One Concept, One Example, One Question

The concept

Use AI for work that needs judgement; use rules for work that needs repeatability.

Before deciding how to use AI for a task, you should first decide if it needs AI at all.

A task can look like an AI problem because it involves messy text, routing decisions, or support for a human judgement. Sometimes it does, yet some work is predictable enough that ordinary workflows will do the job more reliably: rules, templates, validation checks or lookup tables, to name a few.

AI is not free just because the first output is quick, someone still has to explain, check and maintain it, and decide what happens when it behaves unexpectedly. If the task has a known input, a known rule, and a known output, a simpler method is often easier to inspect and therefore easier to trust.

Save AI for the parts of the work that actually need it, instead of turning every improvement idea into a model-shaped problem.

The example

A team wants to reduce the time spent sorting incoming messages. One option is to use AI to read each message and decide where it should go. That may be sensible if the messages are varied, ambiguous, and depend on context: a model might help spot the difference between a routine request, a complaint, a safeguarding concern, and a case that needs a manager's judgement.

However suppose the immediate problem is narrower. The team only needs to flag messages that contain one of a small number of agreed phrases, then route those messages to an existing review queue. In that case, AI may add more uncertainty than value. A simple rule can be tested against examples, explained to colleagues, adjusted when the phrase list changes, and audited when something goes wrong.

If the uncertainty is in the meaning of the message, AI may help as part of a reviewed workflow. If the work is only looking for agreed signs and moving the message to a known place, a rule is probably enough. The judgement then belongs in choosing and reviewing the rule, not in asking a model to make every individual routing decision.

The team may still use AI later, but it has not made AI the default answer before understanding the task.

The question

Ask yourself:

Which part of this task genuinely needs judgement?

Naming that part tells you whether AI should help with it, support a person around it, or stay out of it. If you cannot name it yet, that is the real task: designing the rule, template, checklist, or workflow that makes the work dependable before adding a model to the middle.

Related

• Explain why AI strategy should not start with tools is the wider version of the same argument: useful AI adoption starts with the work, the constraint, and the capability being built, not with the tool.

Elsewhere this week…

• I published A Human in the Loop Needs a Job, which argues that human oversight is only meaningful when the person has a defined decision, usable evidence, authority to intervene, and a route for what they learn to change the system.

• The SecondBrain Framework update tightens the Digital TeamMate operating rules: answer the question asked, distinguish genuine blockers from deferrals, use UK English, commit code changes deliberately, prefer project devcontainers, and keep weekly newsletter and blog-review reminders on the right cadence.

Support

Everything I publish here is free to read, and always will be.

I write, research, build, and maintain this newsletter and my website in my own time. If something here has been useful to you, and you would like to help cover the costs behind that work, you can support me through Ko-fi.

Generative AI is a powerful tool. I openly use it to conduct research, challenge my thinking, generate drafts, and review my work. AI never owns what I think, or my personal judgement. Everything I publish is a reflection of my own thoughts and opinions.