The Half-Built Audit

Open the sidebar and count the chats

Go and open whichever AI tool you use most, and look at the list of chats down the side. Actually count them.

If you’re anything like most people who got started early, there are forty-something, most of them named after the first few words you typed late on a Sunday night. Somewhere in there is the version of that client email that finally landed. You won’t find it in under ten minutes, so on Thursday you’ll write it again from scratch.

I don’t think that’s a discipline problem. I think it’s an inventory problem. And it usually isn’t only the chats.

There’s often a custom GPT someone built in a workshop that nobody has opened since. An automation that works most of the time, where the failures get fixed by hand and never written down. Prompts saved in Notion, and in Slack, and in a Google Doc called “prompts FINAL v3”. Maybe a chatbot on the website whose conversations nobody reads.

None of it talks to anything else. So you don’t really have one AI problem. You have four unfinished ones, and until they’re all on one page, you can’t tell which of them deserves another hour.

Set aside ninety minutes. That’s enough to get the list done.

Find everything before you judge anything

The judging comes later. Right now you’re only making a list, and it’ll be longer than the one in your head.

1. Your own accounts. Open every AI tool you pay for and write down every custom GPT, project, saved instruction, uploaded file and pinned chat.
2. Other people’s accounts. This is the step people skip, and it’s often where half of it lives. Ask your VA, your contractors and anyone who ran a project with you this year: what AI have you set up for us, and whose login is it under? Ask in writing, so the answer comes back in writing.
3. The automation layer. Open Zapier, Make or n8n. List every workflow, and next to each one write the date it last ran cleanly and how many errors it’s thrown.
4. Wherever prompts hide. Search Notion, Google Drive, Slack and your notes app for you are a, act as and prompt. You’ll find things you wrote and completely forgot about.
5. Anything a client can touch. The website chatbot, auto-replies, booking flows, and the AI features that got switched on by default inside your CRM, helpdesk or email tool. Default-on still counts.
6. The card statement. Go back three months. There’s usually at least one subscription you forgot you were paying for.

If you want help with steps one to four, paste this into your main AI tool and let it interview you:

I'm auditing every AI tool, automation and saved prompt in my business.

Here's what I know exists so far: [paste your rough list]

Ask me one question at a time until you have, for each item:
- what it does, in one sentence
- who built it, and when
- which account or login it lives under
- who has actually used it in the last 30 days
- what breaks, and for whom, if it stops working tomorrow

Don't summarise until you've asked about every item. Then prompt me
for the categories I've probably forgotten: browser extensions,
meeting note-takers, anything embedded on my website, anything a
contractor set up, and AI features switched on by default inside
tools I already pay for.

Finish by giving the list back as a table with those columns, sorted
by how much of my week depends on each item.

Four honest questions sort the whole list

Take each item and ask these four questions, in order. Answer them about last month, not about what you were planning to do.

Does anyone use it? Not “could someone use it”. Name the person and the date. A tool nobody uses isn’t an asset in progress. It’s a decision you’ve already made without saying it out loud.

Does it produce something you’d send to a client without editing? Find the last thing it made and read it now. If you rewrote the opening line every single time, the answer is no.

If it broke tomorrow, would you notice? Some automations quietly fail for weeks while the work still gets done, because a person picked up the slack without telling anyone. So ask the person, not the dashboard.

Who owns it? Say the name of someone who still works with you. If the honest answer is “whoever’s free”, or a login that gets passed around a group chat, then nobody owns it.

Score it so you can’t talk yourself out of it

Give each item 0, 1 or 2 for each question. Eight is the top score.

Question	0	1	2
Does anyone use it?	Nobody in 30 days	One person, now and then	A named person, weekly or more
Would you send its output unedited?	Never	After a solid rewrite	Yes, after a read-through
Would you notice if it broke?	No	Within a fortnight	Same day
Who owns it?	Nobody, or someone who left	Shared, unclear	One named person, still here

Keep, fix or kill, and the rule for each

Keep (6 to 8). It isn’t really kept until it’s written down. One page, this week: what it does, where it lives, who owns it, and what to check when it misbehaves. An undocumented keep tends to become next year’s kill.

Fix (3 to 5). You’re only allowed to fix it if you can name the broken part in one sentence. “It doesn’t really work” doesn’t count. Give it two hours and a date. If it’s still broken on that date, it moves to kill.

Kill (0 to 2). Killing means turning it off, cancelling the plan and removing access, not leaving it sitting there. Before you do, copy the prompt text into one file. The tool around it is replaceable. The wording is usually the part someone actually thought hard about.

There are two exceptions. If something scores well on use but zero on ownership, that’s an access risk sitting in someone else’s account, and it gets sorted this week whatever the total. And anything a client can see that scores under 6 gets switched off today. A chatbot answering your customers badly is worse than no chatbot, and you’ll be the last to hear about it.

The first thing to rebuild is almost never the flashy one

Look at how everything scored on question two. I’d be surprised if most items didn’t fail the same way, for the same reason: the tool had no idea who you are, who you’re talking to, or what good looks like in your business. It was guessing, quite competently, at a generic version of you.

Rebuilding the chatbot won’t fix that. Writing the context down once fixes it in every tool at the same time, including the ones you haven’t bought yet.

So the first rebuild is really two pages. One is a business context file: what you sell, to whom, in the words your clients use. The AI Foundations Checklist has a template. The other is a voice file, the rules and real examples that make the output sound like you, and The Voice File has the prompt that builds it.

Then take the worst item on your list, paste both pages in at the top of the prompt, and run it again. The difference between the two outputs tells you what to build next.

When not to bother

If you have fewer than three AI things and all of them get used every week, skip this. That’s not a mess, that’s a small setup, and your time is better spent on the foundations.

If you’re in the middle of a launch or end of financial year, audit only the things clients can see, and park the rest until the week after.

If one person built all of it and they’re leaving in a fortnight, this isn’t an audit, it’s a handover. Sit beside them, record their screen with their permission, and move the logins over.

And if someone has told you the answer is to cancel everything and start clean, I’d push back on that. The prompts and the context took the real effort. The tools around them can be replaced in an afternoon.

What you’ve got now

One list of every AI thing in your business, including the ones sitting in other people’s accounts. A score for each that you could defend to someone who disagrees with you. A decision, in writing, for every item. And the two pages of context that make the next build better than the last one.

Start with step two. Send that message to your contractors before you open a single tool, because their answers take the longest to come back.