/03

Where AI Fits

A workflow-by-workflow test for finding useful AI opportunities before anyone buys software.

Beginner

10 min

Open +

Ask six boring questions about the work. You will learn more than you will from a hundred AI demos.

Leave knowing / Find your best first use cases

The six-question scan

Walk one workflow from start to finish.

What repeats?

Look for recurring reading, writing, sorting, comparing, finding, explaining, or routing.

What is slow?

Find handoffs, queues, searching, copying, review, and waiting for someone to assemble context.

What is messy?

AI is often useful where the input is language, documents, images, audio, or inconsistent human information.

What requires judgment?

Separate assistance from final authority. A task can be AI-assisted without giving AI the decision.

What can be checked?

The safest useful work has an answer, source, schema, rule, test, or human review that can catch mistakes.

What happens if it is wrong?

Start with reversible errors. High-impact decisions need stronger controls and often should stay human.

Fit test

A quick way to classify the opportunity.

Work looks like	AI fit	Why
Read, summarize, compare, draft	Strong starting fit	Language models are built for this kind of pattern work.
Find information across known sources	Strong with retrieval	Useful when answers can point back to evidence.
Route or classify repeated cases	Strong with checks	Clear categories and examples make this measurable.
Multi-step software work	Possible with tools	Needs permissions, state, error handling, and verification.
High-stakes final judgment	Assist, do not hand over	Human accountability and stronger evidence remain essential.
Simple deterministic math or rules	Use normal software first	Do not pay a model to do what a formula can do perfectly.

Score the pilot

A first pilot should say yes to most of these.

It happens often enough to matter.

The current pain is visible and measurable.

The output can be reviewed or verified.

A mistake is reversible.

The required data can be accessed safely.

One person clearly owns the workflow.

We can compare before and after without inventing ROI.

Use this on a workflow

Make the AI argue against using AI.

Try the template

I am considering using AI in this workflow:
[DESCRIBE WORKFLOW]

Analyze it step by step. For every step, tell me:
- What the person does now
- Whether normal software, AI assistance, or an AI agent is the best fit
- What AI would need to see or use
- What could go wrong
- How the output could be verified

Then recommend the smallest useful pilot. If AI is unnecessary, say so clearly.

Notes / Subscribe

No schedule. No content calendar. A note goes out only when there is something worth understanding.