This website uses cookies

Read our Privacy policy and Terms of use for more information.

Your AI tool knows a business that no longer exists. An AI tool’s file card, loaded on day one with brand guidelines, service descriptions and a project brief, sits above three changes: guidelines rewritten, process changed, services updated. None of that ever reached the tool. Why it happens, and one thing to do about it this week.

You set up an AI tool for your business, maybe a custom GPT or a Claude project, and loaded it with everything it needed: your brand guidelines, your service descriptions, your current project brief. At first it was the best thing you had.

Then it started answering out of a version of your business that doesn't exist anymore. The guidelines you rewrote. The process you changed after a project went sideways. None of that ever reached the tool, and the tool never said a word about it.

That quiet drift is the first clue to why your AI tools keep failing. It took me the better part of two years to see what was sitting underneath it.

Fifteen Custom GPTs and One Fragmented Mess

November 2022, generative AI hit, and I went all the way in. Fifteen custom GPTs. Three Claude projects. One fragmented mess.

Every one of them was genuinely good on the day I built it, loaded with the right documents and a system prompt that made perfect sense that week, and every one of them started going stale the moment anything in the business changed, which was constantly.

I'd update a document, or change a process, or finish a project that shifted the whole context the tool had been built around, and slowly, without me really noticing, that tool would drift out of sync with the actual business. It kept referencing an old version of who I was and what I was doing, and I kept working with it, trusting it, not realizing how far apart we'd quietly gotten.

Here's what I learned the slow way. You will forget to update them.

You change a project brief, a set of content guidelines, the way you bring on a new client, and then you forget every place the old version still lives. So the tool keeps answering you confidently, out of a business you stopped running a while back, and you don't catch it, because it sounds exactly as certain as it did back when it was right.

That's a problem. That's a real problem.

When I did notice, I'd either fix it by hand or quietly stop using the thing. I kept a list of the setups I'd built, and at some point I stopped keeping it, because it had started to feel like a record of things I'd abandoned.

The Blame Comes First

When a tool stops pulling its weight, you go looking for someone to blame. I'll tell you the order I went in, in case you recognize it.

I blamed myself first. I figured there was some technical thing everybody else had and I didn't, so I bought the course. When that didn't fix anything, I blamed the tool. Then I blamed the vendor.

Every one of those diagnoses sent me shopping for something, a skill or a better tool or a better company to buy it from, and every purchase landed in exactly the same place as the one before it.

If your AI tools keep failing and you've already started blaming yourself, let me say the thing I needed someone to say to me back then. This wasn't my fault. And it probably isn't yours either.

The models are more than capable right now. Think about what that means for anyone who has set one of these tools up properly and watched it fade anyway, me included. The model behind it could do the work. The setup was careful, and every document in it was current on the day it went in.

It was ready. Their business wasn't.

What "Architecture" Means for AI Tools in a Small Business

In a business, architecture is a plain thing. It's where the information about your business lives, and how the pieces reach each other.

You already have one, even if you've never drawn it, and it's the layer that decides whether your AI tools keep failing.

Your client records sit in one app and your project notes in another, billing lives in a third, and the guidelines that are supposed to steer all of it sit in a document you haven't opened since last year, so the only thing that actually connects any of it is you, walking between them.

Now add AI to that.

Every AI tool you bring in becomes its own island, holding its own copy of your business, its own version of your documents, its own idea of who you are and what you do. The day you change your services, you've got four places to update where there used to be three. Every smart tool you add to a disconnected system multiplies the fragmentation.

So say you have a tool that helps you with copywriting, another that builds your project roadmaps, a project management tool, and another that writes your video scripts, and the output of one is supposed to be the input for the next. None of those tools know about each other. Getting them to hand work along means writing that logic into each tool's instructions by hand, and rewriting it every time something changes.

Data Silos, and the Person Holding Them Together

So you fill the gaps yourself. Every week, the same gaps.

You become the bridge. The connection layer, the one carrying information back and forth between tools that have no idea the others exist.

Call it admin work if you want. It still has to come out of your head every single time, because your head is the only place the whole picture lives.

The more AI tools you add, the harder the business gets to run. You get gains, real ones, and you're creating more work in other places at the same time, and holding all of it together, remembering what lives where and which copy is current and which one quietly went stale, is exhausting. It's really, really exhausting.

This Happened Once Before, With Electricity

This has happened before, and the history is worth a minute of your time.

About a hundred years ago, American factories went through the same thing with electricity. The first move was the obvious one: bolt electric motors onto factories laid out for steam, and keep the old floor plan. Those factories saw little productivity gain. The payoff came about two decades later, once factories tore out the old layout and rebuilt around what electricity could actually do. The economist Paul David traced that lag in a 1990 paper, and it's the clearest version of this mistake I know.

That's where we are with AI. The tools keep arriving. The system they land in stays the same.

The Question I Should Have Been Asking

For a long time the only question I knew how to ask when something broke was which tool I should try next.

The question that changed things was: what system should this tool live inside?

That one swap changes what you notice. You stop reading app comparisons and start watching how information actually moves through your business, where it gets handed from one place to the next, and where it stops moving, which is almost always somewhere you're standing.

You can't automate what you haven't defined. If a process only lives in your head, nothing else can run it: a new hire can't, your website can't, and the AI tool you're already paying for can't either.

One thing to do this week

Pick one outcome your business owes somebody, start to finish, from the moment it gets asked for to the moment it's done. The project a client pays you for. The follow-up after a first call.

Write down how it really happens today, including the steps that aren't really steps, and next to each one, the name of whoever can do it. If you work alone, your name is going to be on most of that page, and that's worth seeing in writing too.

You're done when someone could read that page and tell you where the business stops without you.

It costs nothing, and it puts one piece of your architecture on paper where you can finally look at it.

The Phone Call That Changed How I Work

What finally broke the cycle for me was a phone call with my brother.

By then I'd been trying to hold everything together by wiring my tools to each other, with automations carrying information from one app to the next, and the work was tedious and brittle. It broke often. Every time anything changed, I had to go back in and fix the connections by hand.

My brother was showing me how he worked, with a coding agent running straight out of a folder on his computer, and I watched it create a file, right there in that folder, straight from the conversation. It was a small thing to watch.

That was the whole moment. If the agent could create a file, it could update one. It could move things around. Everything could finally live in one connected place, one single source of truth, and every tool could read from it. I would never have to be the router again.

So that's what I built. I run the business out of one folder now, with a place for everything and the logic of how it all fits written down, and the tools and agents I use all work from that same source. When my brand guidelines change, I describe the change in plain words and the document gets updated once, in the one place everything else reads from.

It took a while to rebuild my work around that, and it was worth every hour. My job on a project has shifted to the plan at the front and the review at the end, and agents handle a lot of what sits in the middle, depending on the task. It's given me room for the kind of planning I used to squeeze in between fixes.

Where This Goes Next

Getting your own work out of your head and off your desk is where this starts.

The same cause shows up in other rooms of the business too. It's why your marketing can sound like it could have come from any business in your category, and why your website can't answer the question a real buyer actually asks.

Whether there's someone capable standing there waiting on you for the context, or there's nobody else at all and the job simply never gets picked up, the constraint is the same one: everything the business needs to know still lives in your head.

And once the business holds its own context, the horizon moves. You can start running work you were never going to be able to afford to hire for.

There is more to this than one essay can hold. The Work Redesign works through it every two weeks.