The short answer
Automation is the wrong answer when the process is undocumented, changes constantly, runs only a handful of times a week, or is really a staffing or data-quality problem wearing a process costume. Automating a bad process only makes it happen faster, and at greater volume.
Five situations where automation makes things worse
The process is not written down. If it lives in one person’s head, what gets automated is not the process — it is one person’s recollection of the process on the day you asked. Write it down first. Sometimes writing it down is the whole fix.
It is really a data-quality problem. Automation applied to inconsistent data produces inconsistent output at speed, and it produces it with an air of authority that makes it harder to spot. Fix the data, then automate.
It is really a staffing problem. Some processes are slow because one person is doing the work of two. A system will not resolve that; it will hide it for about a quarter and then the same person will be maintaining the system as well.
The rules change constantly. If the logic changes monthly, you are not buying an automation, you are buying a maintenance commitment. Occasionally that is still worth it. Usually it is not, and nobody says so at quote time.
The volume is not there. A process that runs three times a week is rarely worth a build. It might be worth a checklist.
The tell
Ask the people who do the work what the process is. Ask them separately.
If you get materially different answers, you do not have a process to automate yet. You have four processes with the same name, and the first job is deciding which one is correct — a management decision, not a technical one, and not one anybody external should be making for you.
Where a language model specifically is the wrong tool
Distinct from automation in general, and worth separating.
If a task has one correct answer and a rule can find it, use the rule. Models are expensive per run, non-deterministic, harder for your team to reason about later, and they fail in ways that look like success. A scheduled job that moves a file does not need a model in front of it, and putting one there is a decision made for the brochure rather than the business.
Models earn their place on judgement and language — summarising, drafting, classifying messy input, handling text that never arrives in the same shape twice. That is a real and valuable category. It is a smaller category than most proposals imply.
What being told no is worth
It is worth the price of the audit.
Plenty of processes that look automatable turn out to be a staffing question, a data question, or a process nobody has ever written down. Finding that out for $500 to $1,500 is a good outcome, and a considerably better one than discovering it after paying for a build that quietly made things worse.
Last updated 18 August 2026 by Hugo, founder of Hugo Signal.
Bring me the process that wastes the most time.
A first call takes thirty minutes. You will leave it knowing whether automation is worth doing at all in your case.