Most small businesses answer “can we use AI for this?” out loud, once, and never write it down. Then the person who heard the answer leaves, or the task changes, or someone new joins and starts using AI for something nobody sanctioned.
The fix isn’t a tool. It’s a page.
Below is the template I use. Copy it, fill it in, use it. You don’t need to credit me and you don’t need to tell me. It’s here because a standard nobody can reach isn’t a standard.
One warning before you start: fill it in for one task, not for your whole business. A general AI policy is a document nobody consults. A page about how your team writes first replies to enquiries is a document someone actually opens.
The template
1. The task
Name one task. Be specific enough that someone could watch it being done.
Example: writing the first reply to a new enquiry that arrives by email.
- What it is:
- How often it happens:
- Who does it now:
2. What the model may draft
List what the model is allowed to produce. Draft means draft — nothing here leaves the building without step 4.
3. What a person must keep
This is the important section. List what never goes to a model, even as a draft.
My default, which you’re welcome to adopt: anything that produces a figure someone will act on. Prices, dates, quantities, tax treatment, deadlines. A badly written sentence announces itself. A wrong number does not.
Who is accountable for this task by name:
4. Good enough to send
The test that decides whether an output ships. Write it as things you can check, not as a feeling.
The mistake almost everyone makes is testing for fluency. The model is always fluent. Test for completeness instead:
- Does it answer everything that was actually asked?
- Does it ask for everything we need in order to proceed? (the most common failure — a fluent reply that omits one question, forcing a second round trip)
- Are all figures, dates and names taken from the source, not generated?
- Would we be comfortable if this were forwarded to someone else unchanged?
- (add your own, specific to this task)
If any box is unticked, it does not go.
5. When it goes wrong
Not if.
- The most likely failure here is:
- Who notices, and how:
- What we do in the first hour:
- What we tell the client:
- Who decides whether to keep using AI for this task afterwards:
6. How we’ll know it worked
Not “it feels faster.”
Count the same thing before and after:
- Number we’re counting: (suggested: how many times work on this task came back to be redone)
- Before, over ___ weeks:
- After, over ___ weeks:
- The number that would make us stop:
Three notes on using it
Filling in section 3 first works better. Deciding what AI may do is hard and arguable. Deciding what it must never touch is usually obvious, and once that boundary exists the rest gets easier.
A blank section 5 is the warning sign. If nobody can say what happens when the output is wrong, the task isn’t ready to be automated. That’s a useful finding, not a failure.
Review it when the task changes, not on a schedule. Calendar reviews get skipped. A standard tied to a task naturally comes up when the task does.
Why this is free
I sell this work — I write these documents for other people’s clients, as a subcontractor, at a fixed price per task.
The template isn’t the valuable part. The valuable part is the hour spent arguing about what belongs in section 3 for your specific business. Handing over the blank form costs me nothing and might save someone a bad week.
If you fill it in and it doesn’t survive contact with your actual work, I’d genuinely like to know which section broke. That’s the kind of thing that makes the next version better.
Hirotoshi Yamaguchi — Guardrail Ops, Gold Coast QLD. I work entirely in writing: no calls, no video.
One workflow, written down — A$300
Pick one task you do over and over. You get back three to four pages separating what AI may draft from what a person must keep, with a completeness check and failure handling. Five business days. No data required, no video calls.
Guardrail Ops — Hirotoshi Yamaguchi, Gold Coast QLD. Get in touch