Generated screens will swallow the software you only look at

Somewhere in your company there’s a spreadsheet next to the ERP, and it’s the schedule the shop works to.

Nobody planned it. The ERP was bought to run the business, the business didn’t fit, and somebody built the spreadsheet to cover the gap. Now the company runs on both, and keeping them in step is a job that isn’t on any org chart.

Every application you use is a set of decisions someone made before you arrived. A production manager wants capacity by plant, so somebody builds that screen. Then capacity by week. Then commitments against capacity. Then labor against capacity. Business software piles up hundreds of reports and still never catches up, because it has to guess in advance who will be looking and what they’ll need to see.

For the entire history of business software, we’ve forced companies to adapt to it, instead of the software adapting to the business. I don’t think that lasts, because software itself is commoditizing. My bet is that the agents, workflows and applications a company uses will be generated for it, pretty much on demand. Screens are already being drawn that way.

When a screen costs next to nothing to make, software that only shows you your records stops being worth a long contract: the ERP’s reporting module, the BI seats, the dashboard add-on, the custom report you raise a ticket for and wait a month on.

The vendors you pay are putting AI inside their own apps, betting the app keeps its place. From where they sit it’s the obvious move, and inside their own walls it will help. But an agent inside your ERP knows what your ERP knows, and the spreadsheet exists because that was never enough.

Screens are getting cheaper to generate. What they draw from isn’t: which job is late, which source won, what an agent is allowed to change. That only exists inside your business. So own the record every screen reads from, across the ERP, the inbox and the spreadsheet, and stop signing long contracts for software that only displays it. Building that record is our business, so weigh my bet with that in mind.

What is a generated interface?

A video game doesn’t draw every wall in its world all the time. It holds enough to rebuild the world, works out what you can see from where you’re standing, and draws that. A generated interface does the same for a business.

A generated interface, also called generative UI, is a screen an AI model draws when you ask a question: a table, a chart, a form or a small working app, built for that question and gone after it. Google started shipping it in its Gemini app in November 2025,¹ and ChatGPT added it in October 2026.²

It isn’t finished. In Google’s study, people rated screens designed by human experts highest, with generated screens close behind and well ahead of a chatbot’s plain text. Google also says a screen can take a minute or more to draw, and its test didn’t count the wait.¹ So hand-built screens still win wherever someone predicted the question.

We found out how hard the predicting is ourselves. Our own tool doesn’t draw screens yet. At a steel fabricator we work with, the model picks from views we designed. The day before we showed the scheduler’s planning view, I told our team it was confusing and not what we needed, and spent the day reworking it. We’d sat with the scheduler in June, and the view still wasn’t right.

Hand-built or generated, every screen reads from the record, so that’s where we put the work first.

Why can’t my ERP’s AI assistant see the whole job?

Because each app holds a slice. Your ERP sees orders and stock, and the scheduling tool sees crews and dates. Each one claims to show you the company, and none of them contains it. The facts that decide whether a job ships on time, like the change order or what the supplier promised on the phone, sit in inboxes and with whoever took the call.

The fabricator’s scheduler plans ahead by dragging placeholder jobs from a forecast into each plant’s schedule inside the ERP. What the ERP doesn’t reliably hold is how much work each plant gets through in a day. The scheduler carries that from experience. When we talked in June, the system’s capacity figure wasn’t one the shop relied on. An agent inside the ERP can read every placeholder and still not know it.

And a generated screen gets wrong whatever the record gets wrong. Take the email the ERP sends listing overdue purchase orders. An order marked overdue might mean the steel is late, or the steel arrived and nobody checked it in, or the order just needs closing. It takes someone who knows the shop to tell them apart.

A generated screen can decide how that order looks. It can’t be allowed to decide which of the three it is. That has to come from the record, and a record with one status for all three hands the same status to every screen.

Can someone approve from a screen that disappears?

A generated screen disappears once the question is answered. That’s fine until someone signs something from it. If a PM approves a change order or a buyer releases a purchase order from a screen that no longer exists, nobody can show what they were looking at.

So save it: the screen, and where each number on it came from, stored with the signature. Treat it like the drawing revision a fabricator signs against. Ask any vendor selling generated screens how they do this, and until they can show you, keep approvals off them.

Which screens stay

On a shop floor, a screen that looks the same every day is a feature. Operators know where the number is without reading it, and a screen that takes a minute to draw has no place there, for now. Those stay, and so do the systems where orders and money are kept.

Everything else can be drawn for whoever is asking: first the reports nobody built, then the everyday views, like capacity by plant or one customer’s open jobs. Each of those is a screen a vendor charges for today.

Before your next renewal

1. List your subscriptions. Mark which ones hold records and which only show them.
2. Keep the second group on short terms. They usually cost more per year, and the premium buys you the exit.
3. Test the AI add-on before you buy it. Ask it one question from last week that needed the inbox or the scheduling spreadsheet. If it can’t answer, you’d be paying for your own data, in the vendor’s app, with a chat box on it.
4. Ask for the exit at renewal. Can you export your data, and the rules and settings you’ve configured, in a format another tool can read?