The philosophy
Sometimes, Not AI
The principle
A Rube Goldberg machine can turn on a light switch. It can send a ball down a track, tip a row of dominoes, release a lever, and guide a mechanical hand toward the switch. The light comes on. The machine worked.
You would still be right to ask why nobody simply flipped the switch.
I see the same kind of overengineering in the rush to put AI into every corner of a business. A problem receives an AI solution because AI is available, impressive, or expected. Soon a straightforward task depends on a model, a prompt, an integration, a review step, and a recovery plan. The solution works, but the business has inherited a machine it never needed.
Knowing when to leave AI out is part of knowing how to use it well.
01
Simple work deserves a simple system
Some problems have clear rules and repeatable answers. A calculation should produce the same result every time. A required field should be checked against a known condition. Information that moves from one system to another should arrive intact. These jobs benefit from dependable logic that can be tested from beginning to end.
AI introduces uncertainty into that kind of work. A generative model interprets an instruction and produces a response rather than following a fixed sequence of rules. That flexibility is useful when language, context, or ambiguity matters. It becomes a liability when there is one correct answer and the system is expected to return it exactly.
Using AI also adds work around the task. Someone has to define acceptable output, test edge cases, monitor performance, review uncertain results, and decide what happens when the model gets something wrong. Those safeguards are part of the system. If their cost exceeds the burden of the original problem, the implementation has moved you backward.
This is the complexity trap. The task looks more advanced while the operation becomes harder to trust.
If you bring me a problem, I am not going to force AI into the recommendation to prove that I understand AI. My responsibility is to solve the problem in a way you can afford, maintain, and depend on.
Sometimes the answer is a better spreadsheet. A clean database can remove confusion that no model should be asked to interpret. A conventional automation can transfer information according to fixed rules. A manual step can remain the best choice when it is infrequent, important, and requires judgment that belongs with a person. Simplicity is an advantage when it matches the work.
I compare the complexity of the problem with the complexity of the proposed system. I want to know whether the rules can be written down, whether the correct result is known in advance, how much variation appears in the input, and what an error would cost. I also look at the frequency of the task. Building and supervising a system for work that happens twice a year rarely gives you much back.
The least complicated answer accounts for more than the size of the first build. It remains understandable once maintenance, oversight, failures, and future changes enter the picture.
02
I look for the least complicated answer
03
AI has to earn the uncertainty it introduces
AI becomes useful when the work contains variation that fixed rules cannot handle reasonably well. It can help make sense of large collections of unstructured material, work across natural language, recognize patterns, or prepare information for a person who needs to make a decision. In those situations, flexibility provides capability that traditional automation cannot easily reproduce.
Even then, I want a clear reason for accepting uncertainty. The value of the output has to justify the review it requires. The consequences of an error need to be understood. There should be a defined point where a person takes over, especially when the decision affects a customer, a commitment, or sensitive information.
I am interested in the smallest role AI needs to play. A model can handle the ambiguous portion of a task while ordinary software manages the rules, records, and controls around it. Keeping those responsibilities separate makes the system easier to test and gives you fewer surprises to manage.
There is no prize for using the most advanced technology in the room. You benefit when the problem is solved reliably and the solution does not demand more attention than the work it replaced.
That is the standard I bring to an AI project. I will recommend AI when its flexibility creates enough value to carry the added cost and uncertainty. I will recommend a simpler tool when the task calls for one. If all you need is to turn on the light, I would rather help you reach the switch than spend your money building the machine around it.
