You’ve seen the demos. A slick interface where an AI agent spins up a server, deploys a container, and configures a database. The crowd goes wild. The executives open their checkbooks. And then you look under the hood and realize they just spent $50 in API tokens to do what a 20-line Python script has been doing flawlessly since 2012.
We are currently living through an era of technological masochism, where organizations are actively forcing artificial intelligence into every conceivable crack of their workflow, regardless of whether it actually makes sense.
Using a multi-billion dollar language model to run a script you could have written in Bash is like buying a Ferrari to deliver newspapers.
Recently, a popular tech blog suggested three Claude skills every organization should have. Two of them involved using the AI to “create” and “deploy” resources. The engineering community’s response was swift and brutal: this is a pointless waste of tokens. Why? Because creating and deploying are deterministic, reproducible, and scriptable. They were solved problems long before LLMs ever ingested their first byte of training data.
The push to use LLMs for every task clashes violently with the reality of software engineering. If a process can be mapped out in a flowchart, you don’t need a probabilistic engine guessing its way through it. You need a script. Scripts don’t hallucinate. Scripts don’t cost $0.03 per execution. Scripts don’t suffer from context-window amnesia.
Forcing an LLM to do a script’s job doesn’t make you an AI pioneer; it makes you a token-burning tourist.
The real leverage of large language models isn’t in replacing existing automation. It’s in addressing the “last mile” of unstructured, ambiguous tasks that no script could ever handle.
Think about the messy, undocumented, human-centric problems in your organization. Parsing a chaotic, unstructured email thread from a frustrated client to extract actionable requirements. Analyzing a decade of ad-hoc customer support transcripts to find hidden friction points. Navigating a legal contract where the meaning is buried in nuance, not just keywords.
You cannot write a deterministic regex for “read the room.” You can’t script “figure out what the user actually meant, not what they typed.” That is the native habitat of the LLM.
The most valuable AI skills aren’t about doing the predictable faster; they’re about making the unpredictable computable.
Organizations need to wake up and distinguish between shiny new toys and genuine productivity gains. If you are building AI agents to execute tasks that have a strict, known sequence of steps, you are wasting time, money, and the fragile trust your team has in this technology. Every time an LLM hallucinates a deployment step that a simple CI/CD pipeline would have executed perfectly, trust in AI dies a little death.
Stop trying to automate the automated. Let scripts be scripts, and let AI tackle the beautiful, messy, unstructured chaos that actually requires intelligence. That’s where the revolution is happening—if you’re sharp enough to see it.
FAQ
Q: Why shouldn't I use an LLM if it eventually gets the deployment right?
A: Because 'eventually' isn't good enough for deterministic tasks. Scripts execute perfectly every time without hallucinating, cost a fraction of a cent, and don't suffer from context window limits. Using an LLM here introduces unnecessary risk and expense.
Q: What's the practical implication?
A: Audit your AI workflows immediately. If a task can be flowcharted or follows a strict sequence of steps, rip out the LLM and replace it with a deterministic script. Save your AI budget for ambiguity.
Q: What's the contrarian take?
A: Most 'AI Agent' startups today are just fragile wrappers around basic scripts that could have been written five years ago. When the novelty wears off and token costs are scrutinized, they will collapse.