I’m about to save our company 2 hours a month in invoice processing and renewal checks… So far, I’ve “only” invested 80 of my own hours into working with Claude Code, and I’m not halfway there yet. I alternate from excitement and amazement to horror and depression every other build, test, and commit.
I keep going because I can’t stop, and somewhere in the middle of my frustration, I realized that the AI tools actually make the foundation way more important.
Even Before the Foundation
We recently had a duplex built on land near us for some relatives to rent and for my office upstairs, and after going through that process, I may never want to build again. I learned a lot more than I wanted about the care that is supposed to go into selecting a foundation site, prepping the ground for it, and keeping the ground cared for while they prep to pour in the middle of winter…
It’s the same with your AI project. Everyone wants to imagine the second floor. No one wants to think about the work before you even start digging the cellar hole.
Your AI Doesn’t Know What It Doesn’t Know (Or What It’s Forgotten…)
I assumed early on that the conversations I’d already had with Claude about our business model and billing approach would carry over perfectly. They didn’t.
That’s when I started reading about context files and writing my own. Markdown files that explain what the business does, what our billing philosophy is, what the rules are (what all of our exceptions are!). All the stuff I sometimes just assume everyone else already knows. Once I built out files that explained what I was trying to build, what success looks like, and what failure looks like, I put them in the working directory and told Claude to read them along with a claude.md file with the ground rules and guardrails (my dos and don’ts.)
Before that, the process was 70% annoying and 30% exciting. Once I’d done the work of actually writing out what I wanted (and why), it flipped to only about 40% annoying.
Context files are necessary, and they don’t just help the AI – they help you.
Your AI Will Forget The Most Important Thing You Told It…
“Why is that showing overage calculation like that? We’ve already talked….” I typed out, a little annoyed, at midnight on one of the weekend nights I spent building this project early on. Just before that, the Claude Code session told me it was running auto-compaction. Once I realized the importance of context, I took care to try and provide some, but then in the heat of the moment, rebuilding version 27 of the billing spreadsheet I was building, I realized that I had “told” Claude Code what I meant and added clarity, but I never updated the context files I took care to build. Version 26 had the detail I wrestled with in versions 24 and 25. Version 27 didn’t keep it, not because of anything I did, but because the AI decided it wasn’t a detail worth preserving.
Update your context files every time you change key details or solve a hard problem.
Your AI Will Do Whatever You Let It…
AI isn’t malicious, it’s just obedient and a little forgetful. If you are a DBA reading this, you already know that your users will eventually do whatever you allow them to do. It’s no different with AI. You are the one in your environment with judgment and wisdom, and you are the one who knows what you actually meant. The horror stories out there you’ve read about AI doing the wrong things? Most of them started with someone who thought they had that situation covered. Start with minimal permissions and firewalling your development, testing, and even what the AI can touch when production-ready.
If it can do something, your AI invention just may do it, so be like a DBA – be a paranoid control freak and don’t give the permissions to do what you don’t want to happen.
Your AI Doesn’t Know Your Data Is A Mess…
Our PSA calls a customer one thing. Their Teams channel uses a slightly different name. The contract is in the name of the parent company. The Excel spreadsheet we use to plan our billing uses a longer name that doesn’t really match the others. Add in typos, exceptions, overrides, different abbreviations for our levels, differences for the same clients at the same levels in different months based on how we wrote out the various spreadsheets or entered free-form text one way for some clients in our PSA tool and one way for others, and maybe you already see where I’m going.
It took me trying to build an AI-powered tool to realize that maybe this data and database consulting company is a bit less data-driven than I’d like to admit. It also dawned on me that data really doesn’t just mean the data explicitly in a SQL Server database. Those Excel spreadsheets we use to drive processes and invoicing? Data. The PSA tool we use each day? Data. In the case of AI, even the free-form fields on the PSA tool with important client details are critical data.
We’re a simple company when it comes to data. Small. Under 25 people, with a handful of systems with critical data – the larger the company, the more data, and the more folks making their own data silos – the more complex. If your data estate is not in order, it won’t matter what kind of context and security you have – your AI initiatives will fall flat at best, and lead you down some dangerous roads at worst.
Clean data is the site prep that has to be done before a single footing is poured, or you are building something that will fall apart with catastrophic results.
Before You Start – Call Dig Safe
Before you even start to think about excavating in America, you call Dig Safe. They’ll let you know if there are hidden pipes, wires, or tanks where you were planning to put that foundation hole.
My project continues, but I’m spending more time cleaning data, building context, and limiting what the tools can do, even if it’s a bit more work on me today. The time I save later from the context and data cleaning will be worth it. And the data loss I’ll avoid by “being a cranky DBA” with the security is worth the effort.
If you’re about to begin your own AI initiative, start with your data, not the exciting tool selection – your data.
If you want help understanding what your data estate actually looks like before you start, that’s why we built our Data Estate Audit with Buck. Four weeks and you’ll know where the tangled wires and pipes are under the ground you’re about to dig on. You leave with a roadmap, a plan, and data that will work with you, not against you.