1.1.1. Deterministic Software vs. Probabilistic AI
💡 First Principle: If a person can write down every rule and the rules never change, ordinary software executes them perfectly and cheaply. When the rules cannot be written down (recognizing a frustrated customer, spotting a fraudulent pattern, summarizing a contract), AI learns an approximation from examples. You trade certainty for the ability to handle ambiguity, and you should only make that trade when ambiguity is the real problem.
Consider a payroll system. Overtime is paid at 1.5 times the hourly rate after 40 hours. That rule is written once, executed millions of times, and is correct every time. Nobody would use AI for it, and any proposal to do so should raise an eyebrow. Now consider routing 50,000 inbound customer emails a day to the right team. The emails are written in thousands of different ways; no one can list every phrasing of "my package arrived damaged." An AI model that has learned from past emails can route 94% of them correctly in milliseconds. That is an enormous gain, and it comes with a permanent fact: 6% will be misrouted unless something catches them.
| Characteristic | Deterministic software | Probabilistic AI |
|---|---|---|
| Source of logic | Rules written by people | Patterns learned from data |
| Same input, same output? | Always | Usually, but not guaranteed (especially generative AI) |
| Handles ambiguous or novel input | Poorly; fails or rejects | Well, within the patterns it has seen |
| Errors | Bugs that can be found and fixed | A statistical error rate that is managed, not eliminated |
| Explaining a decision | Point to the rule | Harder; may require explainability tools |
| Changes over time | Only when someone edits it | Performance can drift as the world changes |
| Best for | Stable, enumerable, auditable logic | Ambiguous, high-variety, pattern-based tasks |
The practical question for a strategist is therefore not "Can AI do this?" but "Can the business absorb this error rate here?" That leads directly to a design choice between automation, where AI acts on its own, and augmentation, where AI recommends and a person decides. Low-cost, reversible errors (a misrouted email, a mediocre product recommendation) suit automation. High-cost or irreversible errors (denying a loan, a medical triage decision, a public statement) call for augmentation with human review.
⚠️ Exam Trap: When an answer option promises that AI will "eliminate errors" or "guarantee accuracy," it is almost certainly a distractor. Correct options talk about acceptable error rates, confidence thresholds, human review for high-impact cases, and monitoring.
Reflection Question: A bank's fraud model flags 2% of transactions, and 30% of the flags turn out to be legitimate purchases. Is the model "broken," or is this a business design question about who reviews flags and how customers are treated while a review happens?
