My AI found four dead systems in my business. Nothing had broken.
This week I gave my AI one instruction: review everything. Every tool, every scheduled job, every project. It runs on Fable 5 now, Anthropic's newest model, and it went through my whole stack in about an hour. 197 tools, 45 scripts, 300 automated tests, 54 active project files. The report came back and the worst findings were not the things that were broken. They were the things that were dead, and had been dead for weeks, without a single complaint from anyone. Including me.
What it found
Four systems, silently gone. My nightly business digest email had not sent in weeks. Two separate bugs had stacked on top of each other, and the second one was almost funny: the tool was blocking its own send because its output violated one of my own writing rules. My operations dashboard had been down since a reboot. A weekly self-improvement job had failed five weeks in a row. And a shared tool that other tools use to call AI models was quietly pinned to a model that got retired in June, so everything that depended on it was getting worse answers with no error message.
Every one of those failures had left evidence in a log file. The logs were right there the whole time. Nobody reads logs until something hurts, and nothing hurt.
The question that actually matters
Here is the uncomfortable part. My business ran fine for those weeks. Clients got served. Emails went out. Money moved. If a system can die and nothing changes, you have to sit with the obvious question: did I ever need it?
I think the honest answer has three parts, because dead is not one thing.
Three kinds of dead
The slow fuse. The digest email is not something I miss on day one. It is awareness. I feel its absence in week six, when I realize I have drifted and I am making decisions on stale information. Some systems fail silently precisely because their value is preventive. Dead does not mean useless. It means the cost shows up later, with interest.
The dead weight. The review also found I have eight versions of one banner tool, six different transcript tools, and five ways to send an email. Built in enthusiasm, kept by inertia. That drawer in your shop where every tool is dusty except the three you actually reach for. Usage is the only honest test of need, and most of that drawer fails it.
The scary middle. The worst category is not dead at all. The AI-calling tool with the retired model looked alive. It answered every request. The answers were just quietly worse. Dead systems are honest once you look. Degraded systems lie to your face. That one I did need, and I had no idea it was hurting me.
Your business has these too
I run an AI company and this still happened to me, so I will not pretend your shop is different. You have subscriptions nobody has logged into since winter. Reports that generate and go unread. A sync that stopped syncing. An answering flow that half works. The test is simple and free: for each one, ask when it last proved it was alive. Not whether it is supposed to run. When it last showed evidence.
Then sort what is left into the three piles. Fix the slow fuses, because those protect you. Cancel the dead weight without guilt, because the dust already voted. And go hunting for the scary middle, because the things that look alive and are quietly wrong are the ones costing you real money right now.
The shorter version
I pointed Fable 5 at my whole infrastructure and told it to review everything. It found four systems that had been dead for weeks with nothing visibly broken, fixed eight things the same day, and left me with a better question than it started with. If it can die and nobody notices, it is either preventive, pointless, or lying to you. Every business owns all three. The difference is whether you find out on your schedule or on a customer's.
I do this exact review for businesses. Ask me about it.
← Back to all posts