If you squint at the first version of ChatGPT, you see a search box, a blinking cursor, and an answer underneath. The most disruptive product of the decade arrived dressed in the most familiar interface of the Cloud era.
New technologies often begin by imitating what came before. The first films drew on theater's aesthetics, the first photographs borrowed poses and genres from painting, and the first iPhone apps looked suspiciously like real-life office counterparts. Imitation makes the unfamiliar easier to adopt by letting habits guide the way.
I argue that enterprise AI is currently in such an imitation phase. Better search, better dashboards, smarter forms, and conversational assistants inside familiar applications. Those are all reasonable first moves to capture value from AI. But we risk mistaking a tweaked version of the past for the genuine new.
AI inherited Cloud-shaped products
Every technology carries the compromises of the conditions under which it was created. Software was expensive to build, risky to change, and hard to integrate. Because building was expensive, a single durable product with a common data model had to serve everyone. Because change was risky, work froze into predefined paths and prescribed workflows. Because integration was brittle, each application could reliably manage its own records, e.g., an order, an invoice, or a payment, but not the relationships among them. Users had to cross those application boundaries themselves, connecting the records and assembling the meaning across multiple screens. Because complexity and variety had to be dealt with, standardization became the means to make complex work repeatable at scale.
Enterprise cloud software has been and remains incredibly successful. Worldwide public cloud spending reached $723 billion in 2025, up from just under $600 billion the year before. The model won.
The familiar shape of SaaS, which has become the default, was never inevitable. It is a result of the economic constraints of the Cloud era.
Lost in user translation
Consider a business user asking why invoices in one region are being paid late. In a conventional enterprise stack, that question often begins with an orientation exercise. Which application, among the 600 the average enterprise now runs? Which process step? Which object? Which report? Which owner? The user must turn the actual problem into the vendor's architecture before the system can help.
That is more than a user-experience inconvenience. It is a translation exercise, and it has been measured. Researchers who followed 137 employees across three Fortune 500 companies found them toggling between applications and websites about 1,200 times a day, spending nearly four hours a week just reorienting after each switch. Roughly 9% of their time at work. In one company, executing a single supply-chain transaction took about 350 switches across 22 different applications. Per person. The software did not run the process. The people carried the process across the software.
None of this is the user's fault. The software decides the sequence. Every capability is a place you have to visit. The dashboards show what is happening, but leave you to work out what it means.
The same is true of records. Enterprises need authoritative records of suppliers, orders, and payments. Without them, there is no trust. But a record is not context. It can tell you what exists without telling you why it matters, what changed, or what can safely happen next.
Too often, the business expert must first become an expert software user.
The application is not the process
AI changes the economics of software. It makes it possible to assemble a bounded experience around a business objective, rather than force that objective through a permanent product shape.
This is what I mean by situated software. It does not begin with the question of which application owns a task. It begins with the outcome the business is trying to achieve and creates a task-specific solution grounded in the company's context.
This does not mean all UIs and workflows will go away. Stable, high-volume, regulated work will continue to need durable records, controls, and interfaces. The change is that today’s application layer won’t be the only viable surface for work to happen.
Return to the late-invoice question. A business user should not have to know whether an order-to-cash problem lives in a process mining view, an ERP screen, or a workflow tool. The system should start with the objective: why are invoices in this region being paid late, what is the impact, and what can be changed safely?
Imagine a workspace built for that objective. It shows how the process works in each region, which systems and owners are involved, the likely business impact, open risks, and relevant controls. It can distinguish between a local adaptation that helps the business and a workaround that has become process debt. The expertise and context sit next to the decision.
The visible form may be a temporary dashboard for a steering meeting, a conversational response, or a workspace custom-built for a single change process. It may exist for an hour or a month. The durable asset is not the UI surface. It is the ability to generate the right surface for the task.
Standardization moves down
Standardizing enterprise software during the cloud era wasn't a mistake, and customization wasn’t its solution. It was how companies made complex work reliable. The problem was applying standardization at the surface, where users had to conform to the product, instead of below the surface, where variation could remain safe.
But if all interfaces are in flux, how does this not end in chaos? For some companies, it will. But situated software is not a move from order to chaos. It is a move from order at the surface to order at the foundation. The experience on top can be as fluid as the moment demands, but the order underneath cannot. Standardization relocates, it does not disappear.
Situated software needs rock-solid foundations. The foundations are your company’s process reality, canonical records, permissions, policies, and validation procedures. Putting an app-generating AI on top of a shaky foundation, without durable constraints, only amplifies the incoherence. Companies that invest here can afford variety on top. Companies that do not will watch AI tear apart their processes.
The foundation has to carry the order
This is where SAP Signavio becomes important. Companies need to know which process each generated experience touches, what data and systems it relies on, which metric defines success, which owner is accountable, what controls apply, and whether the change actually improved the business. If order-to-cash has five different manifestations across five regions, the company still needs a coherent way to compare performance, understand variation, govern changes, and connect improvements to value.
Process intelligence shows how work actually flows. Process observability helps inspect what changes when agents and generated applications act. Company Memory provides the company-specific context, constraints, and operating logic that keep situated software aligned with how the business wants to work. Value management connects all of that activity to validated business outcomes.
The end of imitation as destiny
New technologies usually imitate the old. That is a fair starting point. But imitation is a phase for bridging the old to the new, not a destination, and the long-term winners are rarely the ones who reproduce the old form most efficiently.
Gutenberg is remembered for his Bible, but producing it was so slow and capital-hungry that his financier sued him and took the workshop. The earliest runaway successes of the printing press weren’t the old formats, but new ones, such as administrative papers and indulgence forms with standardized text and blank spaces for names, places, and dates.
The next phase of enterprise software will not be won by the companies that imitate the old stack most efficiently. It will be won by the companies that exit the imitation phase, because they invested in the foundation that makes situated software safe to generate.
Imitation is the first draft. I am excited to see what gets written once software stops imitating itself.
--
With thanks to Lukas N.P. Egger for his input, review, and the discussions that helped sharpen this piece.
References:
- Worldwide public cloud spending reached $723 billion in 2025, up from just under $600 billion the year before.
https://www.gartner.com/en/newsroom/press-releases/2024-11-19-gartner-forecasts-worldwide-public-cloud-end-user-spending-to-total-723-billion-dollars-in-2025
- Researchers who followed 137 employees across three Fortune 500 companies found them toggling between applications and websites about 1,200 times a day, spending nearly four hours a week just reorienting after each switch. Roughly 9% of their time at work… In one company, executing a single supply-chain transaction took about 350 switches across 22 different applications. Per person.
https://hbr.org/2022/08/how-much-time-and-energy-do-we-waste-toggling-between-applications
- Which application, among the 600 the average enterprise now runs?
https://zylo.com/news/2025-saas-management-index