Should you build your admin panel with AI?
4 min read
Ask a coding agent for "an admin page that lists our users and shows what each one pays" and you'll have something on screen before lunch. It will query your users table, call Stripe's customer list, match the two by email and draw a table. That's real progress, and it's free.
So the honest answer to "why not just build it with AI?" is that you can. The question worth asking is what happens in the weeks after that first afternoon. This post walks through where the time goes, and when building it yourself is the right call.
The first version is the cheap part
What the agent writes on day one is the happy path: one sign-in provider, one billing provider,
the first page of results from each, and a join on email.
Then real data arrives.
- Stripe returns at most 100 customers a page. Your table quietly shows the first page.
- Someone signed up as
Sam@Example.comand paid assam@example.com. They show up as two people, one who never pays and one who never signs in. - A guest checkout has an email but no customer record, so it isn't joined to anyone.
- A customer was deleted in Stripe but their subscription is still in your list.
Each of these is a small fix. None of them shows up until a support ticket does, and each one is a prompt, a review and a deploy. Baseboard's join trims and lowercases every address, pages through every result instead of stopping at the first page, files guest purchases under the buyer's email and gives orphaned subscriptions a row of their own. None of that is clever. It's just fix number forty instead of fix number one.
Every provider is its own project
The second provider costs as much as the first. Clerk, Supabase and Auth0 each have their own pagination, rate limits, error shapes and idea of what "banned" means. Paddle, Lemon Squeezy and RevenueCat each model a subscription differently, and you have to turn "every 3 months" into a monthly figure before any of them add up to recurring revenue.
OAuth makes it worse. A connection made by signing in at the provider comes with a token that expires, so something has to refresh it before it does, store the new one and cope when the refresh fails.
An admin panel that covers sign-in, billing, hosting and errors is four of these integrations at minimum. Baseboard connects to every provider on its integrations page, and keeping them working when an API changes is our job rather than yours.
The admin panel holds the most dangerous keys you have
A panel that can refund a payment needs a Stripe key that can refund any payment. One that can roll back a deploy needs a token that can deploy anything. Put those in your app's environment and the admin panel becomes the part of your system an attacker most wants, usually written with the least review because "it's internal".
Building it with an agent adds a twist: to test the code, the agent needs those keys too. They end
up in .env files, in your shell history and in the agent's context.
Baseboard seals each key separately (AES-256-GCM, bound to its workspace, project and provider), and a single broker is the only code that can open one. Agents that work on your Baseboard never see a key at all. The security page has the detail.
Permissions are the feature everyone adds last
The first version of every homemade admin panel has one role: whoever can sign in can do everything. That's fine until you hire someone for support and want them to look up customers but not issue refunds.
Doing this properly means a level per action per person, a confirm step for anything that changes data, and an audit log that says who did what. In Baseboard every member starts read-only, an owner allows each change person by person, and every read and change is written down. See Permissions.
When you should build it yourself
Baseboard isn't the right answer for everyone. Build it yourself if:
- Most of your admin work is your own domain logic, like approving listings or moderating posts, rather than looking up users and payments.
- You need one table from your own database and nothing else. A coding agent will do that well.
- Your rules say every tool must run on your own infrastructure. Baseboard is a hosted service.
There's also a middle path. If you have a few workflows only you can write, write them as a Custom API endpoint and Baseboard shows them next to Stripe and Clerk, under the same permissions and audit log. You build the part that's yours and skip the forty fixes that aren't.
The short version
AI makes the first version of an admin panel nearly free. It doesn't make the second provider, the pagination bug, the support hire or the leaked key any cheaper. If those sound like problems you'd rather not own, try Baseboard on the Free plan. It reads everything and changes nothing, so you can see what it finds before trusting it with more.