👋 The Wake Up Call
Everyone talks about the agent era like the whole software world is ready.
It is not.
The systems that actually run real businesses are still stuck in another century. Medical billing portals. Permit systems. Warehouse dashboards. Ancient back-office tools with bad UX and zero APIs.
So businesses do something absurd. They hire people to become the API.
A human logs in. Clicks three buttons. Copies one field. Pastes another. Repeats that dance 500 times a week. That is not operations. That is payroll-funded software compensation.
Now the door is cracking open. OpenAI’s GPT-5.4 introduced native computer-use capabilities, and OpenAI’s API docs now describe computer use as operating software through the UI by inspecting screenshots and returning actions for code to execute.
That changes the game.
The next gold rush is not another chatbot. It is software that bypasses software.
🦄 The Idea Drop: Portal Pilot
AI that works inside the old portals nobody wants to rebuild.
The Problem:
Small and mid-sized businesses are trapped inside software they hate but cannot leave.
A clinic still has to update payer portals manually. A law firm still uploads case docs into clunky county systems. A local contractor still files permits through interfaces that feel like they were designed during the dial-up era. None of these tools play nicely with modern systems. No API. No webhook. No clean export. Just friction.
That creates a weird labor sink. Teams are not doing high-value work. They are acting as glue between systems that refuse to talk. A shocking amount of “operations” is just professional tab-switching.
Traditional automation breaks here too. Rules-based bots are brittle. One moved button and the whole thing faceplants.
The Solution:
PortalPilot watches a human complete a task once, then turns that workflow into a repeatable agent job.
It logs in. Reads the screen. Finds the right fields. Clicks the right controls. Submits the work. If the layout shifts, it does not panic like a brittle selector-based script. It re-interprets the page and keeps going.
This is the wedge: no integrations, no vendor cooperation, no waiting for “digital transformation” to arrive in 2034. You automate the world exactly as it exists today, not as the SaaS brochure pretends it exists.
The first use case is narrow on purpose. Pick one painful workflow in one ugly portal for one niche. Then own it.
Features
Workflow Recorder
Record one human session and capture the real browser path, field logic, and decision points.Ghost Operator
Run the workflow in the background with new inputs, even when the interface shifts slightly.Approval Desk
Surface exceptions cleanly: “15 renewals ready. 2 missing data. Approve the rest.”
Business Model: Charge on hours replaced, not seats.
If you remove 40 hours of monthly portal work, pricing at $499 to $999 is easy to justify.
🚀 MVP Blueprint
We are testing for retention, not perfection.
1. The Tech Stack
Next.js for the customer dashboard and job history
GPT-5.4 for vision and computer-use reasoning across UI tasks
Playwright as the browser control layer
Browserbase if you want managed browser infrastructure fast. Browserbase positions itself as browser infrastructure for computer-use agents.
Skyvern if you want a faster starting point for browser-based workflow automation. Skyvern describes itself as automating browser workflows with LLMs and computer vision.
2. Core Features
Record a workflow once
Re-run it with new variables
Log proof of completion with screenshot + timestamp
3. Validation Plan & Estimates
Week 1: Pick one vertical. Medical billing is strong. Local permitting is even better because the pain is disgusting and constant.
Week 2: Shadow the ugliest workflow in that business. Find the one staff members dread.
Week 3: Build the recorder and replay flow for that one job.
Week 4: Add exception handling and approval queue.
Budget:
Cheap if you stay focused. Browser infra + model usage + hosting can stay lean for an MVP, especially if you are running a single narrow workflow first. GPT-5.4 pricing is listed on OpenAI’s model page, and Browserbase/Skyvern both offer developer-facing starting points rather than requiring enterprise-scale setup.
Success metric:
The customer does not manually touch that portal for 7 straight days.
Ask for the MVP @ NexTribes.
🧠 Founder Lesson: The moat is the mess
In the old world, bad software looked like a dead market.
In this world, bad software is the market.
The harder a workflow is to automate, the fewer founders will touch it. That is your opening. While everyone else is building wrappers for polished SaaS tools, you are going after the grimy systems businesses actually depend on.
The moat is not code. The moat is your willingness to deal with the ugly stuff.
In the agentic world, "No API" is your moat.
⚡ Quick Tips
The Tool: Browserbase for managed browser sessions, or Skyvern if you want an opinionated automation head start.
The Competitor: UiPath at the top end. Your wedge is not “more features.” It is faster setup, tighter vertical focus, and simpler pricing.
The Book: Company of One by Paul Jarvis. Niche first. Land grab later.
Bye Bye.