# Design the delegation, not the screen. Written by Bart Van Hecke ยท Imported 2026-10-20 **Type:** article AI agents now do part of the work inside SaaS products. Users no longer click through every step; they set a goal, define the limits and approve what matters. That shifts the interface from a set of screens to a control plane for partly autonomous work. The design question changes with it: not "where does this button go?", but "which part of this process deserves autonomy, and how does the user stay in control?" *Agentic SaaS turns product design into the design of trust. Here's what changes, and what doesn't.* SaaS is shifting from software where people perform actions to software where people define intent, limits and accountability. Agents do a growing share of the work. The interface stops being the place where work happens and becomes the place where work is directed, checked and corrected. Pixelmojo calls this the move from UX to AX: Agentic Experience. I agree with the direction. I disagree with how fast most people think it's arriving. ## Agents clicking through your UI are a bridge, not the destination The popular image of an AI agent is one that operates existing software like a human: reading screens, clicking buttons, filling in forms. That capability matters. It's how agents will reach legacy systems that were never built for them. It is rarely the end state. UI automation is slower, more fragile and harder to audit than a well-designed tool interface. Computer use is an escape hatch for software that isn't agent-ready yet. The products that win will expose their capabilities directly. ## The interface becomes a control plane Chat everywhere is not the answer. Chat becomes one input method among several, mostly in the background. A strong agentic product needs six things: - **A place to state the goal.** What outcome does the user want? - **A visible plan.** What will the agent do, with which data and which tools? - **Checkpoints.** Mandatory approval where money, reputation, privacy or irreversible changes are at stake. - **Live status with a way to intervene.** The user can pause, redirect or take over at any moment. - **An audit trail.** What happened, why, based on which source, with what result, and how to undo it. - **An exception queue.** The place where human expertise is actually needed. Pixelmojo's six AX patterns map onto the same timeline: intent preview and an autonomy dial before the action, rationale and confidence signals during it, audit, undo and escalation after it. Different vocabulary, same conclusion. The screen is no longer a set of forms. It's the governance layer for partly autonomous work. ## Your product becomes a capability When agents call software instead of people, a SaaS product stops being only a destination. It becomes a capability that other agents discover and invoke. MCP is an early expression of this: a standard way for models to reach context and tools. That moves differentiation away from a polished CRUD interface and toward: - unique, reliable data - a sound domain model - safe, clearly bounded actions - permissions and approval flows - consistent tool contracts - observability and recoverability APIs don't become less important. They become product. A tool now has to be technically correct, legible to a model, safe under ambiguity and inspectable by a human. That's a design problem, not only an engineering one. ## Progressive autonomy beats maximum autonomy The best agentic products will not be the most autonomous ones. They'll offer **progressive autonomy**: - **Observe.** The agent watches and learns the workflow. - **Suggest.** It proposes actions; the human decides. - **Execute with review.** It acts; the human approves before anything sticks. - **Act within a mandate.** It works independently, inside limits it has earned. Trust is calibrated over time, per task and per risk level. The product that designs this transition most elegantly wins the user's confidence. Trust is the currency of agentic SaaS. ## What this means for product designers Classic UX optimizes the path a user takes. Agentic UX optimizes the quality of delegation. The designer's role shifts accordingly. Less time goes into perfecting a linear happy flow. More goes into: - **Workflow design.** Which part of this process deserves autonomy, and which doesn't? - **Trust design.** When does the system act, explain, ask or stop? - **Policy design.** How do business rules and risk levels become behavior a user understands? - **Designing for three users at once.** The end user, the agent and the operator who monitors and intervenes. - **State and exception design.** Uncertainty, failure, recovery and handoff, not just success. - **Behavior evaluation.** Defining "good enough" with product and engineering: accuracy, cost, speed, reversibility. The deliverables change too. Flows and high-fidelity screens are no longer the full specification. Add tool schemas, permission matrices, agent state models, escalation criteria, example traces and evaluation sets. Figma remains useful. It's no longer the whole spec. Pixelmojo flags screen-level UI, motion and microcopy as the most exposed to commoditization. I'd put it differently. Those skills don't disappear; they stop being the differentiator. The differentiator is deciding what the system should be allowed to do. ## The reality check Demos are far ahead of production. Deloitte found that only about 11% of companies have agents running in production. Figma's AI report shows most designers feel faster with AI, yet fewer than half say quality improved and only around a third trust the output. Teams report duplicated agent work, added oversight and time spent fixing results. That gap is the point. Generating interfaces, code and agent actions is getting cheap. Knowing what's good, what's safe and what scales is not. The bottleneck has moved from production to judgment. ## Where I come in I work with SaaS startups and scaleups on exactly this layer: deciding where autonomy belongs, designing the control plane around it, and building the design system and component architecture that keep it consistent as agents and teams ship faster. If your product is adding agents and you're not sure the experience earns your users' trust, [let's talk](https://superflow.design/). ## Want this applied to your product? Book a call. [Book a call](https://superflow.design/contact)