Why does something need to be an “AI project” or not? I still don’t know why we aren’t building around the business problem. If there is some LLM in the architecture, great. If not, who cares? This is not a knock at you by the way, because I think you are seeing the same thing and I’m sure someone is hiring you saying “we need AI here.” Just something I’m also wrestling with myself.
Thank you for taking the time to leave your comment. This is how I see it. AI or not AI is key for three main reasons.
First, "normal" software being deterministic (if, then, else) is inherently lower risk by comparison to the new world of non-deterministic processing via an LLM.
The second point involves cost - The more autonomy you want in an AI system, the higher the price. The operating costs of running AI systems are much more variable (mostly to the upside) than other systems.
The third point is jurisdiction - something I will write about next. There's many parts of the world where LLM processing outside of a nation state are simply prohibited. The META region in particular faces this challenge.
When you pull just these three elements together, they form the foundation of decision making that must be part of Question Zero - otherwise, the downstream impact is frankly, career wrecking for many people.
Right okay this makes a ton of sense! Usually I don’t think it makes sense that everyone know what’s under the hood. But the jurisdiction thing is super important and the heads up about those darn API costs. I think most AI products would still be mostly deterministic with a judgment layer, but even then, you need evals, a bunch of noise to approve…
Beyond honesty in your work, does the AI/not-AI distinction really matter? If the AI hype gets businesses (and cities) thinking seriously about automation, that's a win. My worry is practical: pitch an honest non-AI solution and you may lose to the vendor who shoves AI in the middle for no reason. I spend a lot of my time telling my clients that they don't need an AI solution but can use AI to build the automation.
This maps to clinical AI because the label changes the safety case.
A deterministic workflow tool, an LLM judgment layer, and an agent that can spend tokens, call tools, or move state need different evals and owners.
In an OR-adjacent setting, "is this actually agentic?" becomes a practical question: what action can the system take, what cost or jurisdiction boundary follows it, and who owns the exception when the model's path is nondeterministic?
Why does something need to be an “AI project” or not? I still don’t know why we aren’t building around the business problem. If there is some LLM in the architecture, great. If not, who cares? This is not a knock at you by the way, because I think you are seeing the same thing and I’m sure someone is hiring you saying “we need AI here.” Just something I’m also wrestling with myself.
Thank you for taking the time to leave your comment. This is how I see it. AI or not AI is key for three main reasons.
First, "normal" software being deterministic (if, then, else) is inherently lower risk by comparison to the new world of non-deterministic processing via an LLM.
The second point involves cost - The more autonomy you want in an AI system, the higher the price. The operating costs of running AI systems are much more variable (mostly to the upside) than other systems.
The third point is jurisdiction - something I will write about next. There's many parts of the world where LLM processing outside of a nation state are simply prohibited. The META region in particular faces this challenge.
When you pull just these three elements together, they form the foundation of decision making that must be part of Question Zero - otherwise, the downstream impact is frankly, career wrecking for many people.
Right okay this makes a ton of sense! Usually I don’t think it makes sense that everyone know what’s under the hood. But the jurisdiction thing is super important and the heads up about those darn API costs. I think most AI products would still be mostly deterministic with a judgment layer, but even then, you need evals, a bunch of noise to approve…
Beyond honesty in your work, does the AI/not-AI distinction really matter? If the AI hype gets businesses (and cities) thinking seriously about automation, that's a win. My worry is practical: pitch an honest non-AI solution and you may lose to the vendor who shoves AI in the middle for no reason. I spend a lot of my time telling my clients that they don't need an AI solution but can use AI to build the automation.
This maps to clinical AI because the label changes the safety case.
A deterministic workflow tool, an LLM judgment layer, and an agent that can spend tokens, call tools, or move state need different evals and owners.
In an OR-adjacent setting, "is this actually agentic?" becomes a practical question: what action can the system take, what cost or jurisdiction boundary follows it, and who owns the exception when the model's path is nondeterministic?