You’ve probably noticed the panic setting in across enterprise software teams. The mandate comes down from above: “We need to add AI.” So, what do you do? You bolt a chatbot onto your legacy SaaS platform, maybe add a summarization feature, and call it an innovation sprint.
But let’s be brutally honest: Adding an AI assistant to a legacy SaaS platform isn’t innovation; it’s just a digital band-aid on a broken workflow.
A recent government directive in China—specifically the Ministry of Industry and Information Technology’s push for “Agent Software”—just forced the tech world to confront a reality we’ve been ignoring. The next generation of software isn’t about adding a fancy generative AI feature to your existing UI. It’s about completely reorganizing what software actually does.
For decades, we’ve built software by chopping human work into isolated, discrete Functions. A query is a function. An approval is a function. Submitting a form is a function. As product managers, we design pages, menus, and buttons around these functions. Even when the LLM wave hit, we didn’t change this logic. We just turned those functions into AI tasks: “Summarize this,” “Generate that.”
But the real disruption begins when AI can continuously pursue a single goal. It reads the context, finds the missing pieces, calls different tools across different systems, and pushes a task forward over days. At that point, you’re no longer designing a Function. You’re designing Work.
Functions don’t disappear; they become the fuel for autonomous Work. If your product only offers isolated features, it will be reduced to a background API.
Think about what happens to your precious UI. Today, if a user wants to close a deal, they open your CRM. If they want to check inventory, they open your ERP. We’ve built our entire industry around software as the destination. But when an Agent can orchestrate work across the CRM, ERP, and knowledge base simultaneously, the user’s entry point is no longer your software. It’s the completion of the task itself.
Your product will not be “eaten” by AI. Your UI will simply be bypassed. The new “user” isn’t a human clicking a button; it’s another software agent calling your API. If your capabilities aren’t recognizable, callable, and controllable by an AI, you are invisible.
But here is the twist nobody is talking about. The more capable your agent becomes, the more critical it is to design for what it cannot do.
I saw this firsthand in enterprise bidding workflows. An Agent can read an RFP, extract scoring criteria, match company qualifications, and draft a technical proposal. You can even give it the tools to fill out and submit the final forms. But “technically able to submit” and “business authorized to submit” are two entirely different universes.
When software just gives answers, boundaries are easy. When software takes action, you have to untangle Capability, System Permission, Business Authority, Formal Effect, and Responsibility. Just because an API call succeeds doesn’t mean the business effect is legally bound. And you absolutely cannot push the final liability onto an AI.
A great agent product isn’t defined by how much it can do, but by exactly where it knows to stop.
Governance can no longer be a hidden permissions page in the backend settings. It has to be baked into the runtime. You have to design exactly when the system gives a suggestion, when it forms a candidate result, and when it forces a human to take the wheel.
If we are truly moving from one-and-done prompts to continuous work loops, we need to radically upgrade our software with four layers of capability. I don’t care about new buzzwords; I care about these four questions:
1. Can it recognize the context? (Workspace & Context) If the Agent doesn’t know exactly what work it is currently inside of, giving it more tools just increases the blast radius of its mistakes.
2. Can it connect the past to the present? (Knowledge & Memory) Was a fact confirmed yesterday still valid today? Work spans days. The system must remember what to keep and what has expired.
3. Can it push the loop forward? (Work Loop) A task isn’t a single prompt. The system must look at the execution result, handle exceptions, and decide the next step across multiple systems.
4. Can it govern itself? (Responsibility Boundary) What can be automated? What requires human confirmation? If something goes wrong, can you reconstruct exactly what happened?
We are standing at the edge of a massive identity crisis for product managers. We are no longer designing screens. We are designing how work gets done across systems, and where agents are allowed to stop.
Interfaces will change. Agent forms will change. But the question is already unavoidable: Are you still designing a list of functions, or are you designing a product that can actually take on a piece of work?
Stop polishing your UI. Start designing your boundaries.
FAQ
Q: Will AI agents completely replace traditional SaaS interfaces?
A: No. Legacy systems hold the ground truth of business data and rules. Agents won't eat them; they will just bypass the UI and call their APIs. The human UI becomes a secondary entry point, not the primary one.
Q: What does this mean for product managers right now?
A: Stop asking 'what feature does the user need?' and start asking 'what continuous work does this product execute?' You are no longer designing screens; you are designing work loops and responsibility boundaries.
Q: Is adding an AI assistant to my current product a waste of time?
A: Mostly, yes. Slapping a chatbot on a clunky, multi-step workflow is just putting lipstick on a dinosaur. If you don't reorganize your software around the Work to be done, the agent will just expose how broken your UX really is.