Everyone wants to know: "Should we use AI for this?"
The answer is not yes or no. It's "it depends."
It depends on two things:
1. Interplay of Cost of Error and Human Judgment Required
2. Things AI will fail at
The Matrix
The interplay between Human Judgement Required and Cost of error determines everything. As the Cost of Error and Human Judgment Requirement rises, AI's value collapses. Similarly, if Cost of error and Human judgement requirement are low, AI is a fit case, but it needs to justify its place with traditional algorithmic automation. AI fits well between these two extremes.
The Failure Modes: Why AI Gets Overconfident
Three specific reasons AI fails when you use it wrong:
1. Training Data Blindness
A research team asked an AI to identify Indus Valley Civilization script. The model had never seen it. It confidently generated analysis. It didn't say "I don't know." It hallucinated as you are asking it something completely outside its training.
This happens in your company too. Ask an AI about an internal process it's never seen. A market that didn't exist during training. A regulatory requirement specific to your jurisdiction. It will sound smart while being completely wrong.
2. Context Collapse
A hiring algorithm looked at successful employees and worked backward. It couldn't see that the factor was "hired during economic boom with low unemployment" or "worked for previous CEO who specialized in finding diamonds in the rough." It saw patterns in the data and learned to replicate them. Perfectly. And discriminatorily.
Undocumented context kills AI decisions. If the reason lives in someone's head or an email thread, AI can't see it.
3. Liability Transfer
An AI chatbot approved a refund policy. The airline approved the AI. The customer is now demanding a refund. Who's liable?
The customer doesn't care that "the AI decided." They care about getting their refund. And the airline's lawyers now care about whether it was negligent to let an AI approve policy changes.
The Decision Test
Before deploying AI to any decision, ask three questions:
If this is wrong, who bears the cost?
- You bear it → Use humans with AI research
- A customer bears it → Use humans with AI research
- Nobody bears much → Use AI freely
Is this in the AI's training data?
- Yes → AI can help
- No → Humans must decide
- You're unsure → Ask yourself: "Could a company train an AI to do this?" If yes, it's in the data. If no, it's not.
Does this require judgment we haven't documented?
- Yes → Humans lead
- No → AI can handle it
The Pattern Most Miss
The best companies don't ask: "Can AI do this?"
They ask: "Should AI do this?"
One gets you innovation. The other keeps you out of the news.
Posts
