AI isn't slow because of the model.
Software is still designed for humans.
Three a16z partners reached the same conclusion in the same month. The prompt box isn't the completion of the AI revolution — it's the bottleneck.
Everyone thinks "just use better prompts," right?
Honestly, that's how most teams use AI today. Keep a ChatGPT tab open, type when you need something, copy-paste the result into another tool. Feels like enough AI usage, right?
Here's the uncomfortable math. The AI software market is $300–400B a year. The US labor market is $13 trillion. That's a 30x gap.
Marc Andrusko (a16z) says this gap exists because AI is still stuck as a tool that "answers what you ask." For agents to actually replace labor, the software interface itself needs to change. That shift is happening across three axes simultaneously.
Gartner Forecast
By end of 2026, 40% of enterprise apps will embed task-specific AI agents — up from under 5% in 2025.
Three shifts happening simultaneously
| Axis | Old Paradigm | Agentic Paradigm |
|---|---|---|
| Interface | Prompt box | Execution engine + approval gates |
| Software design | Visual hierarchy for humans | Machine-legible for agents |
| System | System of Record (SoR) | Agent execution layer |
Shift 1 — "Death of the Prompt Box"
Andrusko predicts future apps will require minimal prompting. Instead of users typing requests, apps will observe user behavior and proactively suggest the next action, asking for approval.
An AI CRM should work like this: before a salesperson opens the CRM, the agent has already identified dormant leads. It analyzes email history, pulls calendar data, drafts outreach emails, and asks "Should I send this?"
The human only does one thing: click approve. Andrusko says future power users will measure efficiency by "what percentage of my tasks ran without my direct intervention this month?" — he calls this the "Approvals-Free Rate."
Shift 2 — Software gets rewritten for agents
Stephanie Zhang's analogy lands well. Traditional journalism puts conclusions in the opening paragraph to capture human attention. Agents don't need that — they'll read content buried five pages deep.
This changes software design fundamentally. Good UI, intuitive navigation, and SEO-optimized titles used to matter. Agents don't need any of that. Agents need APIs, context, and permission to act.
Indeed already moved in this direction. After building a machine-legible design system that agents can read, they generated 4,300 AI prototypes in 4 months. JSON-based systems used 80% fewer tokens than Markdown — that's what made it possible.
Zhang also calls out a new category: "GEO tools." These optimize content so it surfaces when users ask ChatGPT for product recommendations. SEO is optimization for humans. GEO is optimization for agents.
Shift 3 — The first real threat to Systems of Record
Sarah Wang's take is the boldest. Salesforce, ServiceNow, and other Systems of Record are facing their first genuine threat in history.
Why haven't SoRs been disrupted until now? UI lock-in. The more employees use a system, the more muscle memory builds up, and that becomes a moat. Agents don't need UI — all they need is an API and context.
In ITSM (IT Service Management), the change is already visible. Agent-native newcomers like Resolve and Traversal are outpacing incumbents. The reason is simple: they iterate weekly, sometimes daily, building trust through consistent execution.
a16z describes the shift directly: "Control moves from the database to the intelligent execution environment where employees actually work." SoR gets demoted to data persistence, while the agent layer becomes where real work happens.
What you can do with your team right now
- Make your data machine-legible
Restructure process docs, CRM data, and content libraries so agents can consume them. JSON beats Markdown by 80% in token efficiency for agents. - Convert one repetitive task to an approval gate workflow
Find something your team reviews then approves. Flip it: agent drafts, human approves. That's the fastest way to feel what an agentic workflow actually is. - Understand GEO (Generative Engine Optimization)
Check how your business appears in AI search (ChatGPT, Perplexity). If you're only doing SEO, GEO is still an open first-mover opportunity. - Audit your SoR's agent strategy
Does your Salesforce, ServiceNow, or Jira offer agent APIs? What are the agent-native alternatives? You don't need to replace them now — but you need to know where the agent layer connects. - Define the agent supervisor role
Once agents run autonomously, someone needs to orchestrate and oversee them. a16z calls this an "AI workflow designer" — and they're calling it one of the next critical hires.
Further Reading
Big Ideas 2026: The Agentic Interface The original podcast with three a16z partners. 15 minutes. a16z.com
Is Software Losing Its Head? a16z's deep dive into how defensibility shifts as software goes headless. a16z.com
Agent UX: UI Design for AI Agents Plan-and-Execute, Confidence Signaling, Progressive Delegation — practical patterns for agent-first UI. fuselabcreative.com
Agentic Design Systems: The Complete Guide How to build design systems that agents can actually read. Includes Indeed's 4,300 prototypes case study. intodesignsystems.com
Agentic Interfaces Inside Your Product Practical framework for agentic interface implementation. Notion, Intercom, Microsoft 365 case studies. theinteractive.studio
