I think I’ve just crossed one of those slightly strange lines where a system stops feeling like a piece of software I built and starts feeling like something completely different.
I can now ask my business questions.
Not metaphorically.
I can type:
“What’s moving fastest right now?”
Or:
“Which angle is winning this month?”
Or even:
“How many Go To Black Incredible Dark Maximisers have I sold this month?”
And ChatGPT can go into my 1J2 Control Room data, inspect the relevant records and answer me.
That’s considerably more interesting than the Google Sheet that makes it possible.
It started as a normal dashboard
I’ve been building the 1J2 Control Room as the central place for all the systems I now use.
My TikTok affiliate operation alone produces a ridiculous amount of information. There are videos, hourly view snapshots, products, orders, commission, content angles, transcripts, Creative Studio projects, mini-site visitors, shop clicks and strategy tasks.
Originally, the job of the Control Room was straightforward: bring that information together and show me what was happening.
That meant dashboard cards. Tables. Rankings. Traffic comparisons. Video movers.
All useful.
But dashboards have a fundamental limitation.
You have to know where to look.
If I want to know how a particular product is doing, I have to go to one section. If I want to know whether an old TikTok has suddenly started moving again, I need another. If I want to compare content angles, there is another screen for that.
The information exists, but I am still the person doing the interrogation.
The Google Sheet was supposed to be a simple bridge
I recently built a read-only export of the Control Room into Google Sheets.
The workbook contains things like:
- Dashboard
- Video Movers
- Latest Videos
- Video History
- Weekday Baselines
- Angles
- Products
- Recent Orders
- Mini Sites
- Mini Site Clicks
- Creative Studio
- Strategy
- System information
The export runs automatically from the same underlying database that powers the Control Room.
I originally thought of this as little more than a convenient way of giving ChatGPT visibility of my systems without opening up the production database.
It turns out that was underselling it enormously.
The interface became the question
Once ChatGPT could read the workbook, I stopped needing to think in terms of dashboard pages.
I could just ask the thing I actually wanted to know.
For example:
“How is today’s bike video doing compared with yesterday’s?”
That potentially involves the latest video records, historic metric snapshots and the age of each video.
I don’t need to assemble that comparison manually.
I can ask:
“Which videos have accelerated in the last hour?”
That is a Video Movers and Video History question.
I can ask:
“What angle is winning?”
That is an Angles question.
I can ask:
“Which products are getting views but not producing orders?”
Now we are crossing Products, video performance and orders.
And that is where this suddenly became different from a dashboard.
The interface is no longer the navigation.
The interface is the question.
A good example came from a bottle of tanning cream
I asked how many Go To Black Incredible Dark Maximisers I had sold this month.
The information was already sitting in the Recent Orders tab.
Each row contained the order date, product, quantity, status, GMV, commission and associated video.
So the job was simply to find the August orders for that product, ignore the ineligible ones and total the quantities.
The answer was 10 qualifying units.
There were 14 units across all August rows, but four of them belonged to orders marked ineligible.
That sounds like an extremely mundane calculation.
That’s exactly why I like it.
I didn’t build a special “Go To Black sales this month” report.
I didn’t need to.
The underlying information was structured well enough that the question could become the report.
It also exposed a flaw immediately
Then I asked:
“What’s in my work queue?”
And we discovered something interesting.
The Control Room itself knew the answer, but the Google export didn’t.
The dashboard has an “Attention needed” area which can contain things like draft articles, untagged assets, unmatched orders, missing transcripts, enquiries and Strategy tasks.
But I hadn’t exported that assembled queue.
That meant ChatGPT could see several of the ingredients, but it could not truthfully tell me what the Control Room itself considered to be my current work queue.
That was exactly the kind of mistake I wanted this architecture to expose.
So we fixed it.
The Control Room export now has a dedicated Work Queue dataset using the same underlying rules as the dashboard.
The important part is that ChatGPT is explicitly told not to reconstruct or guess the queue from other information.
If I ask what is waiting, it should read the Work Queue.
If the answer is 600 untagged assets, tell me 600 untagged assets.
If tomorrow it is six draft articles, tell me six draft articles.
Facts first.
This changes how I think about building admin systems
Traditionally I would build an interface for every useful view of the data.
That still matters. I like my Control Room and I’m not about to replace it with a blank chat box.
Visual dashboards are brilliant for scanning.
But conversational access changes what needs to be pre-built.
I no longer have to predict every question Future Me might ask.
I need to make sure the underlying data is clean, well-labelled and accessible.
Once that exists, completely new questions become possible without building another PHP page.
That is a fairly fundamental architectural shift.
The really interesting questions aren't simple totals
The sales calculation is useful, but the questions I’m most interested in are the ones that combine several parts of the system.
Things like:
- Which old videos are suddenly coming back to life?
- Is today's traffic genuinely unusual for a Tuesday?
- Which video has the strongest momentum rather than the highest lifetime views?
- Which content angles generate the best orders per video?
- What is getting plenty of traffic but failing to convert?
- Which existing TikTok should I make another variation of?
- What has changed since the last Control Room snapshot?
- What is working that I should do more of?
Those are analyst questions.
And increasingly I don’t need to sit there behaving like the analyst.
Voice was supposed to make it feel even more futuristic
There was one part of the experiment that was less successful.
I wanted to wake up in the morning, open ChatGPT Voice and simply say:
“What’s in my work queue?”
That would be lovely.
In practice, the current voice experience was unreliable for this. It struggled with some of the wording and, more importantly, didn’t have the same dependable connected-data behaviour I was getting through normal chat.
At one point it didn’t even respond properly to me saying hello.
So that part can wait.
The significant thing is that the underlying system already works.
Typing or dictating a question into the normal conversation gives me the experience I was actually trying to create.
I don't think this is a dashboard anymore
That is probably the biggest conclusion from the build.
I started by creating a dashboard.
Then I created a Google Sheet so another system could read the dashboard data.
What I have now is beginning to look much more like an API for understanding my own business, except the query language is English.
The database still does the recording.
The Control Room still does the visualisation.
The Google export provides a safe, read-only representation of the important information.
And ChatGPT becomes the interrogation layer over the top.
None of those pieces is particularly revolutionary on its own.
Put them together, though, and the experience is completely different.
I can ask my business what is happening.
And increasingly, it can answer.
That is a build I’m rather pleased with.
