What automates and what does not
Domain-free rules for deciding whether a step belongs to the AI or to a person, and the four conditions a step must meet to be automated safely.
- Status
- stable
- Trust
- human-reviewedhuman:rrizwan98 ·
- Owner
- human:rrizwan98
- Approved
- human:rrizwan98 ·
Tagging a step is a judgement, and these are the rules that make it the same judgement every time, whatever the business.
The four conditions
A step can be automated when all four hold. Miss one and it stays with a person, or the step gets redesigned until it holds.
| Condition | What it means | |
|---|---|---|
| 1 | Repeatable | The step happens the same way often enough to be worth writing down. A step that is different every time is a decision wearing a process costume. |
| 2 | Checkable | There is a way to tell a good result from a bad one without asking the person who produced it. A test, a rule, a rubric, a comparison. |
| 3 | Reachable | The tools and data the step needs can be reached with permission: an API, a file, a database, an inbox. A step that needs someone to walk to a desk is not reachable yet. |
| 4 | Reversible or gated | Either a mistake can be undone cheaply, or the step sits behind a human gate. Irreversible and ungated is the one combination that is never allowed. |
What stays with a person, in every business
- Taste and judgement. Which of two good options fits the brand, the client, the moment.
- Relationships. The conversation where trust is built or lost.
- Price and terms. What something is worth to this buyer, and what you will accept.
- Risk and liability. Anything where being wrong costs money, safety or reputation.
- The exception. The case the written procedure did not anticipate. Automation should recognise it and stop, not improvise through it.
- The final yes. Someone signs before it goes live, leaves the building, or reaches a customer.
What almost always automates
- Intake and triage. Turning a message, a form or a ticket into a structured record with the missing questions asked.
- Gathering. Collecting the documents, logs, prices, references or history a step needs.
- Drafting. The first version of anything: a spec, a page, a reply, a report, a plan.
- Checking. Running the rules against the work: tests, checklists, scores, comparisons.
- The fix loop. Reading a failed check and correcting the work until it passes.
- Packaging and handover. The delivery note, the deploy, the summary, the log of what was done.
- Chasing. Status, reminders, follow-ups, anything whose whole content is "where is this".
Three traps
"It needs judgement" is often "nobody wrote it down." Ask what the person actually considers. If the answer is a list, it is a rule and it can be written. If the answer is "it depends, you get a feel for it", it is judgement.
Automating a broken step makes it fail faster. If a step is unclear, unowned, or done differently by each person, fix it as a process first. A system built on it will enforce the confusion perfectly.
The last mile is where the time goes. Drafting is the easy 20%. Checking, fixing, packaging and handing over are the 80%, and those are exactly the steps people leave out of the count. Include them, or the number will flatter you.
Where the safety comes from
Automating a step does not mean trusting the AI to behave. The process lives in a server, and the server refuses what the process forbids: out-of-order work, missing evidence, self-reported scores, placeholder text, an edit after approval. Read how I work for what that looks like in practice.
That is why the four conditions include "checkable": a step nobody can check is a step nobody can enforce.
Count your own workflow
How much of my own work can be automated: the four-step procedure that turns any run of work, in any industry, into a tagged step table and an honest percentage.
Worked counts
The same counting method applied to five different businesses, each with its own step table, its own number and its own human gates.