferences:
Most problems don't need AI. They need better data models, cleaner code, or simpler UX. AI is often a $500,000 solution to a $5,000 problem.
My Framework: The 5 Red Flags
🚩 Red Flag #1: Deterministic Requirements
If your problem has clear rules and predictable outputs, you don't need AI.
Real Example:
A logistics company wanted to use AI for route optimization. Their constraints were: trucks can't exceed weight limits, drivers can't work over 8 hours, and certain roads have time restrictions. This is a graph problem, not an AI problem. We built it with Dijkstra's algorithm in 2 weeks instead of 6 months.
🚩 Red Flag #2: Explainability Requirements
If you need to explain every decision to regulators or customers, AI might be a liability.
Real Example:
A healthcare startup wanted AI for insurance claim approvals. Legal required detailed explanations for every rejection. GPT-4's "because the pattern suggests" doesn't hold up in court. We built a decision tree system instead - boring but defensible.
🚩 Red Flag #3: Latency Constraints Under 100ms
If you need consistent sub-100ms responses, AI will frustrate you.
Real Example:
High-frequency trading firm wanted AI for order execution. Even with edge deployment and optimization, we couldn't guarantee <50ms latency. Their existing algorithm was already at 5ms. AI would have made them slower and poorer.
🚩 Red Flag #4: Perfect Accuracy Requirements
If 99.9% accuracy isn't enough, AI will disappoint you.
Real Example:
A nuclear facility wanted AI for safety monitoring. "99.9% accurate" means 1 in 1000 failures. In nuclear safety, that's unacceptable. Rule-based systems with redundancy achieve 99.9999% reliability. Sometimes boring is better.
🚩 Red Flag #5: Limited or Biased Training Data
If you have less than 10,000 quality examples, AI will hallucinate.
Real Example:
A startup wanted AI to generate legal contracts for a niche industry. They had 50 example contracts. The AI started inventing clauses that sounded legal but were nonsense. We pivoted to a template system with variable insertion. Less sexy, more reliable.
The $2M I Saved By Saying "No"
Remember that financial fraud detection project? Here's what I told the client:
"Your current rule-based system catches 94% of fraud with 0.1% false positives. It processes transactions in 10ms and costs $10k/month to run. An AI solution might get you to 96% detection, but with 0.5% false positives, 200ms latency, and $100k/month in API costs. That's 5x more angry customers calling about frozen cards for a 2% improvement in catch rate."
They kept their existing system and used the saved budget to hire two data analysts who improved the rules and got detection to 96% anyway.
When AI IS The Right Choice
I'm not anti-AI. I've built my career on it. But AI shines in specific scenarios:
- ✅ Unstructured data: Text, images, audio that need understanding
- ✅ Pattern recognition: Finding insights humans would miss
- ✅ Creative generation: Content, code, designs that need variety
- ✅ Natural interaction: Chatbots, voice assistants, conversational interfaces
- ✅ Prediction with uncertainty: Forecasting where 85% accuracy is valuable
The Questions to Ask Before Starting Any AI Project
- What's the simplest solution that could work?
Start there. You can always add AI later.
- Can I afford to be wrong 5-10% of the time?
If no, reconsider AI.
- Do I have the data volume and quality to train effectively?
If no, gather data first.
- Will the ROI justify the complexity?
Calculate real costs including maintenance.
- Can my team maintain this in 2 years?
If everyone who understands it might leave, simplify.
The Competitive Advantage of Restraint
Here's the counterintuitive truth: In the AI gold rush, the companies that show restraint often win. They:
- Ship faster with simpler solutions
- Spend less on infrastructure
- Have fewer catastrophic failures
- Build trust by solving real problems
My reputation isn't built on implementing AI everywhere. It's built on implementing AI where it matters and having the wisdom to recognize the difference.
The Bottom Line
AI is a powerful tool, not a universal solution. The best AI engineers know when to use it and, more importantly, when not to. Your job isn't to use the most advanced technology – it's to solve problems effectively.
Next time someone says "we need AI for this," ask them why. If they can't explain it without using the words "innovative" or "cutting-edge," you probably don't need AI.
Sometimes the most innovative thing you can do is choose the boring solution that works.
