Same Input, Same Output? Not With an LLM. Ever.
Large language models break the one rule every business system is built on. Not occasionally — by design, and forever.
FOR: Operators · weighing where AI belongs in their workflows
Quick answer
Large language models are probabilistic by design, not deterministic. A setting called temperature makes the model pick a less-likely word on purpose, so the same input can give you a different output every run — that is a feature of the engine, not a bug. Combined with the model’s trained instinct to please, it will also invent a source rather than admit it does not know one. The practical rule: use LLMs for drafts, summaries, and first-pass triage where variance is fine; keep numbers, commands, and client-facing outputs on deterministic code.
We are all comfortable with one rule of business software: same input, same output. A report pulls the same numbers every time. An invoice adds up to the same total every time. That predictability is the paradigm every business system is built on.
Large language models break that rule. Not occasionally — by design, and forever.
The temperature dial
Every LLM has a setting called temperature — a dial for how factual versus how creative it is. At 0.0 it plays it as straight as it can. At 1.0 it is at its most creative. It’s easiest to visualize it as 0 to 1, with 0.7 as a baseline for the standard conversation. OpenAI, Anthropic, Google may have different scales, but the core idea is the same.
Under the hood, that dial changes how the model picks its next word. Near 0.0 it takes the single most-likely option every time. Turn it up and it will happily reach for a word that was only the tenth-best guess. Same question in, different answer out.
For a business process, that is the problem. Automation is supposed to be repeatable and auditable. Drop an LLM into the middle of a workflow and you have added a step that can answer differently on Tuesday than it did on Monday.
The sycophant problem
Now add a second trait. LLMs are trained to keep you happy — there are guardrails pushing them to be agreeable and helpful above almost all else. So a model would rather invent an answer than admit it does not know. Ask it for a source and it will give you one, confidently. Whether that source actually exists is a separate question. Left unchecked, “helpful” quietly becomes making things up. We call it hallucinating. Really, it is a sycophant telling you what you want to hear.
Drawing the line
So the real skill is not wiring an LLM into everything. It is drawing a clear line between what has to be deterministic and what genuinely needs thought. Numbers, commands, anything that lands in front of a client — deterministic, every time. Drafts, summaries, first-pass triage — those are fair candidates for thought.
That line is the main thinking point behind my auto-remediation and LLM pipeline project: which steps have to be deterministic, and which ones actually require thought. These ideas guide the many architectural decisions required for creating a “Harness”, something I’ll get into another time.
— Thomas
This is one operator’s technical read of how LLMs behave inside a business workflow - not a universal prescription. Model behavior varies by vendor, model version, and configuration; test yours before you trust it with anything client-facing.
→ Talk to us about where AI fits