Arin Bhowmick, SAP’s chief design officer, shifts the software-death debate from apps to operating context. Rather than treating the model’s capabilities as the whole story, he asks where enterprise software’s enduring value actually sits:

Despite the cliché premise, I found it insightful, an honest take from one of the “Big Four” on AI and SaaS. In it, PwC’s TMT leader Dallas Dolen says enterprise software isn’t dead. His clients are redirecting their software budgets toward AI, and the systems going in on top of the stack they already run, keeping what runs the business and building the intelligence on top. They are betting that the value sits in what they have spent years assembling.

There’s a good reason for that. An AI model can answer almost anything, but on its own, it knows nothing about your company: how it runs, which data to trust, what it is allowed to touch…

The model is the “cheap” part, while the expensive one is the business knowledge around it, the processes, exceptions and rules a company encodes over decades. That context provided by software is what makes an AI agent useful.

The existing stack becomes the policy layer for whatever intelligence sits above it. That context determines more than what the agent knows. It sets the boundaries of action: when the system can proceed, when a person must approve, what happens when the evidence conflicts, and who is accountable afterward. Bhowmick turns those boundaries into the design brief:

Once design “owns” the judgment a system applies on its own, a new set of questions lands on every designer’s desk:

• When should an agent act?
• When should it stop and ask first?
• What does it show before it does something it can’t undo?
• When it isn’t sure, does it say so, or hand you a confident guess?
• When it gets something wrong, who answers for it?
• How does a person take back control when they need to?

Often, the most useful move an agent can make is to admit when it’s unsure and hand the decision back. We have to design that handoff so it feels natural. If it reads as the software confessing a mistake, people lose trust.