Will We Even Need Admin Panels?

In September I updated a live website without logging in to anything.

I asked Claude to publish two articles to tools.belchamber.us, one on writing an MCP server that doesn't leak and one on quantizing a 70B model for one consumer GPU. Google's Gemini had drafted both inside my content app, and I'd already read them. Claude exported them as drafts, laid out every production step as a numbered choice, and I answered each one in the chat window with a reply a few characters long, usually just the option's code. The last gate, switching the posts from draft to live, waited for my word like all the others. They went live on September 17. When I noticed the listing page showed no excerpts for them, I said so, and the fix was live about ten minutes later. I never opened a content management system, never typed a password and never went looking for the "Publish" button that some redesign had moved into a new menu.

That's not a party trick. It's how more and more of the work on my own sites happens now. And it raises a question I didn't expect to be asking this soon: if I can just talk to my systems, what are admin panels for?

My password manager is starting to feel neglected, by the way. I've been meaning to check in on it, but that would mean logging in.

The login tax nobody puts on the invoice

Here's a partial list of the admin interfaces I work in, across my own businesses, client work and enterprise work. More and more of them take a conversation instead of a tour:

• Content management systems like WordPress, Drupal and Sitecore. Posting and editing content, adding and removing users. The work takes seconds; finding the right screen took minutes.
• Salesforce Marketing Cloud. Everything from data extensions to journeys to the automations in Automation Studio. If you've used it, you know it's a platform where one change can live four menus deep.
• Observability platforms like New Relic and Dynatrace. Setting up synthetic monitoring, with a bonus I didn't see coming: the cryptic terms and threshold conditions come back in plain English. I describe what "broken" means, and the AI translates it into the vendor's dialect.
• Ticketing in Jira and ServiceNow. Reading, summarizing, updating and editing any ticket, without opening one of those famously busy forms.
• Infrastructure as code with Terraform. Asking what the current state of the infrastructure is, then changing it through the AI, is cleaner and faster. It also keeps environments in parallel. We'd built a layer of YAML config files that drove scripts running Terraform dynamically, one per environment, and keeping five of them in step by hand is exactly the kind of chore that quietly drifts. Now the change is described once and applied across all of them.
• Google's consoles: Google Cloud runs by command now, and Google Analytics and Tag Manager both have APIs that work the same way. For my own Analytics, the last step is an access grant I have to click inside Analytics myself, which is its own small lesson about where logins still live.
• And more: privacy and consent platforms like Ketch, ad servers from open-source ones like Revive Adserver to Kevel. The list seems endless.

None of these is bad software. Some of them have interfaces that are, let's say, character-building. But the real cost isn't any single console. It's the wandering: sign in, find the right account, find the right property, open three menus, discover the setting moved in last quarter's redesign, and finally make a change that takes about four seconds.

Add up those minutes across a week and it's a real line item, even though it never shows up on an invoice.

What replaced the clicking

The change wasn't "AI does everything." That would be expensive, slow and occasionally creative in ways you don't want near a production website.

What I've been building is closer to a single dashboard that knows the difference between a routine call and a judgment call.

• Routine work runs as plain, deterministic calls. Publishing an approved post, checking whether a site is up, reading a setting, running a deploy. Those are fixed operations with a right answer, so in my setup they run as ordinary scripts and API calls. The operation itself spends no AI tokens, and it does the same thing every time.
• AI steps in only when thinking is actually needed. Drafting, summarizing, spotting what's wrong with a page, deciding which of three approaches fits. That's where a model earns its keep.

In my experience, most of what a business asks of its systems falls into the first bucket. Once you route the routine work away from the AI, the conversation stays cheap and the results stay predictable. The AI becomes the front desk that knows which department to call, not the department doing every job.

Underneath, this only works because the platforms expose an API. Google's consoles are the easy example: there's the gcloud command-line tool for Google Cloud, the Google Analytics Admin API and the Tag Manager API. Standards like the Model Context Protocol are making it easier to plug an AI assistant into those same doors. The admin screen was always just one way in. It's no longer the only one.

The credentials haven't disappeared, either. They live in the tooling I set up once, not in the chat. "No login required" means I don't do the logging in, not that nothing is secured.

It isn't one AI doing all of this, either. Claude does a lot of it, but Google's Antigravity generates the imagery for my sites from written briefs and has shipped deploys of its own (one of which taught us to add a gate: nothing goes live unless it's on the main branch, a slip a Claude session repeated the same day). GitHub Copilot reviews code, has built a few features, and writes commit messages for the code I write myself, which saves time and leaves a better trail for the other agents to read. Part of the job is knowing which tool to hand which task.

So what are admin panels still good for?

Plenty. I don't think they vanish. I think they change jobs.

So my prediction is that admin consoles become read-only windows. You'll visit them to look, to verify and occasionally to rescue. The day-to-day work (the changes, the publishing, the routine checks) moves into conversation.

Click to enlarge

Site Ops, the dashboard I built to watch my own five websites. One screen, no logins, and when something needs attention it says what.

The part where I tell you to keep checking

Talking to your systems doesn't mean you stop paying attention to them. Two habits make this safe:

1. Approve anything new before it goes public. Every new AI-written article on my sites waits for my yes, just like those two did. The approval moved into the chat. It didn't go away.
2. Review the code the AI builds as thoroughly as your goals and time allow. Much of the code behind that dashboard was written with AI help, and a confident-sounding conversation is not a reason to trust it. Tests and automated checks help, but they come in addition to review, not instead of it.

Where this series goes next

No Login Required will get into the specifics: the story of that no-login website update, how the "don't think unless you need to" routing works in practice, what it took to move five websites off WordPress, and the automated checks that run before anything reaches me.

The first three follow-ups are on Brandager, for owners weighing what this means for the software they pay for:

• The Login Tax: What Admin Screens Really Cost
• Enterprise Platforms Without the Guided Tour
• Buy Software That Has an API

If you're a business owner spending real hours a week inside vendor dashboards, you're probably wondering where your version of this starts. That's exactly what my one-on-one AI guidance sessions are for. Let my Claude help your Claude. We'll look at the systems you actually use, find the routine work worth taking off your plate, and set it up so you stay in charge of the decisions that matter. Book a free 30-minute consult to start.